(19)
(11)EP 3 449 465 B1

(12)EUROPEAN PATENT SPECIFICATION

(45)Mention of the grant of the patent:
13.05.2020 Bulletin 2020/20

(21)Application number: 17720738.8

(22)Date of filing:  24.04.2017
(51)Int. Cl.: 
G06T 7/73  (2017.01)
(86)International application number:
PCT/EP2017/059648
(87)International publication number:
WO 2017/186635 (02.11.2017 Gazette  2017/44)

(54)

RETINAL IMAGE PROCESSING

NETZHAUTBILDVERARBEITUNG

TRAITEMENT D'IMAGE DE LA RÉTINE


(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR

(30)Priority: 26.04.2016 US 201615139182

(43)Date of publication of application:
06.03.2019 Bulletin 2019/10

(73)Proprietor: Optos Plc
Dunfermline, Scotland, KY11 8GR (GB)

(72)Inventor:
  • FLEMING, Alan Duncan
    Edinburgh Scotland EH13 9JR (GB)

(74)Representative: Hoffmann Eitle 
Patent- und Rechtsanwälte PartmbB Arabellastraße 30
81925 München
81925 München (DE)


(56)References cited: : 
  
  • ALAN D FLEMING ET AL: "Automatic detection of retinal anatomy to assist diabetic retinopathy screening", PHYSICS IN MEDICINE AND BIOLOGY, INSTITUTE OF PHYSICS PUBLISHING, BRISTOL GB, vol. 52, no. 2, 21 January 2007 (2007-01-21), pages 331-345, XP020113152, ISSN: 0031-9155, DOI: 10.1088/0031-9155/52/2/002
  • G Linda Shapiro ET AL: "Computer VIsion - chapter 10 Image Segmentation", , March 2000 (2000-03), XP055199831, Retrieved from the Internet: URL:https://courses.cs.washington.edu/cour ses/cse576/book/ch10.pdf [retrieved on 2015-07-02]
  
Note: Within nine months from the publication of the mention of the grant of the European patent, any person may give notice to the European Patent Office of opposition to the European patent granted. Notice of opposition shall be filed in a written reasoned statement. It shall not be deemed to have been filed until the opposition fee has been paid. (Art. 99(1) European Patent Convention).


Description

[Technical Field]



[0001] The present invention generally relates to the field of medical image processing and, more particularly, to the processing of a retinal image to determine the location within the image of an anatomical feature of the retina.

[Background]



[0002] Many eye diseases can be detected and diagnosed by imaging the retina. In a busy clinic it is not always possible to record even basic information such as whether the image is of the left or right eye. Automated analysis and manipulation of retinal images is a growing field. Automated analysis by computerised algorithms can provide assistance to clinicians in the form of disease detection or improved presentation modes. Many automated methods require positional information of the basic anatomical landmarks to aid their operation. This positional information may be input into software which may, for example: (i) project a retinal image into a uniform format which allows accurate dimensional measurements of the retina to be made; (ii) analyse multiple retinal field-of-view image sets, for example to identify fields-of-view and confirm that at all required fields-of-view are present; (iii) register multiple retinal images so that the location of disease visible in one modality can be identified in another; (iv) montage multiple retinal images to present clinicians with a wider view of the retina in one image; (v) present the user with a fly-through presentation of the retina; and/or (vi) perform automated disease determination; this can be so that the software operates only on the relevant area (for example, the optic disc in glaucoma, the macula in age-related macular degeneration) or so that the software ignores irrelevant areas (for example, the optic disc in diabetic retinopathy).

[0003] Although methods have been developed for determining the locations of anatomical features of the retina such as the optic disc and fovea in fundus reflectance photographs (mainly red and green light), the technical problem of automatically and reliably determining the location of such features in a retinal images of other modalities, and for ultra-wide field views, presents a difficult challenge, owing to artefacts and pathology, variations in eye position and variations between image modalities.

[0004] The article titled "Automatic detection of retinal anatomy to assist diabetic retinopathy screening" by A. D. Fleming at al., published in Physics in Medicine and Biology 52 (2007) at pages 331-345, discloses methods for the robust location of the optic disc and fovea in retinal images. The elliptical form of the major retinal blood vessels is used to obtain approximate locations, which are refined based on the circular edge of the optic disc and the local darkening at the fovea. The methods have been tested on 1056 sequential images from a retinal screening programme. Positional accuracy was better than 0.5 of a disc diameter in 98.4% of cases for optic disc location, and in 96.5% of cases for fovea location. The methods are sufficiently accurate to form an important and effective component of an automated image grading system for diabetic retinopathy screening.

[0005] Reference is also made to "Computer Vision - Chapter 10 Image Segmentation" by G. Linda Shaprio at al., pages 28-37 (March 2000).

[Summary]



[0006] In view of the shortcomings of conventional retinal image-processing methods and devices summarised above, the present inventor has devised a non-transitory computer-readable storage medium storing computer program instructions which, when executed by a processor, cause the processor to process image data defining an image of a retina to determine a location of an anatomical feature of the retina in the image by: receiving the image data; calculating, for each of a plurality of pixels of the received image data, a respective local orientation vector indicative of the orientation of any blood vessel present in the image at or adjacent the pixel; calculating a normalised local orientation vector for each of the plurality of pixels by normalising the local orientation vector calculated for each of the pixels so that the magnitude of the normalised local orientation vector at each of the pixels takes a common predetermined value; operating on an array of accumulators, wherein each accumulator in the array is associated with a respective pixel of the received image data, by (i) for each of the plurality of pixels, adding a respective value to an accumulator associated with a pixel of the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in a predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels, and (ii) smoothing the values in the accumulators; and determining the location of the anatomical feature in the image of the retina using the location of a pixel of the received image data which is associated with an accumulator having accumulated an accumulated value, which accumulated value is within a predetermined range of an extremum of the accumulated values in the accumulators.

[0007] The inventor has further devised an image-processing apparatus comprising a processor and a memory, the memory storing instructions executable by the processor whereby the processor is operative to process image data defining an image of a retina to determine a location of an anatomical feature of the retina in the image by: calculating, for each of a plurality of pixels of the received image data, a respective local orientation vector indicative of the orientation of any blood vessel present in the image at or adjacent the pixel; calculating a normalised local orientation vector for each of the plurality of pixels by normalising the local orientation vector calculated for each of the pixels so that the magnitude of the normalised local orientation vector at each of the pixels takes a common predetermined value; operating on an array of accumulators, wherein each accumulator in the array is associated with a respective pixel of the received image data, by (i) for each of the plurality of pixels, adding a respective value to an accumulator associated with a pixel of the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in a predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels, and (ii) smoothing the values in the accumulators; and determining the location of the anatomical feature in the image of the retina using the location of a pixel of the received image data which is associated with an accumulator having accumulated an accumulated value, which accumulated value is within a predetermined range of an extremum of the accumulated values in the accumulators.

[Brief Description of the Drawings]



[0008] Embodiments of the invention will now be explained in detail, by way of example only, with reference to the accompanying figures, in which:

Fig. 1 is a schematic illustration of an image-processing apparatus according to embodiments of the present invention;

Fig. 2 is a first part of a flow diagram illustrating a process by which the image-processing apparatus of the first embodiment processes image data defining an image of a retina to determine a location of an anatomical feature of the retina in the image;

Fig. 3A shows an example of an Optomap™ image of a retina;

Fig. 3B shows a vessel-enhanced image of the Optomap™ image in Fig. 3A;

Fig. 3C shows a smoothed vessel-enhanced image of the Optomap™ image in Fig. 3A;

Fig. 4 is a continuation of the flow diagram in Fig. 2;

Fig. 5 is a flow diagram illustrating process S50 in Fig. 4;

Fig. 6 is a schematic illustrating the relationship between pixels psij and poij of a retinal image and an accumulator aij of an accumulator array described herein;

Fig. 7 is an alternative continuation of the flow diagram in Fig. 2, which illustrates a process by which the image-processing apparatus according to a second embodiment processes image data defining an image of a retina to determine a location of an anatomical feature of the retina in the image;

Fig. 8A shows an example of an Optomap™ green plane image of a retina;

Fig. 8B shows a response map generated by processing the retinal image in Fig. 8A in accordance with the first embodiment;

Fig. 8C shows a response map generated by processing the retinal image in Fig. 8A in accordance with the second embodiment;

Fig. 9A shows an example of another Optomap™ image of a retina;

Fig. 9B shows a response map generated by processing the retinal image in Fig. 9A in accordance with the first embodiment;

Fig. 9C shows a response map generated by processing the retinal image in Fig. 9A in accordance with the second embodiment; and

Fig. 10 is an alternative continuation of the flow diagram in Fig. 2, which illustrates a process by which the image-processing apparatus according to a variant of the first embodiment processes image data defining an image of a retina to determine a location of an anatomical feature of the retina in the image.


[Detailed Description of Embodiments]


[Embodiment 1]



[0009] Figure 1 illustrates an image-processing apparatus according to a first embodiment, which is configured to process received image data defining an image of a retina so as to determine a location (or one of a number of likely locations) of one or more anatomical features of the imaged retina. The anatomical feature may be any anatomical feature of the retina towards which two or more blood vessels converge (e.g. the optic disc, a vascular junction or an overlap of blood vessels in the retina), or feature such as the fovea, which has a well-defined location with respect to a different anatomical feature (e.g. the optic disc) exhibiting vascular convergence. By way of an example, the anatomical feature of interest in the present embodiment is the optic disc.

[0010] The image-processing apparatus of the present embodiment forms part of a programmable signal processing apparatus, as illustrated schematically in Fig. 1. The signal processing apparatus 100 shown in Fig. 1 comprises an input/output (I/O) section 110 for receiving image data defining an image of the retina, and for outputting the determined location of the optic disc or a result of further processing operations based thereon. The signal processing apparatus 100 further comprises a processor 120, a working memory 130 and an instruction store 140 storing computer-readable instructions which, when executed by the processor 120, cause the processor 120 to perform the processing operations hereinafter described to determine the location of the optic disc in the retinal image. The I/O section 110, the working memory 130 and the instruction store 140 are communicably connected to the processor 120, as shown in Fig. 1. The instruction store 140 may comprise a ROM which is pre-loaded with the computer-readable instructions. Alternatively, the instruction store 140 may comprise a RAM or similar type of memory, and the computer-readable instructions can be input thereto from a computer program product, such as a computer-readable storage medium 150 such as a CD-ROM, etc. or a computer-readable signal 160 carrying the computer-readable instructions. The signal processing apparatus 100 may, for example, form part of a personal computer (PC) or a network server, and may comprise other components that are well-known (e.g. display, input device(s), etc.). In the present embodiment, the combination 170 of the hardware components shown in Fig. 1, comprising the processor 120, the working memory 130 and the instruction store 140, is configured to implement the functionality of the image-processing apparatus of the present embodiment, which will now be described in detail with reference to Figs. 2 to 6.

[0011] Figure 2 is a flow chart illustrating a first part of a process by which the image-processing apparatus 170 processes image data defining an image of a retina to determine the location of the optic disc, as an example of an anatomical landmark whose location in a retinal image is useful in a variety of different applications.

[0012] Firstly, in step S10, the processor 120 receives image data defining an image of the retina comprising a vascular structure via the I/O section 110. The acquired image data may define an retinal image which may be one of a number of different types that are known to those skilled in the art, for example a red and green (RG) reflectance image, an autofluorescence (AF) image, a fluorescein angiogram (FA) image, a reflectance scanning laser ophthalmoscope (SLO) image of the retina, or a colour fundus photograph of the retina, and may provide an ultra-wide-field view of the retina.

[0013] Before proceeding to step S20, the processor 120 may pre-process the received retinal image data to enhance the vasculature in the image, using techniques known to those skilled in the art, for example as described in chapter 8 of the book "Automated Image Detection of Retinal Pathology" by H. Jelinek and M. J. Cree (CRC Press, January 1, 2009). For example, the complex-valued Gabor kernel

may be used to generate convolved images, with the value of θ taking each value 2/N for a = 0... N - 1, where N is the number of orientations used in the evaluation which, may, for example, be between 8 and 12 (although reducing N to 6 in the pre-processing of RG images was not found to be detrimental). Values used for the other parameters in the Gabor kernel g may, for example, be:

γ = 0.37, λ = 7s, ψ = 0.

[0014] The parameter s sets the scale. The value of s = 2 was found to be well-suited for the pre-processing of Optomap™ images. However, the ideal value of s depends on the magnification of the image (or the real dimension in microns per pixel) and this is variable between imaging modalities and image formats. Also, a range of values of s may be used to account for vessels with a range of diameters. In this case, the resulting vessel enhancements at each value of s may be combined, for example, by pixel-by-pixel addition.

[0015] In this example, an array of complex-valued images is generated by convolution of the image by each Gabor kernel. Each element of this array is generated with one of the kernels:



[0016] At each pixel, the orientation index is found which gave the maximum absolute response:



[0017] This is the locally dominant (discretised) angle for pixels on a vessel, i.e. the orientation of the vessel.

[0018] The complex vessel-enhanced image, V, in which blood vessels appear brighter than the background, has a value at each pixel which is selected from one of the images in array U:



[0019] An example of a vessel-enhanced image generated by processing an Optomap™ image as shown in Fig. 3A in this way is illustrated in Fig. 3B.

[0020] A smoothed vessel image may, as in the present embodiment, then be generated by evaluating, twice, the maximum of the convolution by directional Gabor kernels. In this case, the input to the first execution is the vessel-enhanced image downsized by a factor of two, thus effectively doubling the size of the kernel. The input image to the second evaluation is the result of the first evaluation downsized by a factor of two. An example of a smoothed vessel-enhanced image generated in this way is illustrated in Fig. 3C.

[0021] Another form of linear filtering, non-linear filtering and/or normalisation may alternatively be applied at the pre-processing stage. For example, a low-pass filter may be used to reduce noise. Additionally or alternatively, a high-pass filter may be used to remove gradual image intensity changes such as lighting variations. A high-pass filter may also aid more consistent image normalisation. Normalisation is any method serving to reduce the variation between images of some parameter such as global mean brightness, global mean contrast or image histogram, or to reduce the variation across a single image of parameters such as local image brightness, local contrast or local histogram.

[0022] In step S20 in Fig. 2, the processor 120 calculates, for each of a plurality of pixels pij (where i and j represent the column and row of the pixel array, respectively) of the image which, as noted above, may have been processed to enhance the vasculature, a respective local orientation vector indicative of the orientation of any blood vessel present in the image at or adjacent the pixel (in other words, any blood vessel that may be present in the image at, or in the vicinity of (e.g. within a predetermined number of one or more pixels from), the pixel). Such a local orientation vector may be found in a number of different using techniques known to those skilled in the art. Local orientation and edge strength can be evaluated using first-order or second-order derivatives of image intensity, for example. First-order derivatives can be calculated by convolution with a Sobel or other kernel (as discussed above). Local orientation is then the arctangent of the ratio of the y-component and the x-component of the gradient. Second-order derivatives can be calculated by repeated convolution with a Sobel or other kernel. The local orientation is then the arctangent of the y-component and the x-component of the eigenvector with the largest eigenvalue of the Hessian matrix:

where z is image intensity.

[0023] For example, the local orientation vector may be calculated by finding the eigenvector of the Hessian matrix at each pixel that has the largest eigenvalue. Alternatively, the local orientation vector may, as in the present embodiment, take the form of the pixel value gradient vector, which tends to have a large magnitude in the boundary region of a blood vessel in the retinal image, where the pixel values change relatively rapidly one pixel to the next in the direction perpendicular to the local orientation of the blood vessel segment (i.e. along or opposite to the flow direction of the blood vessel; in other words, the direction along which the blood vessel extends in the image), and is generally small away from the blood vessels. The pixel value gradient vector at a pixel in a region of the retinal image showing a part of a blood vessel (particularly an edge portion of the blood vessel) is thus indicative of the orientation of the blood vessel.

[0024] Thus in the present embodiment, the processor 120 calculates in step S20, for each of the plurality of pixels pij of the smoothed image, a respective pixel value gradient vector g(i,j) (in other words, the gradient of image intensity) at the pixel. The pixels may be considered to hold respective values of a discrete function f(i,j) whose value varies with position (i,j) on the array of pixels, with the pixel value gradient vector g at any pixel pointing in the direction of greatest increase of the function at that pixel, the magnitude of the pixel value gradient vector, |g|, being the slope of the function in that direction. As noted above, in an image of a retina, the magnitude of the pixel value gradient vector will generally be highest at the edges of a blood vessel, where there is a rapid change in pixels values in the direction perpendicular to the direction along which the blood vessel extends.

[0025] Any well-known technique for calculating the pixel value gradient vector g may be used, for example as described in "Digital Image Processing" by R. C. Gonzalez and R. E. Woods (Pearson, August 31, 2007) or "Digital Image Processing and Computer Vision" by R. J. Schalkoff (John Wiley & Sons, September 2, 1992). Typical methods use Sobel, Prewitt or Roberts operators, which can be used to determine the gradient x- and y-components, the gradient magnitude and the gradient orientation. The plurality of pixels pij at which the pixel value gradient vector g(i,j) is calculated in step S20 may encompass all of the pixels of the smoothed image data, or only some of those pixels. The processor 120 may, for example, employ a mask defining one or more regions of the smoothed image (e.g. a peripheral region of the image), in which region(s) no calculation of the pixel value gradient vector g is to be performed.

[0026] The process may, as in the present embodiment, then proceed to an optional step S30, wherein the processor 120 selects pixels psij from the plurality of pixels pij such that the magnitude of the local orientation vector (in this embodiment, the calculated pixel value gradient vector g(i,j)) at each of the selected pixels psij exceeds a threshold. This optional step disqualifies pixels at which the local orientation vector (i.e. |g|) is small (i.e. pixels unlikely to be located within a blood vessel or at a blood vessel edge) from the further processing operations described below, thereby saving computational resources and, moreover, allowing the location of the center of optic disc to be estimated with higher accuracy. The processor 120 may, for example, select the pixels psij from the plurality of pixels pij such that the magnitude of the calculated pixel value gradient vector g(i,j) at each of the selected pixels psij exceeds a predetermined percentile of pixel value gradient vector magnitudes of the pixels pij. The predetermined percentile may be the 50th percentile, or it may, as in the present embodiment, more preferably be the 80th percentile.

[0027] Proceeding to step S40 in Fig. 4, the processor 120 calculates a normalised local orientation vector for each of the selected pixels psij by normalising the local orientation vector calculated for each of the selected pixels psij so that the magnitude of the normalised local orientation vector at each of the selected pixels psij takes a predetermined value that is the same for all of the selected pixels psij. Where step S30 is omitted, the processor 120 calculates in step S60 a normalised local orientation vector for each of the plurality of pixels pij by normalising the local orientation vector calculated for each of the pixels pij so that the magnitude of the normalised local orientation vector at each pixel takes a predetermined value that is the same for all of the pixels.

[0028] As the local orientation vector is, by way of example, the pixel value gradient vector in the present embodiment, the processor calculates in step S40 a normalised pixel value gradient vector gn(i,j) for each of the selected pixels psij by normalising the pixel value gradient vector g(i,j) calculated for each of the selected pixels psij so that the magnitude of the normalised pixel value gradient vector, |gn(i,j)|, at each of the selected pixels psij takes a predetermined value that is the same for all of the selected pixels psij. Thus, |gn(i,j)| is the same for all of the selected pixels psij, while the orientation of gn will depend on the local pixel value gradient at each selected pixel psij and will therefore generally differ between the selected pixels. In embodiments like the present, where the anatomical feature of the retina whose location is to be determined is the optic disc, the processor 120 preferably calculates the normalised pixel value gradient vector gn (or other form of local orientation vector, as noted above) for each of the selected pixels psij so that that the magnitude of said vector (i.e. R) at each of the selected pixels psij is between 0.4 DD and 2.5 DD, where DD is the diameter of the optic disc in the retinal image (a retinal distance unit widely used in ophthalmology), which may be expressed in terms of a number of pixels, for example. The value of DD may readily be determined by e.g. inspecting a retinal image, and subsequently used in the automatic processing of the same image or other images that have been obtained under similar conditions, using the techniques described herein. In other words, a suitable value of the unit DD to be used for automatically processing a batch of retinal images may be determined by manually examining a set of training images which are representative of the images in the batch to obtain a suitable measure of DD, which could then be used in the processing of the batch of images.

[0029] Where the image data received in step S10 defines an autofluorescence image of the retina or a fluorescein angiogram image of the retina, the processor 120 preferably calculates gn(i,j) (or, more generally, the aforementioned local orientation vector) for each of the selected pixels psij so that the magnitude of the calculated vector (i.e. R) at each of the selected pixels psij is between 0.8 DD and 2.4 DD. On the other hand, where the image data received in step S10 define a reflectance SLO image of the retina or a RG or colour fundus photograph of the retina, the processor 120 preferably calculates gn(i,j) (or, more generally, the aforementioned local orientation vector) for each of the selected pixels psij so that the magnitude of the calculated vector (R) at each of the selected pixels psij is between 0.4 DD and 1.2 DD.

[0030] In step S50, the processor 120 performs operations described hereinafter using an array of accumulators. Each accumulator, aij, of the array is configured to receive real values (which may or may not be integers, and may be positive or negative) and calculate an accumulated value that is indicative of an accumulation of the received values. For example, each accumulator aij may, as in the present embodiment, sum positive integer values (also referred to hereinafter as "votes") which it receives such that the accumulated value is a sum of the received values. Each accumulator aij may add each received value to a sum of any previously received values calculated thereby (in other words, the accumulator aij may update its accumulated value on receiving each new value), or it may temporarily store the received values before calculating their sum at the end of the operations in step S50. However, in other embodiments, each accumulator aij may alternatively be initialised to store an initial value which is a sufficiently large positive number, and add subsequently received negative integer values to the stored value (or deduct received positive values from the stored value), thereby effectively decrementing the stored value with each value it receives. In this case, the accumulated value at the end of the operations in step S50 is also indicative of an accumulation of the received values.

[0031] In the present embodiment, the array of accumulators is implemented by the processor 120 executing the aforementioned instructions in the instruction store 140 to appropriately address and manage (including writing to, reading from and otherwise processing information stored in) storage elements in the working memory 130. Each accumulator aij is associated with a respective pixel pij of the received image data, for example by the processor 120 in the working memory 130 an association, link or pointer relating each accumulator aij to a corresponding pixel pij in the image data.

[0032] In step S50, the processor 120 operates on the array of accumulators by adding, for each of the plurality of pixels, a respective value to an accumulator aij associated with a pixel poij of the plurality of pixels in the received image data that is offset (i.e. spaced apart) from said pixel by the predetermined value in a predetermined direction relative to the direction of the local orientation vector at the said pixel, and by smoothing the values in the accumulators aij. The process in step S50 will now be described in more detail with reference to Figs. 5 and 6.

[0033] Referring firstly to Fig. 5, in step S51, the processor initialises the array of accumulators so that the accumulated value stored in each accumulator aij is common constant value; in this embodiment, zero. In the aforementioned alternative embodiments, where the accumulators subtract received positive values from (or, equivalently, add received negative values to) an initial (starting) value, each accumulator aij is set to store the initial value in step S51.

[0034] In step S52, an index I, which is used to reference each of the selected pixels psij in turn, is set to an initial value of 1.

[0035] Then, in step S53, the processor 120 processes a first of the selected pixels psij (as I = 1 at this stage) by adding a value (which, by way of example is 1 in the present embodiment) to an accumulator aij of the accumulator array which is associated with a pixel poij of the received image data that is offset (i.e. spaced apart) from the selected pixel psij in the image data by the predetermined value R in a predetermined direction relative to the direction of the local orientation vector at the selected pixel. As the local orientation vector is the pixel value gradient vector g(i,j) in this embodiment, the predetermined direction is perpendicular to direction of the pixel value gradient vector g(i,j) at the selected pixel psij. In other embodiments, the predetermined direction may be perpendicular to the direction of the local orientation vector, or be at another predefined angle relative to the direction of the local orientation vector, depending on how the local orientation vector is calculated. The offset considered in step S53 will generally be in a direction along which the blood vessel illustrated by the selected pixel extends at the selected pixel, as illustrated in Fig. 6.

[0036] More particularly, Fig. 6 illustrates the relationship between pixels psij and poij of the retinal image data and an accumulator aij of the accumulator array A. In Fig. 6, the image data is illustrated in the upper part of the figure, and defines an image of a retina having an optic disc OD and various blood vessels BV which converge on the optic disc OD. The lower part of Fig. 6 illustrates part of the accumulator array A comprising the aforementioned accumulator aij. As illustrated in Fig. 6, pixel poij is taken to be a pixel which lies on a line that is perpendicular to the direction of g(i,j) at psij, and whose distance from psij along that line is closest to R (in other words, the pixel to which a vector R points, where R is a vector displacement in the image plane from psij calculated such that |R| = |gn(i,j)| = R and R · g(i,j) = 0). The processor 120 determines which of the pixels in the retinal image data has the calculated offset R (rounded to the nearest pixel) in relation to the first selected pixel psij on the basis of the normalised pixel value gradient vector gn(i,j) calculated for the first selected pixel psij in step S40. Unless the processor 120 employs a mask to exclude candidate locations for pixel poij, it may, as in the present embodiment, furthermore determine which of the pixels in the retinal image has an offset of -R (rounded to the nearest pixel) in relation to the first selected pixel psij, also on the basis of the normalised pixel value gradient vector gn(i,j) calculated for the first selected pixel psij in step S40.

[0037] By the process in step S53, the processor 120 effectively registers a "vote" with each of two accumulators of the accumulator array A corresponding to respective pixels (at R and -R relative to each pixel) that are taken to be a candidates for the location of the center of the optic disc.

[0038] The process then proceeds to step S54, wherein the processor 120 determines whether the index I has reached the limit L, L being the number of pixels psij having been selected in step S30. If the counter I has not yet reached the limit L, then the counter I is incremented by 1 in step S55, and the process loops back to step S53, where the next selected pixel psij is processed as described above. In this way, each pixel in the image data at which the pixel value gradient magnitude exceeds the threshold contributes a vote in an accumulator corresponding to a pixel that is offset from the aforementioned pixel by a distance R, in a direction which is normal to the local gradient and thus in a direction parallel to that of the blood vessel convergence.

[0039] Once all of the selected pixels psij have been processed (I = L in step S54), the process proceeds to step S56, wherein the processor 120 smooths the accumulated values in the accumulators to generate a response map. After a large number of selected pixels have been processed, a higher number of votes will have been registered in accumulators associated with pixels in the image towards which blood vessels converge, and this will be reflected in the response map. The array of accumulated values may be smoothed by applying any image-smoothing (or low-pass filtering) operation to the array. Examples of suitable image-smoothing methods are described in "Digital Image Processing" by R. C. Gonzalez and R. E. Woods (Pearson, August 31, 2007) and "Digital Image Processing and Computer Vision" by R. J. Schalkoff (John Wiley & Sons, September 2, 1992). The smoothing operation can be a linear operation such as convolution by a kernel which could have, for instance, a Gaussian shape or an array of similar values (to make a moving-average filter). Smoothing or low-pass filtering can also be performed in the frequency domain. The smoothing operation could also be a non-linear operation such as a median filter or a morphological filter. The processor 120 may, as in the present embodiment, smooth the values in the accumulators using a Gaussian kernel which preferably has a standard deviation of 0.1 DD to 0.5 DD, and more preferably 0.2 DD to 0.4 DD, and yet more preferably 0.3 DD. The smoothed values are indicative of the probability distribution for the location of the anatomical feature in the image of the retina (i.e. the center of the optic disc in the present embodiment).

[0040] Referring again to Fig. 4, after the completion of S50, the process proceeds to step S60, where the processor 120 determines the location of the optic disc in the image of the retina using the location of a pixel in the received image data which is associated with an accumulator of the array having accumulated an accumulated value which is a local or a global maximum among the accumulated values in the array of accumulators (the global maximum being the highest accumulated value of all the accumulators). More generally, the processor 120 may determine the location of the optic disc using the location of a pixel in the received image data which is associated with an accumulator having accumulated a value that is within a predetermined range of values below a local or a global maximum among the accumulated values in the array of accumulators (the range, where it is non-zero, spanning e.g. 5% or 10% of the maximum value). The locations of two or more local maxima determined from the response map may be used to determine the location(s) of the anatomical feature(s) of interest. For example, multiple maxima may be chosen by taking all maxima which are above a certain threshold. Multiple maxima would be found if the aim is to find anatomical features of interest having multiple instances (such as vessel junctions or crossovers), or if the maxima are considered as candidates for a feature of interest with a single instance (such as the optic disc) and a choice between these is to be deferred. A predetermined region of interest in the response map may be used to limit the choice of maxima. For example, the region of interest could be the region over which the feature to be detected is known from experience to be very likely to occur.

[0041] In some embodiments (for example, where the accumulated values are negative or where the accumulators decrement an initially set count with each received "vote"), the processor 120 may determine the location of the optic disc in the retinal image using the location of a pixel of the received image data which is associated with an accumulator of the array having accumulated an accumulated value which is a local or a global minimum among the accumulated values in the array of accumulators. Thus, the processor 120 may more generally determine the location of the optic disc in the image of the retina using the location of a pixel of the received image data which is associated with an accumulator of the array having accumulated an accumulated value which is a (local or global) extremum of the accumulated values in the array of accumulators.

[Embodiment 2]



[0042] In the present embodiment, the processor 120 operates not on one but on N accumulator arrays, where N ≥ 2. Each of these arrays of accumulators is associated with a respective one of N different quantised directions along which the pixel value gradient vector g may be oriented, and is labelled with a corresponding index n, where n = 1, ... N. The possible directions along which g may be oriented are taken to fall within the range of angles [0, π], and this range is quantised into N sub-ranges or "quantised directions". For example, where N = 4, any vector g(i,j) may be classified as being oriented along a first quantised direction corresponding to angular range [0, π/4), a second quantised direction corresponding to angular range [π/4, π/2), a third quantised direction corresponding to angular range [π/2, 3π/4), or a fourth quantised direction corresponding to angular range [3π/4, π]. The function round[Nmod(θ,π)/π - 0.5], where θ = arctan(gy/gx), gy is the y-component of g and gx is the x-component of g, may be used to quantise the direction of g. The accumulator array associated with the first quantised direction is assigned n = 1, the accumulator array associated with the second quantised direction is assigned n = 2, accumulator array associated with the third quantised direction is assigned n = 3, and accumulator array associated with the fourth quantised direction is assigned n = 4. In the present embodiment, an accumulator in the ith column and the jth row of the nth accumulator array (associated with the nth quantised direction) is denoted aijn. Each accumulator aijn in each accumulator array is associated with a respective pixel in the retinal image data, so that there are N accumulators (one in each of the N accumulator arrays) associated with each pixel in the image data.

[0043] Figure 7 is a flow diagram illustrating the process by which the image-processing apparatus of the present embodiment processes image data defining an image of a retina to determine the location of the optic disc, which comprises steps S10 to S30 and the exemplary form of step S40 of the first embodiment described above (which is labelled S40' in Fig. 7), whose description will not be repeated here, as well as a sequence of operations performed on the N arrays of accumulators, which will now be described. It should be noted that the structure and operation of the present embodiment is the same as the first embodiment, apart from the differences hereinafter described. Furthermore, the possible modifications to the first embodiment mentioned above may also be made to the present embodiment.

[0044] In step S71, each accumulator of the N arrays of accumulators is initialised in the same way as in step S51 of Fig. 5, so that the accumulated value stored in each accumulator aij is zero.

[0045] In step S72, an index I, which is used to reference each of the selected pixels psij in turn, is set to an initial value of 1.

[0046] Then, in step S73, the processor 120 processes a first of the selected pixels psij (as I = 1 at this stage) by quantising the direction of either g(i,j) or gn(i,j) at the first selected pixel psij into one of the N quantised directions. Thus, one of the values of n (from 1 to N) may be associated with the first selected pixel psij, depending on the direction of g at that pixel, and this allows the corresponding accumulator array (having the same value of n) to be identified and accessed.

[0047] In step S74, the processor 120 adds a constant value (e.g. 1, as in the first embodiment) to an accumulator aijn of an accumulator array that is associated with the same quantised direction as the quantised direction of g(i,j) at the first selected pixel psij, wherein the accumulator to which the value is added is associated with a pixel poij of the received image data that is (as in the first embodiment) offset from the first selected pixel psij by the predetermined value R in a direction perpendicular to the direction of g(i,j) at the first selected pixel psij. Thus, in the present embodiment, which accumulator receives the "vote" for the selected pixel depends also on the direction of g(i,j) at the selected pixel, with the accumulator associated with the same quantised direction as the quantised direction of g(i,j), and which is associated with a pixel poij of the received image data that is offset from the selected pixel by distance R in a direction perpendicular to the direction of g(i,j) at the selected pixel, receiving the vote.

[0048] The process then proceeds to step S75, wherein the processor 120 determined whether the index I has reached the limit L, L being the number of pixels psij having been selected in step S30. If the counter I has not yet reached the limit L, then the counter I is incremented by 1 in step S76, and the process loops back to step S73, where the next selected pixel psij is processed as described above. In this way, each selected pixel in the image data at which |g| exceeds the threshold contributes a vote in an accumulator aijn which is associated with the same quantised direction n as the quantised direction of g at the selected pixel, and which corresponds to a pixel that is offset from the selected pixel by a distance R, in a direction which is normal to the local gradient at the selected pixel and thus in a direction parallel to that of the blood vessel.

[0049] Once all of the selected pixels psij have been processed (I = L in step S75), the process proceeds to step S77, wherein the processor 120 smooths the accumulated values in the plurality of arrays of accumulators. The smoothing operation may be performed on accumulated values within one or more of the arrays, using the techniques described in connection with the first embodiment. Thus, Gaussian blurring of one or more accumulator arrays may be performed. Additionally or alternatively, the smoothing operation may be performed on values accumulated in the N accumulators aijn having the same indices i and j but different indices n (from 1 to N), using the same techniques.

[0050] In step S80, the processor 120 calculates, for each set of N accumulators aijn that are provided in different respective accumulator arrays and associated with the same pixel poij of the received image data, a product of the respective accumulated values that have been accumulated in the N accumulators aijn. The product is effectively a "soft AND" operation across the different gradient orientations, so that the pixels in the image associated with high values of this product tend to be where blood vessel converge from a range of different orientations. In other words, wherever this product is non-zero, there must have been a contribution from each of the N directions associated with the N accumulator arrays. In some applications (for example, where the anatomical feature of interest is the optic disc, towards which many different blood vessels converge), the use of multiple arrays of accumulators consequently may suppress some undesirable responses in the response map, such as vessel segments (where only two vessels converge) and vessel junctions (where only three or four vessels converge), allowing the feature of interest to be determined more easily in the subsequent data processing.

[0051] Figure 8A shows an Optomap™ green plane image of a retina, while Figs. 8B and 8C illustrate response maps generated for N = 1 and N = 8, respectively. These results demonstrate that the use of multiple (in this case, 8) accumulator arrays allows visually greater discrimination of the optic disc location compared to the case where only a single accumulator array is used. For a direct comparison of the map for N = 8 with the map for N = 1, the 8th root of the values in the map for N = 8 has been taken. However, taking the Nth root is not generally required because this does not affect the location of the highest peak. Figure 9A shows another Optomap™ image of a retina, in which the location of optic disc is more difficult to detect, as it is partially obscured by eye lashes. Figures 9B and 9C illustrate response maps obtained by processing the Optomap™ image in Fig. 9A using the above-described technique with N = 1 and N = 8, respectively. In this case, the peak response is at the correct location for N = 8 but at the wrong location for N = 1.

[0052] The number of arrays of accumulators, N, is preferably such that 2 ≤ N ≤ 32, and more preferably 4 ≤ N ≤ 16. If N = 1 (as in the first embodiment) then the method is simplified in that there is no need to quantise the gradient orientation and there is no need for the "soft AND" combination of accumulator arrays. However, this simplification comes at a cost of a decrease in "signal-to-noise ratio" by a factor of 2 as compared to the case where N = 8, for example. Using values of N that are too high, on the other hand, increase the computational cost and may not allow the location of the optic disc or other anatomical feature to be determined reliably, as the product calculated in step S80 may be zero for all pixel locations. For the detection of the optic disc, it is preferable that N = 8. For the detection of vessel junctions, a smaller value of N may be used, for example 3 or 4.

[0053] Referring again to Fig. 7, the process proceeds from step S80 to step S90, wherein the processor 120 determines the location of optic disc center in the image of the retina using the location of a pixel of the received image that is associated with a set of N accumulators, for which set the calculated product of accumulated values is a local or a global maximum among the calculated products of accumulated values. More generally, the processor 120 may determine the location of the optic disc using the location of a pixel of the received image data that is associated with a set of N accumulators, for which set the calculated product of accumulated values is within a predetermined range (which may be zero) of values below a (local or global) maximum among the calculated products of accumulated values. The locations of two or more local maxima determined from the response map may be used to determine the location(s) of the anatomical feature(s) of interest. For example, multiple maxima may be chosen by taking all maxima which are above a certain threshold. Multiple maxima would be found if the aim is to find anatomical features of interest having multiple instances (such as vessel junctions or crossovers), or if the maxima are considered as candidates for a feature of interest with a single instance (such as the optic disc) and a choice between these is to be deferred till after further processing. A predetermined region of interest in the response map may be used to limit the choice of maxima. For example, the region of interest could be the region over which the feature to be detected is known from experience to be very likely to occur.

[Modifications and Variations]



[0054] Many further modifications and variations can be made to the embodiments described above.

[0055] For example, the first embodiment described above with reference to Figs. 2 to 6 may, as illustrated in Fig. 10, be modified to include an additional step after the exemplary form of step S40 described above (which is labelled S40' in Fig. 10) and before step S52, namely an additional step S100, wherein the processor 120 calculates a second normalised pixel value gradient vector gn2(i,j) for each of the selected pixels psij by normalising the pixel value gradient vector g calculated for each of the selected pixels psij so that |gn2(i,j)| at each of the selected pixels psij takes a second common predetermined value R2. The first embodiment may, as also illustrated in Fig. 10, be further modified to include an additional step S120 that is performed for each value of index I, wherein the processor 120 adds, for the Ith selected pixel, a respective value to the accumulator which is associated with a pixel in the received image data that is offset from the Ith selected pixel by the second predetermined value R2 in the direction perpendicular to the direction of the pixel value gradient vector g at the Ith selected pixel. In other respects, this variant is the same as the first embodiment. The addition in step S120 may alternatively be made to a second accumulator layer that is associated with the second predetermined value R2, with the corresponding elements in the two accumulator arrays being combined after step S54, by some hard or soft OR operation such as element-by-element addition.

[0056] More generally, more than two normalised pixel value gradient vectors may be calculated prior to step S52, preferably each having a magnitude within the preferred ranges of |gn(i,j)| set out above. In these cases, the values may be accumulated in a single accumulator array, or in a corresponding number of accumulator arrays before being combined in the manner set out above. Using more than one calculated normalised pixel value gradient vector in the process may allow the location of the anatomical feature of interest such as the optic disc to be determined more reliably.

[0057] The above-described modifications may also be made to the second embodiment described above.

[0058] As a further modification, which may be made to either of the above-described embodiments, or combined with any of the above-described modifications to these embodiments, the selection of pixels in step S30 may be omitted so that the processor 120 calculates, each of a plurality of pixels pij of the received (and optically pre-processed) image data, a respective pixel value gradient vector g(i,j) at the pixel. In this variant, the processor 120 would then calculate gn(i,j) for each of the plurality of pixels pij by normalising the pixel value gradient vector calculated for each of the pixels pij so that the |gn(i,j)| at each of the pixels takes a common predetermined value R. The processor 120 would operate on the array of accumulators by adding, for each of the plurality of pixels pij, a respective value to an accumulator associated with a pixel poij of the received image data that is offset from the pixel of the plurality of pixels pij by the predetermined value R in a direction perpendicular to the direction of g(i,j) at the pixel of the plurality of pixels, and smoothing the values in the accumulators. In this variant, the processor may add, as the respective value, a respective weighting to the accumulator associated with the pixel poij, the weighting being indicative of the magnitude of g(i,j) at the pixel of the plurality of pixels. The processor would then determine the location of the anatomical feature (e.g. optic disc) in the image of the retina using the location of a pixel of the received image data which is associated with an accumulator having accumulated an accumulated value which is within a predetermined range of an extremum of the accumulated values in the accumulators, as described in more detail above. In this variant, there is a contribution from each of the plurality of pixels to the accumulator, although the size of this contribution from each pixel will depend on the magnitude of g(i,j) at that pixel.

[0059] Furthermore, the ordering of some of the method steps in the embodiments and modifications thereof described above may be varied. For example, the ordering of step S40' and S100, and similarly S53 and S120, in Fig. 10 may be reversed, or these steps may be performed in parallel. Furthermore, the initialisation of accumulator array A may be performed at any point in the process before step S53 in the process illustrated in Figs. 4, 5 and 10, and at any point before step S73 in the process illustrated in Fig. 7 (the initialisation of course being outside the loop over index I).


Claims

1. A non-transitory computer-readable storage medium storing computer program instructions which, when executed by a processor, cause the processor to process image data defining an image of a retina to determine a location of an anatomical feature of the retina in the image by:

receiving the image data;

calculating, for each of a plurality of pixels of the received image data, a respective local orientation vector indicative of the orientation of any blood vessel present in the image at or adjacent the pixel;

calculating a normalised local orientation vector for each of the plurality of pixels by normalising the local orientation vector calculated for each of the pixels so that the magnitude of the normalised local orientation vector at each of the pixels takes a common predetermined value;

operating on an array of accumulators, wherein each accumulator in the array is associated with a respective pixel of the received image data, by:

for each of the plurality of pixels, adding a respective value to an accumulator associated with a pixel of the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in a predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels; and

smoothing the values in the accumulators; and

determining the location of the anatomical feature in the image of the retina using the location of a pixel of the received image data which is associated with an accumulator having accumulated an accumulated value, which accumulated value is within a predetermined range of an extremum of the accumulated values in the accumulators.


 
2. A non-transitory computer-readable storage medium according to claim 1, wherein the instructions, when executed by the processor, cause the processor to operate on a plurality, N, of arrays of accumulators, where N is an integer equal to at least 2, and wherein:

each of the arrays of accumulators is associated with a respective one of N quantised directions for the local orientation vector;

each of the accumulators in each of the N arrays is associated with a respective pixel in the received image data;

the processor operates on the plurality of arrays of accumulators by:

for each of the plurality of pixels:

quantising the direction of the calculated local orientation vector at the pixel into one of the N quantised directions; and

adding a respective value to an accumulator of an accumulator array that is associated with the same quantised direction as the quantised direction of the local orientation vector at the pixel, the accumulator being associated with a pixel of the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in the predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels; and

smoothing the values in the accumulators of the plurality of arrays;

the processor further calculates, for each set of N accumulators that are provided in different respective accumulator arrays and associated with the same pixel of the received image data, a product of the respective accumulated values that have been accumulated in the N accumulators; and

the processor determines the location of the anatomical feature in the image of the retina using the location of a pixel of the received image data associated with at a set of N accumulators, for which set the calculated product of accumulated values is within a predetermined range of an extremum of the calculated products of accumulated values.


 
3. A non-transitory computer-readable storage medium according to claim 2, wherein 2 ≤ N ≤ 32, and more preferably 4 ≤ N ≤ 16.
 
4. A non-transitory computer-readable storage medium according to any preceding claim, wherein the instructions, when executed by the processor, further cause the processor to calculate a second normalised local orientation vector for each of the plurality of pixels by normalising the local orientation vector calculated for each of the plurality of pixels so that the magnitude of the second normalised local orientation vector at each of the pixels takes a second common predetermined value, and to further operate on each array of accumulators by:
for each pixel of the plurality of pixels, adding a respective value to an accumulator associated with a pixel in the received image data that is offset from the pixel of the plurality of pixels by the second predetermined value in the predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels.
 
5. A non-transitory computer-readable storage medium according to any preceding claim, wherein the instructions, when executed by the processor, cause the processor to process the image data defining the image of the retina to determine the location of the anatomical feature of the retina in the image by performing a further process of:
selecting pixels from the plurality of pixels such that the magnitude of the calculated local orientation vector at each of the selected pixels exceeds a threshold, and by:

calculating the normalised local orientation vector for each of the selected pixels by normalising the local orientation vector calculated for each of the selected pixels so that the magnitude of the normalised local orientation vector at each of the selected pixels takes a common predetermined value; and

operating on each array of accumulators by:
for each of the selected pixels, adding a respective value to an accumulator associated with a pixel of the received image data that is offset from the selected pixel in the image by the predetermined value in the predetermined direction relative to the direction of the local orientation vector at the selected pixel.


 
6. A non-transitory computer-readable storage medium according to claim 5, wherein the instructions, when executed by the processor, cause the processor to select the pixels from the plurality of pixels such that the magnitude of the calculated local orientation vector at each of the selected pixels exceeds the 50th percentile, and more preferably the 80th percentile, of local orientation vector magnitudes of the pixels in the image data.
 
7. A non-transitory computer-readable storage medium according to any preceding claim, wherein the instructions, when executed by the processor, cause the processor to operate on each array of accumulators by:
for each pixel of the plurality of pixels, adding, as the respective value, a respective weighting to the accumulator associated with a pixel in the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in the predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels, the weighting being indicative of the magnitude of the local orientation vector at the pixel of the plurality of pixels.
 
8. A non-transitory computer-readable storage medium according to any preceding claim, wherein the anatomical feature of the retina comprises at least one of an optic disc, a vascular junction connecting at least two blood vessels, an overlap of blood vessels of the retina, and a fovea.
 
9. A non-transitory computer-readable storage medium according to claim 8, wherein the anatomical feature of the retina comprises the optic disc, and the instructions, when executed by the processor, cause the processor to calculate the normalised local orientation vector for each pixel of the plurality of pixels so that that the magnitude of the normalised local orientation vector at each pixel of the plurality of pixels is between 0.5 DD and 2.5 DD, where DD is the diameter of the optic disc.
 
10. A non-transitory computer-readable storage medium according to claim 9, wherein the image is one of an autofluorescence image of the retina and a fluorescein angiogram image of the retina, and the instructions, when executed by the processor, cause the processor to calculate the normalised local orientation vector for each pixel of the plurality of pixels so that the magnitude of the normalised local orientation vector at each pixel of the plurality of pixels is 0.8 DD to 2.4 DD.
 
11. A non-transitory computer-readable storage medium according to claim 9, wherein the image is one of a reflectance scanning laser ophthalmoscope image of the retina and a fundus photograph of the retina, and the instructions, when executed by the processor, cause the processor to calculate the normalised local orientation vector for each pixel of the plurality of pixels so that the magnitude of the normalised local orientation vector at each pixel of the plurality of pixels is 0.4 DD to 1.2 DD.
 
12. A non-transitory computer-readable storage medium according to claim 8, wherein the anatomical feature of the retina comprises the optic disc, and the instructions, when executed by the processor, cause the processor to smooth the values in the accumulators using a kernel having a standard deviation of 0.1 DD to 0.5 DD, and more preferably 0.2 DD to 0.4 DD, where DD is the diameter of the optic disc.
 
13. An image-processing apparatus comprising a processor and a memory, the memory storing instructions executable by the processor whereby the processor is operative to process image data defining an image of a retina to determine a location of an anatomical feature of the retina in the image by:

calculating, for each of a plurality of pixels of the received image data, a respective local orientation vector indicative of the orientation of any blood vessel present in the image at or adjacent the pixel;

calculating a normalised local orientation vector for each of the plurality of pixels by normalising the local orientation vector calculated for each of the pixels so that the magnitude of the normalised local orientation vector at each of the pixels takes a common predetermined value;

operating on an array of accumulators, wherein each accumulator in the array is associated with a respective pixel of the received image data, by:

for each of the plurality of pixels, adding a respective value to an accumulator associated with a pixel of the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in a predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels; and

smoothing the values in the accumulators; and

determining the location of the anatomical feature in the image of the retina using the location of a pixel of the received image data which is associated with an accumulator having accumulated an accumulated value, which accumulated value is within a predetermined range of an extremum of the accumulated values in the accumulators.


 
14. An apparatus according to claim 13, wherein the instructions, when executed by the processor, cause the processor to operate on a plurality, N, of arrays of accumulators, where N is an integer equal to at least 2, and wherein:

each of the arrays of accumulators is associated with a respective one of N quantised directions for the local orientation vector;

each of the accumulators in each of the N arrays is associated with a respective one of the pixels in the received image data;

the instructions, when executed by the processor, cause the processor to:

operate on the plurality of arrays of accumulators by:
for each of the plurality of pixels:

quantising the direction of the calculated local orientation vector at the pixel into one of the N quantised directions; and

adding a respective value to an accumulator of an accumulator array that is associated with the same quantised direction as the quantised direction of the local orientation vector at the pixel, the accumulator being associated with a pixel of the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in the predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels; and

smoothing the values in the accumulators of the plurality of arrays;

further calculate, for each set of N accumulators that are provided in different respective accumulator arrays but are associated with the same pixel of the received image data, a product of the respective accumulated values that have been accumulated in the N accumulators; and

determine the location of the anatomical feature in the image of the retina using the location of a pixel of the received image data associated with one of the sets of N accumulators, for which set of accumulators the calculated product of accumulated values is within a predetermined range of an extremum of the calculated products of accumulated values.


 
15. An apparatus according to claim 14, wherein 2 ≤ N ≤ 32, and more preferably 4 ≤ N ≤ 16.
 
16. An apparatus according to any of claims 13 to 15, wherein the instructions, when executed by the processor, further cause the processor to calculate a second normalised local orientation vector for each of the plurality of pixels by normalising the local orientation vector calculated for each of the plurality of pixels so that the magnitude of the second normalised local orientation vector at each of the pixels takes a second common predetermined value, and to further operate on each array of accumulators by:
for each pixel of the plurality of pixels, adding a respective value to an accumulator associated with a pixel in the received image data that is offset from the pixel of the plurality of pixels by the second predetermined value in the predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels.
 
17. An apparatus according to any of claims 13 to 16, wherein the instructions, when executed by the processor, cause the processor to process the image data defining the image of the retina to determine the location of the anatomical feature of the retina in the image by performing a further process of:
selecting pixels from the plurality of pixels such that the magnitude of the calculated local orientation vector at each of the selected pixels exceeds a threshold, and by:

calculating the normalised local orientation vector for each of the selected pixels by normalising the local orientation vector calculated for each of the selected pixels so that the magnitude of the normalised local orientation vector at each of the selected pixels takes a common predetermined value; and

operating on each array of accumulators by:
for each of the selected pixels, adding a respective value to an accumulator associated with a pixel of the received image data that is offset from the selected pixel in the image by the predetermined value in the predetermined direction relative to the direction of the local orientation vector at the selected pixel.


 
18. An apparatus according to claim 17, wherein the instructions, when executed by the processor, cause the processor to select the pixels from the plurality of pixels by selecting pixels from the plurality of pixels such that the magnitude of the calculated local orientation vector at each of the selected pixels exceeds the 50th percentile, and more preferably the 80th percentile, of local orientation vector magnitudes of the pixels in the image.
 
19. An apparatus according to any of claims 13 to 18, wherein the instructions, when executed by the processor, cause the processor to operate on each array of accumulators by:
for each pixel of the plurality of pixels, adding, as the respective value, a respective weighting to the accumulator associated with a pixel in the received image data that is offset from the pixel of the plurality of pixels by the predetermined value in the predetermined direction relative to the direction of the local orientation vector at the pixel of the plurality of pixels, the weighting being indicative of the magnitude of the local orientation vector at the pixel of the plurality of pixels.
 
20. An apparatus according to any of claims 13 to 19, wherein the anatomical feature of the retina comprises at least one of an optic disc, a vascular junction connecting at least two blood vessels, an overlap of blood vessels of the retina, and a fovea.
 
21. An apparatus according to claim 20, wherein the anatomical feature of the retina comprises the optic disc, and the instructions, when executed by the processor, cause the processor to calculate the normalised local orientation vector for each pixel of the plurality of pixels so that the magnitude of the normalised local orientation vector at each pixel of the plurality of pixels is between 0.5 DD and 2.5 DD, where DD is the diameter of the optic disc.
 
22. An apparatus according to claim 21, wherein the image is one of an autofluorescence image of the retina and a fluorescein angiogram image of the retina, and the instructions, when executed by the processor, cause the processor to calculate the normalised local orientation vector for each pixel of the plurality of pixels so that the magnitude of the normalised local orientation vector at each pixel of the plurality of pixels is 0.8 DD to 2.4 DD.
 
23. An apparatus according to claim 21, wherein the image is one of a reflectance scanning laser ophthalmoscope image of the retina and a fundus photograph of the retina, and the instructions, when executed by the processor, cause the processor to calculate the normalised local orientation vector for each pixel of the plurality of pixels so that the magnitude of the normalised local orientation vector at each pixel of the plurality of pixels is 0.4 DD to 1.2 DD.
 
24. An apparatus according to claim 20, wherein the anatomical feature of the retina comprises the optic disc, and the instructions, when executed by the processor, cause the processor to smooth the values in the accumulators using a kernel having a standard deviation of 0.1 DD to 0.5 DD, and more preferably 0.2 DD to 0.4 DD, where DD is the diameter of the optic disc.
 


Ansprüche

1. Nicht-transitorisches, computerlesbares Speichermedium, das Computerprogramm-Anweisungen speichert, die bei Ausführung durch einen Prozessor den Prozessor veranlassen, Bilddaten zu prozessieren, die ein Bild einer Retina definieren, um einen Ort eines anatomischen Merkmals der Retina im Bild zu bestimmen, durch:

Empfangen der Bilddaten;

Berechnen, für jedes einer Vielzahl von Pixeln der empfangenen Bilddaten, eines entsprechenden lokalen Orientierungsvektors, der für die Orientierung eines Blutgefäßes indikativ ist, das im Bild an oder angrenzend dem Pixel vorhanden ist;

Berechnen eines normalisierten lokalen Orientierungsvektors für jedes der Vielzahl von Pixeln durch Normalisieren des für jedes der Pixel berechneten Lokalorientierungsvektors, so dass die Größe des normalisierten Lokalorientierungsvektors an jedem der Pixel einen gemeinsamen, vorbestimmten Wert annimmt;

Arbeiten, an einem Feld von Akkumulatoren, wobei jeder Akkumulator im Feld mit einem entsprechenden Pixel der empfangenen Bilddaten assoziiert ist, durch:

für jedes der Vielzahl von Pixeln, Addieren eines entsprechenden Wertes zu einem Akkumulator, der mit einem Pixel der empfangenen Bilddaten assoziiert ist, das gegenüber dem Pixel der Vielzahl von Pixeln um einen vorbestimmten Wert in einer vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors beim Pixel der Vielzahl von Pixeln versetzt ist; und

Glätten der Werte in den Akkumulatoren; und

Bestimmen des Orts des anatomischen Merkmals im Bild der Retina unter Verwendung des Orts eines Pixels der empfangenen Bilddaten, welches mit einem Akkumulator assoziiert ist, der einen akkumulierten Wert akkumuliert hat, welcher akkumulierte Wert innerhalb eines vorbestimmten Bereichs eines Extrems der akkumulierten Werte in den Akkumulatoren ist.


 
2. Nicht-transitorisches computerlesbares Speichermedium gemäß Anspruch 1, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, an einer Vielzahl, N, von Feldern von Akkumulatoren zu arbeiten, wobei N eine Ganzzahl zumindest gleich 2 ist, und wobei:

jedes der Felder von Akkumulatoren mit einer entsprechenden von N quantisierten Richtungen für den Lokalorientierungsvektor assoziiert ist;

jeder der Akkumulatoren in jedem der N Felder mit einem entsprechenden Pixel in den empfangenen Bilddaten assoziiert ist;

der Prozessor an der Vielzahl von Feldern von Akkumulatoren arbeitet, durch:
für jedes der Vielzahl von Pixeln:

Quantisieren der Richtung des berechneten Lokalorientierungsvektors bei dem Pixel in eine der N quantisierten Richtungen; und

Addieren eines entsprechenden Wertes zum Akkumulator eines Akkumulatorfelds, welches mit derselben quantisierten Richtung wie die quantisierte Richtung des Lokalorientierungsvektors an dem Pixel assoziiert ist, wobei der Akkumulator mit einem Pixel der empfangenen Daten assoziiert ist, die gegenüber dem Pixel der Vielzahl von Pixeln um den vorbestimmten Wert in der vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors beim Pixel der Vielzahl von Pixeln versetzt ist; und

Glätten der Werte in den Akkumulatoren der Vielzahl von Feldern;

der Prozessor weiter für jeden Satz von N Akkumulatoren, die an unterschiedlichen jeweiligen Akkumulatorfeldern vorgesehen und mit demselben Pixel der empfangenen Bilddaten assoziiert sind, ein Produkt der entsprechenden akkumulierten Werte, die in den N Akkumulatoren akkumuliert worden sind, berechnet; und

der Prozessor den Ort des anatomischen Merkmals im Bild der Retina unter Verwendung des Orts eines Pixels der empfangenen Bilddaten, welche mit einem Satz von N Akkumulatoren assoziiert sind, für welche das berechnete Produkt akkumulierter Werte innerhalb eines vorbestimmten Bereichs eines Extrems der berechneten Produkte von akkumulierten Werten liegt, bestimmt.


 
3. Nicht-transitorisches, computerlesbares Speichermedium gemäß Anspruch 2, wobei 2 ≤ N ≤ 32 und bevorzugterer Weise 4 ≤ N ≤ 16.
 
4. Nicht-transitorisches, computerlesbares Speichermedium gemäß einem vorstehenden Anspruch, wobei die Anweisungen bei Ausführung durch den Prozessor weiter den Prozessor veranlassen, einen zweiten normalisierten Lokalorientierungsvektor für jedes der Vielzahl von Pixeln zu berechnen, durch Normalisieren des Lokalorientierungsvektors, der für jedes der Vielzahl von Pixeln berechnet wird, so dass die Größe des zweiten normalisierten Lokalorientierungsvektors an jedem der Pixel einen zweiten gemeinsamen vorbestimmten Wert annimmt, und weiter an jedem Feld von Akkumulatoren zu arbeiten, durch:
für jedes Pixel der Vielzahl von Pixeln, Addieren eines entsprechenden Werts zu einem mit einem Pixel in den empfangenen Bilddaten assoziierten Akkumulators, das gegenüber dem Pixel der Vielzahl von Pixeln um den zweiten vorbestimmten Wert in der vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors versetzt ist, an dem Pixel der Vielzahl von Pixeln.
 
5. Nicht-transitorisches computerlesbares Speichermedium gemäß einem vorstehenden Anspruch, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, die Bilddaten zu prozessieren, die das Bild der Retina definieren, um den Ort des anatomischen Merkmals der Retina im Bild zu bestimmen, durch Durchführen eines weiteren Prozesses von:
Auswählen von Pixeln aus der Vielzahl von Pixeln, so dass die Größe des berechneten Lokalorientierungsvektors an jedem der ausgewählten Pixel einen Schwellenwert übersteigt, und durch:

Berechnen des normalisierten Lokalorientierungsvektors für jedes der ausgewählten Pixel durch Normalisieren des Lokalorientierungsvektors, der für jedes der ausgewählten Pixel berechnet wird, so dass die Größe des normalisierten Lokalorientierungsvektors an jedem der ausgewählten Pixel einen gemeinsamen vorbestimmten Wert annimmt; und

Arbeiten an jedem Feld von Akkumulatoren durch:
für jedes der ausgewählten Pixel, Addieren eines entsprechenden Werts zu einem mit einem Pixel der empfangenen Bilddaten assoziierten Akkumulator, das gegenüber dem ausgewählten Pixel im Bild um den vorbestimmten Wert in der vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors an dem ausgewählten Pixel versetzt ist.


 
6. Nicht-transitorisches computerlesbares Speichermedium gemäß Anspruch 5, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, die Pixel aus der Vielzahl von Pixeln so auszuwählen, dass die Größe des berechneten Lokalorientierungsvektors an jedem der ausgewählten Pixel das 50-ste Percentil, und bevorzugterer das 80-ste Percentil von Lokalorientierungsvektor-Größen der Pixel in den Bilddaten übersteigt.
 
7. Nicht-transitorisches, computerlesbares Speichermedium gemäß einem vorstehenden Anspruch, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, an jedem Feld von Akkumulatoren zu arbeiten, durch:
für jedes Pixel der Vielzahl von Pixeln, Addieren, als dem entsprechenden Wert, einer entsprechenden Gewichtung zum Akkumulator, der mit einem Pixel in den empfangenen Bilddaten assoziiert ist, das gegenüber dem Pixel der Vielzahl von Pixeln um den vorbestimmten Wert in der vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors am Pixel der Vielzahl von Pixeln versetzt ist, wobei die Gewichtung für die Größe des Lokalorientierungsvektors an dem Pixel der Vielzahl von Pixeln indikativ ist.
 
8. Nicht-transitorisches computerlesbares Speichermedium gemäß einem vorstehenden Anspruch, wobei das anatomische Merkmal der Retina ein Sehnervenkopf, eine Gefäßverbindung, die zumindest zwei Blutgefäße verbindet, eine Überlappung von Blutgefäßen der Retina und eine Fovea umfasst.
 
9. Nicht-transitorisches computerlesbares Speichermedium gemäß Anspruch 8, wobei das anatomische Merkmal der Retina den Sehnervenkopf umfasst, und die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, den normalisierten Lokalorientierungsvektor für jedes Pixel der Vielzahl von Pixeln so zu berechnen, dass die Größe des normalisierten Lokalorientierungsvektors an jedem Pixel der Vielzahl von Pixeln zwischen 0,5 DD und 2,5 DD ist, wobei DD der Durchmesser des Sehnervenkopfs ist.
 
10. Nicht-transitorisches computerlesbares Speichermedium gemäß Anspruch 9, wobei das Bild ein Autofluoreszenz-Bild der Retina oder ein Fluoreszin-Angiogramm-Bild der Retina ist und die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, den normalisierten Lokalorientierungsvektor für jedes Pixel der Vielzahl von Pixeln so zu berechnen, dass die Größe des normalisierten Lokalorientierungsvektors an jedem Pixel der Vielzahl von Pixeln 0,8 DD bis 2,4 DD beträgt.
 
11. Nicht-transitorisches, computerlesbares Speichermedium gemäß Anspruch 9, wobei das Bild entweder ein Reflektanz-Scanlaser-Ophthalmoskopbild der Retina ist oder ein Fundusfoto der Retina ist und die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, den normalisierten Lokalorientierungsvektor für jedes Pixel der Vielzahl von Pixeln so zu berechnen, dass die Größe des normalisierten Lokalorientierungsvektors an jedem Pixel der Vielzahl von Pixeln 0,4 DD bis 1,2 DD beträgt.
 
12. Nicht-transitorisches computerlesbares Speichermedium gemäß Anspruch 8, wobei das anatomische Merkmal der Retina den Sehnervenkopf umfasst, und die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, die Werte in den Akkumulatoren unter Verwendung eines Kernels mit einer Standardabweichung von 0,1 DD bis 0,5 DD und bevorzugterer 0,2 DD bis 0,4 DD zu glätten, wobei DD der Durchmesser des Sehnervenkopfs ist.
 
13. Bildverarbeitungs-Einrichtung, die einen Prozessor und einen Speicher umfasst, wobei der Speicher Anweisungen speichert, die durch den Prozessor ausführbar sind, wodurch der Prozessor operativ ist, Bilddaten zu prozessieren, die ein Bild einer Retina definieren, um einen Ort eines anatomischen Merkmals der Retina im Bild zu bestimmen, durch:

Berechnen, für jedes der Vielzahl von Pixeln der empfangenen Bilddaten, eines entsprechenden Lokalorientierungsvektors, der für die Orientierung eines Blutgefäßes indikativ ist, das im Bild an oder angrenzend dem Pixel vorhanden ist;

Berechnen eines normalisierten Lokalorientierungsvektors für jedes der Vielzahl von Pixeln durch Normalisieren des für jedes der Pixel berechneten Lokalorientierungsvektors, so dass die Größe des normalisierten Lokalorientierungsvektors an jedem der Pixel einen gemeinsamen vorbestimmten Wert annimmt;

Arbeiten an einem Feld von Akkumulatoren, wobei jeder Akkumulator in dem Feld mit dem entsprechenden Pixel der empfangenen Bilddaten assoziiert ist, durch:

für jedes der Vielzahl von Pixeln, Addieren eines entsprechenden Wertes zu einem mit einem Pixel der empfangenen Bilddaten assoziierten Akkumulator, das gegenüber dem Pixel der Vielzahl von Pixeln um den vorbestimmten Wert in einer vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors am Pixel der Vielzahl von Pixeln versetzt ist; und

Glätten der Werte in den Akkumulatoren; und

Bestimmen des Orts des anatomischen Merkmals im Bild der Retina unter Verwendung des Orts eines Pixels der empfangenen Bilddaten, welches mit einem Akkumulator assoziiert ist, der einen akkumulierten Wert akkumuliert hat, welcher akkumulierte Wert innerhalb eines vorbestimmten Bereichs eines Extrems der akkumulierten Werte in den Akkumulatoren ist.


 
14. Einrichtung gemäß Anspruch 13, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, an einer Vielzahl N von Feldern von Akkumulatoren zu arbeiten, wobei N eine Ganzzahl gleich zumindest 2 ist, und wobei:

jedes der Felder von Akkumulatoren mit einer entsprechenden von N quantisierten Richtungen für den Lokalorientierungsvektor assoziiert ist;

jeder der Akkumulatoren in jedem der N Felder mit einem entsprechenden der Pixel in den empfangenen Bilddaten assoziiert ist;

die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen:
an der Vielzahl von Feldern von Akkumulatoren zu arbeiten, durch:
für jedes der Vielzahl von Pixeln:

Quantisieren der Richtung des berechneten Lokalorientierungsvektors am Pixel in einer von N quantisierten Richtungen; und

Addieren eines entsprechenden Werts zu einem Akkumulator eines Akkumulator-Felds, das mit derselben quantisierten Richtung wie die quantisierte Richtung des Lokalorientierungsvektors am Pixel assoziiert ist, wobei der Akkumulator mit einem Pixel der empfangenen Bilddaten assoziiert ist, das gegenüber dem Pixel der Vielzahl von Pixeln um den vorbestimmten Wert in der vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors am Pixel der Vielzahl von Pixeln versetzt ist; und

Glätten der Werte in den Akkumulatoren der Vielzahl von Feldern;

weiter für jeden Satz von N Akkumulatoren, die in unterschiedlichen jeweiligen Akkumulator-Feldern vorgesehen sind, aber mit demselben Pixel der empfangenen Bilddaten assoziiert sind, ein Produkt der jeweiligen akkumulierten Werte zu berechnen, die in den N Akkumulatoren akkumuliert worden sind; und

den Ort des anatomischen Merkmals im Bild der Retina unter Verwendung des Orts eines Pixels der empfangenen Bilddaten zu bestimmen, die mit einem der Sätze von N Akkumulatoren assoziiert sind, für welche ein Satz von Akkumulatoren das berechnete Produkt akkumulierter Werte innerhalb eines vorbestimmten Bereichs eines Extrems der berechneten Produkte akkumulierter Werte ist.


 
15. Einrichtung gemäß Anspruch 14, wobei 2 ≤ N ≤ 32 und bevorzugterer Weise 4 ≤ N ≤ 16.
 
16. Einrichtung gemäß einem von Ansprüchen 13 bis 15, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor weiter veranlassen, einen zweiten normalisierten Lokalorientierungsvektor für jedes der Vielzahl von Pixeln zu berechnen, durch Normalisieren des für jedes der Vielzahl von Pixeln berechneten Lokalorientierungsvektors, so dass die Größe des zweiten normalisierten Lokalorientierungsvektors an jedem der Pixel einen zweiten gemeinsamen vorbestimmten Wert annimmt, und weiter an jedem Feld von Akkumulatoren zu arbeiten, durch:
für jedes Pixel der Vielzahl von Pixeln, Addieren eines entsprechenden Werts eines mit einem Pixel in den empfangenen Bilddaten assoziierten Akkumulators, das gegenüber dem Pixel der Vielzahl von Pixeln um den zweiten vorbestimmten Wert in der vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors an dem Pixel der Vielzahl von Pixeln versetzt ist.
 
17. Einrichtung gemäß einem der Ansprüche 13 bis 16, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, die das Bild der Retina definierenden Bilddaten zu prozessieren, um den Ort des anatomischen Merkmals der Retina im Bild zu bestimmen, durch Durchführen eines weiteren Prozesses von:
Auswählen von Pixeln aus der Vielzahl von Pixeln, so dass die Größe des berechneten Lokalorientierungsvektors an jedem der ausgewählten Pixel einen Schwellenwert übersteigt, und durch:

Berechnen des normalisierten Lokalorientierungsvektors für jedes der ausgewählten Pixel durch Normalisieren des für jedes der ausgewählten Pixel berechneten Lokalorientierungsvektors, so dass die Größe des normalisierten Lokalorientierungsvektors an jedem der ausgewählten Pixel einen gemeinsamen vorbestimmten Wert annimmt; und

Arbeiten an jedem Feld von Akkumulatoren durch:
für jedes der ausgewählten Pixel, Addieren eines entsprechenden Werts zu einem mit einem Pixel der empfangenen Bilddaten assoziierten Akkumulator, das gegenüber dem ausgewählten Pixel im Bild um den vorbestimmten Wert in der vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors an dem ausgewählten Pixel versetzt ist.


 
18. Einrichtung gemäß Anspruch 17, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, die Pixel aus der Vielzahl von Pixeln auszuwählen, durch Auswählen von Pixeln aus der Vielzahl von Pixeln so, dass die Größe des berechneten Lokalorientierungsvektors an jedem der ausgewählten Pixel das 50-ste Perzentil und bevorzugterer Weise das 80-ste Perzentil der Lokalorientierungsvektor-Größen der Pixel im Bild übersteigt.
 
19. Einrichtung gemäß einem der Ansprüche 13 bis 18, wobei die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, an jedem Feld von Akkumulatoren zu arbeiten, durch:
für jedes Pixel der Vielzahl von Pixeln, Addieren, als den entsprechenden Wert, einer entsprechenden Gewichtung zum mit einem Pixel in empfangenen Bilddaten assoziierten Akkumulator, das gegenüber dem Pixel der Vielzahl von Pixeln um den vorbestimmten Wert in der vorbestimmten Richtung relativ zur Richtung des Lokalorientierungsvektors am Pixel der Vielzahl von Pixeln versetzt ist, wobei die Gewichtung für die Größe des Lokalorientierungsvektors am Pixel der Vielzahl von Pixeln indikativ ist.
 
20. Einrichtung gemäß einem der Ansprüche 13 bis 19, wobei das anatomische Merkmal der Retina einen Sehnervenkopf, eine Gefäßverbindung, die zumindest zwei Blutgefäße verbindet, eine Überlappung von Blutgefäßen der Retina, und eine Fovea umfasst.
 
21. Einrichtung gemäß Anspruch 20, wobei das anatomische Merkmal der Retina den Sehnervenkopf umfasst, und die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, den normalisierten Lokalorientierungsvektor für jedes Pixel der Vielzahl von Pixeln so zu berechnen, dass die Größe des normalisierten Lokalorientierungsvektors an jedem Pixel der Vielzahl von Pixeln zwischen 0,5 DD und 2,5 DD ist, wobei DD der Durchmesser des Sehnervenkopfs ist.
 
22. Einrichtung gemäß Anspruch 21, wobei das Bild ein Autofluoreszenz-Bild der Retina oder ein Fluoreszin-Angiogramm-Bild der Retina ist und die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, den normalisierten Lokalorientierungsvektor für jedes Pixel der Vielzahl von Pixeln so zu berechnen, dass die Größe des normalisierten Lokalorientierungsvektors an jedem Pixel der Vielzahl von Pixeln 0,8 DD bis 2,4 DD beträgt.
 
23. Einrichtung gemäß Anspruch 21, wobei das Bild entweder ein Reflektanz-Scanlaser-Ophthalmoskopbild der Retina ist oder ein Fundusfoto der Retina ist und die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, den normalisierten Lokalorientierungsvektor für jedes Pixel der Vielzahl von Pixeln so zu berechnen, dass die Größe des normalisierten Lokalorientierungsvektors an jedem Pixel der Vielzahl von Pixeln 0,4 DD bis 1,2 DD beträgt.
 
24. Einrichtung gemäß Anspruch 20, wobei das anatomische Merkmal der Retina den Sehnervenkopf umfasst, und die Anweisungen bei Ausführung durch den Prozessor den Prozessor veranlassen, die Werte in den Akkumulatoren unter Verwendung eines Kernels mit einer Standardabweichung von 0,1 DD bis 0,5 DD und bevorzugterer 0,2 DD bis 0,4 DD zu glätten, wobei DD der Durchmesser des Sehnervenkopfs ist.
 


Revendications

1. Support de stockage lisible par ordinateur non transitoire stockant des instructions de programme d'ordinateur qui, lorsqu'elles sont exécutées par un processeur, amènent le processeur à traiter des données d'image définissant une image d'une rétine pour déterminer un emplacement d'une caractéristique anatomique de la rétine dans l'image en :

recevant les données d'image ;

calculant, pour chacun d'une pluralité de pixels des données d'image reçues, un vecteur d'orientation locale respectif indicatif de l'orientation de tout vaisseau sanguin présent dans l'image au niveau du pixel ou adjacent au pixel ;

calculant un vecteur d'orientation locale normalisé pour chacun de la pluralité de pixels par normalisation du vecteur d'orientation locale calculé pour chacun des pixels de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chacun des pixels prend une valeur prédéterminée commune ;

agissant sur un réseau d'accumulateurs, chaque accumulateur dans le réseau étant associé à un pixel respectif des données d'image reçues, par :

pour chacun de la pluralité de pixels, ajout d'une valeur respective à un accumulateur associé à un pixel des données d'image reçues qui est décalé du pixel de la pluralité de pixels de la valeur prédéterminée dans une direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels ; et

lissage des valeurs dans les accumulateurs ; et

déterminant l'emplacement de la caractéristique anatomique dans l'image de la rétine à l'aide de l'emplacement d'un pixel des données d'image reçues qui est associé à un accumulateur ayant accumulé une valeur accumulée, laquelle valeur accumulée se trouve dans une plage prédéterminée d'un extremum des valeurs accumulées dans les accumulateurs.


 
2. Support de stockage lisible par ordinateur non transitoire selon la revendication 1, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à agir sur une pluralité, N, de réseaux d'accumulateurs, où N est un nombre entier égal à au moins 2, et dans lequel :

chacun des réseaux d'accumulateurs est associé à une direction respective parmi N directions quantifiées pour le vecteur d'orientation locale ;

chacun des accumulateurs dans chacun des N réseaux est associé à un pixel respectif dans les données d'image reçues ;

le processeur agit sur la pluralité de réseaux d'accumulateurs en :

pour chacun de la pluralité de pixels :

quantifiant la direction du vecteur d'orientation locale calculé au niveau du pixel dans l'une des N directions quantifiées ; et

ajoutant une valeur respective à un accumulateur d'un réseau d'accumulateurs qui est associé à la même direction quantifiée que la direction quantifiée du vecteur d'orientation locale au niveau du pixel, l'accumulateur étant associé à un pixel des données d'image reçues qui est décalé du pixel de la pluralité de pixels de la valeur prédéterminée dans la direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels ; et

lissant les valeurs dans les accumulateurs de la pluralité de réseaux ;

le processeur calcule en outre, pour chaque ensemble de N accumulateurs qui sont agencés dans différents réseaux d'accumulateurs respectifs et associés au même pixel des données d'image reçues, un produit des valeurs cumulées respectives qui ont été accumulées dans les N accumulateurs ; et

le processeur détermine l'emplacement de la caractéristique anatomique dans l'image de la rétine en utilisant l'emplacement d'un pixel des données d'image reçues associé à un ensemble de N accumulateurs, pour lequel ensemble, le produit calculé de valeurs accumulées se trouve dans une plage prédéterminée d'un extremum des produits calculés de valeurs accumulées.


 
3. Support de stockage lisible par ordinateur non transitoire selon la revendication 2, dans lequel 2 ≤ N ≤ 32, et plus préférentiellement 4 ≤ N ≤ 16.
 
4. Support de stockage lisible par ordinateur non transitoire selon l'une quelconque des revendications précédentes, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent en outre le processeur à calculer un second vecteur d'orientation locale normalisé pour chacun de la pluralité de pixels en normalisant le vecteur d'orientation locale calculé pour chacun de la pluralité de pixels de sorte que l'amplitude du second vecteur d'orientation locale normalisé au niveau de chacun des pixels prend une seconde valeur prédéterminée commune, et à agir en outre sur chaque réseau d'accumulateurs en :
pour chaque pixel de la pluralité de pixels, ajoutant une valeur respective à un accumulateur associé à un pixel dans les données d'image reçues qui est décalé du pixel de la pluralité de pixels de la seconde valeur prédéterminée dans la direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels.
 
5. Support de stockage lisible par ordinateur non transitoire selon l'une quelconque des revendications précédentes, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à traiter les données d'image définissant l'image de la rétine pour déterminer l'emplacement de la caractéristique anatomique de la rétine dans l'image en effectuant un autre processus consistant à :
sélectionner des pixels parmi la pluralité de pixels de telle sorte que l'amplitude du vecteur d'orientation locale calculé au niveau de chacun des pixels sélectionnés dépasse un seuil, et en :

calculant le vecteur d'orientation locale normalisé pour chacun des pixels sélectionnés en normalisant le vecteur d'orientation locale calculé pour chacun des pixels sélectionnés de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chacun des pixels sélectionnés prend une valeur prédéterminée commune ; et

agissant sur chaque réseau d'accumulateurs par :
pour chacun des pixels sélectionnés, ajout d'une valeur respective à un accumulateur associé à un pixel des données d'image reçues qui est décalé du pixel sélectionné dans l'image de la valeur prédéterminée dans la direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel sélectionné.


 
6. Support de stockage lisible par ordinateur non transitoire selon la revendication 5, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à sélectionner les pixels parmi la pluralité de pixels de telle sorte que l'amplitude du vecteur d'orientation locale calculé au niveau de chacun des pixels sélectionnés dépasse le 50ième centile, et plus préférentiellement le 80ième centile, d'amplitudes de vecteur d'orientation locale des pixels dans les données d'image.
 
7. Support de stockage lisible par ordinateur non transitoire selon l'une quelconque des revendications précédentes, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à agir sur chaque réseau d'accumulateurs en :
pour chaque pixel de la pluralité de pixels, ajoutant, en tant que valeur respective, une pondération respective à l'accumulateur associé à un pixel dans les données d'image reçues qui est décalé du pixel de la pluralité de pixels de la valeur prédéterminée dans la direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels, la pondération étant indicative de l'amplitude du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels.
 
8. Support de stockage lisible par ordinateur non transitoire selon l'une quelconque des revendications précédentes, dans lequel la caractéristique anatomique de la rétine comprend au moins l'un parmi un disque optique, une jonction vasculaire reliant au moins deux vaisseaux sanguins, un chevauchement de vaisseaux sanguins de la rétine, et une fovéa.
 
9. Support de stockage lisible par ordinateur non transitoire selon la revendication 8, dans lequel la caractéristique anatomique de la rétine comprend le disque optique, et les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à calculer le vecteur d'orientation locale normalisé pour chaque pixel de la pluralité de pixels de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chaque pixel de la pluralité de pixels est comprise entre 0,5 DD et 2,5 DD, où DD est le diamètre du disque optique.
 
10. Support de stockage lisible par ordinateur non transitoire selon la revendication 9, dans lequel l'image est l'une parmi une image d'autofluorescence de la rétine et une image d'angiographie à la fluorescéine de la rétine, et les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à calculer le vecteur d'orientation locale normalisé pour chaque pixel de la pluralité de pixels de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chaque pixel de la pluralité de pixels est de 0,8 DD à 2,4 DD.
 
11. Support de stockage lisible par ordinateur non transitoire selon la revendication 9, dans lequel l'image est l'une parmi une image ophtalmoscopique à balayage laser en réflexion de la rétine et une photographie de fond de la rétine, et les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à calculer le vecteur d'orientation locale normalisé pour chaque pixel de la pluralité de pixels de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chaque pixel de la pluralité de pixels est de 0,4 DD à 1,2 DD.
 
12. Support de stockage lisible par ordinateur non transitoire selon la revendication 8, dans lequel la caractéristique anatomique de la rétine comprend le disque optique, et les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à lisser les valeurs dans les accumulateurs en utilisant un noyau ayant un écart type de 0,1 DD à 0,5 DD, et plus préférentiellement de 0,2 DD à 0,4 DD, où DD est le diamètre du disque optique.
 
13. Appareil de traitement d'image comprenant un processeur et une mémoire, la mémoire stockant des instructions exécutables par le processeur, de sorte que le processeur est opérationnel pour traiter des données d'image définissant une image d'une rétine pour déterminer un emplacement d'une caractéristique anatomique de la rétine dans l'image en :

calculant, pour chacun d'une pluralité de pixels des données d'image reçues, un vecteur d'orientation locale respectif indicatif de l'orientation de tout vaisseau sanguin présent dans l'image au niveau du pixel ou adjacent au pixel ;

calculant un vecteur d'orientation locale normalisé pour chacun de la pluralité de pixels par normalisation du vecteur d'orientation locale calculé pour chacun des pixels de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chacun des pixels prend une valeur prédéterminée commune ;

agissant sur un réseau d'accumulateurs, chaque accumulateur dans le réseau étant associé à un pixel respectif des données d'image reçues, par :

pour chacun de la pluralité de pixels, ajout d'une valeur respective à un accumulateur associé à un pixel des données d'image reçues qui est décalé du pixel de la pluralité de pixels de la valeur prédéterminée dans une direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels ; et

lissage des valeurs dans les accumulateurs ; et

déterminant l'emplacement de la caractéristique anatomique dans l'image de la rétine à l'aide de l'emplacement d'un pixel des données d'image reçues qui est associé à un accumulateur ayant accumulé une valeur accumulée, laquelle valeur accumulée se trouve dans une plage prédéterminée d'un extremum des valeurs accumulées dans les accumulateurs.


 
14. Appareil selon la revendication 13, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à agir sur une pluralité, N, de réseaux d'accumulateurs, où N est un nombre entier égal à au moins 2, et dans lequel :

chacun des réseaux d'accumulateurs est associé à une direction respective parmi N directions quantifiées pour le vecteur d'orientation locale ;

chacun des accumulateurs dans chacun des N réseaux est associé à un pixel respectif des pixels dans les données d'image reçues ;

les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à :
agir sur la pluralité de réseaux d'accumulateurs en :

pour chacun de la pluralité de pixels :

quantifiant la direction du vecteur d'orientation locale calculé au niveau du pixel dans l'une des N directions quantifiées ; et

ajoutant une valeur respective à un accumulateur d'un réseau d'accumulateurs qui est associé à la même direction quantifiée que la direction quantifiée du vecteur d'orientation locale au niveau du pixel, l'accumulateur étant associé à un pixel des données d'image reçues qui est décalé du pixel de la pluralité de pixels de la valeur prédéterminée dans la direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels ; et

lissant les valeurs dans les accumulateurs de la pluralité de réseaux ;

calculer en outre, pour chaque ensemble de N accumulateurs qui sont agencés dans différents réseaux d'accumulateurs respectifs mais sont associés au même pixel des données d'image reçues, un produit des valeurs cumulées respectives qui ont été accumulées dans les N accumulateurs ; et

déterminer l'emplacement de la caractéristique anatomique dans l'image de la rétine en utilisant l'emplacement d'un pixel des données d'image reçues associé à l'un des ensembles de N accumulateurs, pour lequel ensemble d'accumulateurs, le produit calculé de valeurs accumulées se trouve dans une plage prédéterminée d'un extremum des produits calculés de valeurs accumulées.


 
15. Appareil selon la revendication 14, dans lequel 2 ≤ N ≤ 32, et plus préférentiellement 4 ≤ N ≤ 16.
 
16. Appareil selon l'une quelconque des revendications 13 à 15, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent en outre le processeur à calculer un second vecteur d'orientation locale normalisé pour chacun de la pluralité de pixels en normalisant le vecteur d'orientation locale calculé pour chacun de la pluralité de pixels de sorte que l'amplitude du second vecteur d'orientation locale normalisé au niveau de chacun des pixels prend une seconde valeur prédéterminée commune, et à agir en outre sur chaque réseau d'accumulateurs en :
pour chaque pixel de la pluralité de pixels, ajoutant une valeur respective à un accumulateur associé à un pixel dans les données d'image reçues qui est décalé du pixel de la pluralité de pixels de la seconde valeur prédéterminée dans la direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels.
 
17. Appareil selon l'une quelconque des revendications 13 à 16, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à traiter les données d'image définissant l'image de la rétine pour déterminer l'emplacement de la caractéristique anatomique de la rétine dans l'image en effectuant un autre processus consistant à :
sélectionner des pixels parmi la pluralité de pixels de telle sorte que l'amplitude du vecteur d'orientation locale calculé au niveau de chacun des pixels sélectionnés dépasse un seuil, et en :

calculant le vecteur d'orientation locale normalisé pour chacun des pixels sélectionnés en normalisant le vecteur d'orientation locale calculé pour chacun des pixels sélectionnés de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chacun des pixels sélectionnés prend une valeur prédéterminée commune ; et

agissant sur chaque réseau d'accumulateurs par :
pour chacun des pixels sélectionnés, ajout d'une valeur respective à un accumulateur associé à un pixel des données d'image reçues qui est décalé du pixel sélectionné dans l'image de la valeur prédéterminée dans la direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel sélectionné.


 
18. Appareil selon la revendication 17, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à sélectionner les pixels parmi la pluralité de pixels en sélectionnant des pixels parmi la pluralité de pixels de telle sorte que l'amplitude du vecteur d'orientation locale calculé au niveau de chacun des pixels sélectionnés dépasse le 50ième centile, et plus préférentiellement le 80ième centile, d'amplitudes de vecteur d'orientation locale des pixels dans l'image.
 
19. Appareil selon l'une quelconque des revendications 13 à 18, dans lequel les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à agir sur chaque réseau d'accumulateurs en :
pour chaque pixel de la pluralité de pixels, ajoutant, en tant que valeur respective, une pondération respective à l'accumulateur associé à un pixel dans les données d'image reçues qui est décalé du pixel de la pluralité de pixels de la valeur prédéterminée dans la direction prédéterminée par rapport à la direction du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels, la pondération étant indicative de l'amplitude du vecteur d'orientation locale au niveau du pixel de la pluralité de pixels.
 
20. Appareil selon l'une quelconque des revendications 13 à 19, dans lequel la caractéristique anatomique de la rétine comprend au moins l'un parmi un disque optique, une jonction vasculaire reliant au moins deux vaisseaux sanguins, un chevauchement de vaisseaux sanguins de la rétine, et une fovéa.
 
21. Appareil selon la revendication 20, dans lequel la caractéristique anatomique de la rétine comprend le disque optique, et les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à calculer le vecteur d'orientation locale normalisé pour chaque pixel de la pluralité de pixels de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chaque pixel de la pluralité de pixels est comprise entre 0,5 DD et 2,5 DD, où DD est le diamètre du disque optique.
 
22. Appareil selon la revendication 21, dans lequel l'image est l'une parmi une image d'autofluorescence de la rétine et une image d'angiographie à la fluorescéine de la rétine, et les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à calculer le vecteur d'orientation locale normalisé pour chaque pixel de la pluralité de pixels de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chaque pixel de la pluralité de pixels est de 0,8 DD à 2,4 DD.
 
23. Appareil selon la revendication 21, dans lequel l'image est l'une parmi une image ophtalmoscopique à balayage laser en réflexion de la rétine et une photographie de fond de la rétine, et les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à calculer le vecteur d'orientation locale normalisé pour chaque pixel de la pluralité de pixels de sorte que l'amplitude du vecteur d'orientation locale normalisé au niveau de chaque pixel de la pluralité de pixels est de 0,4 DD à 1,2 DD.
 
24. Appareil selon la revendication 20, dans lequel la caractéristique anatomique de la rétine comprend le disque optique, et les instructions, lorsqu'elles sont exécutées par le processeur, amènent le processeur à lisser les valeurs dans les accumulateurs en utilisant un noyau ayant un écart type de 0,1 DD à 0,5 DD, et plus préférentiellement de 0,2 DD à 0,4 DD, où DD est le diamètre du disque optique.
 




Drawing

































REFERENCES CITED IN THE DESCRIPTION



This list of references cited by the applicant is for the reader's convenience only. It does not form part of the European patent document. Even though great care has been taken in compiling the references, errors or omissions cannot be excluded and the EPO disclaims all liability in this regard.

Non-patent literature cited in the description