(19)
(11)EP 2 765 437 B1

(12)EUROPEAN PATENT SPECIFICATION

(45)Mention of the grant of the patent:
06.05.2020 Bulletin 2020/19

(21)Application number: 13154916.4

(22)Date of filing:  12.02.2013
(51)Int. Cl.: 
G01R 33/56  (2006.01)
G01R 33/565  (2006.01)

(54)

A SIMPLE METHOD TO DENOISE RATIO IMAGES IN MAGNETIC RESONANCE IMAGING

EINFACHES VERFAHREN ZUM ENTRAUSCHEN VON VERHÄLTNISBILDERN BEI DER MAGNETRESONANZABBILDUNG

PROCÉDÉ SIMPLE DE DÉBRUITAGE DE RAPPORT D'IMAGES EN IMAGERIE PAR RÉSONANCE MAGNÉTIQUE


(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

(43)Date of publication of application:
13.08.2014 Bulletin 2014/33

(73)Proprietors:
  • Siemens Healthcare GmbH
    91052 Erlangen (DE)
  • École Polytechnique Fédérale de Lausanne (EPFL)
    1015 Lausanne (CH)

(72)Inventors:
  • Roche, Alexis
    1008 Jouxtens-Mézery (CH)
  • O'Brien, Kieran
    1004 Lausanne (CH)

(74)Representative: Maier, Daniel Oliver 
Siemens AG Postfach 22 16 34
80506 München
80506 München (DE)


(56)References cited: : 
  
  • VAN DE MOORTELE P F ET AL: "T1 weighted brain images at 7 Tesla unbiased for Proton Density, T2* contrast and RF coil receive B1 sensitivity with simultaneous vessel visualization", NEUROIMAGE, ACADEMIC PRESS, ORLANDO, FL, US, vol. 46, no. 2, 1 June 2009 (2009-06-01), pages 432-446, XP026467914, ISSN: 1053-8119, DOI: 10.1016/J.NEUROIMAGE.2009.02.009 [retrieved on 2009-02-20]
  • MARQUES J P ET AL: "MP2RAGE, a self bias-field corrected sequence for improved segmentation and T1-mapping at high field", NEUROIMAGE, ACADEMIC PRESS, ORLANDO, FL, US, vol. 49, no. 2, 15 January 2010 (2010-01-15), pages 1271-1281, XP026796256, ISSN: 1053-8119 [retrieved on 2009-10-09]
  • HYVARINEN A ET AL: "Independent component analysis: algorithms and applications", NEURAL NETWORKS, ELSEVIER SCIENCE PUBLISHERS, BARKING, GB, vol. 13, no. 4-5, 1 June 2000 (2000-06-01) , pages 411-430, XP004213197, ISSN: 0893-6080, DOI: 10.1016/S0893-6080(00)00026-5
  • KIERAN O'BRIEN ET AL: "A simple method to denoise MP2RAGE", PROCEEDINGS OF THE INTERNATIONAL SOCIETY FOR MAGNETIC RESONANCE IN MEDICINE, 21ST ANNUAL MEETING AND EXHIBITION, SALT LAKE CITY, UTAH, USA, 20-26 APRIL 2013, vol. 21, 6 April 2013 (2013-04-06), page 269, XP055061678,
  
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


[0001] The invention relates to a method for automatically removing background noise in magnetic resonance (MR) imaging, and to a device for carrying out said method. In particular, the present invention is interested in removing noise in Magnetization-Prepared 2 Rapid Gradient-Echoes (MP2RAGE) ratio images.

[0002] At high static magnetic fields B0 (i.e. for B0 ≥ 3T) used for MR imaging, the transmit and receive radio frequency (RF) coils of a magnetic resonance imaging (MRI) scanner are very inhomogeneous. This causes unwanted intensity variations across an image which can impair clinical diagnosis. One method for removing this unwanted intensity variation is the MP2RAGE image acquisition technique that has been proposed by Van de Moortele et al., in Neuroimage 46, 432 (2009). The MP2RAGE image acquisition technique is an extension of the standard Magnetization-Prepared Rapid Gradient-Echo (MPRAGE) acquisition technique and is able to provide a bias-free T1 contrast with heavily reduced T2*- weighting and Proton Density (PD) - weighting compared to the standard MPRAGE.

[0003] The MP2RAGE image acquisition technique comprises:
  • generating a double-echo sequence, namely a first rapid gradient echo block at a first inversion time TI1 separated by a delay TB from a second rapid gradient echo block at a second inversion time TI2 for generating two different standard images at different inversion times and with different contrasts. Said two standard images have thus different weighting but the same receive/transmit field inhomogeneities and might be used for overcoming image quality degradation arising from spatial inhomogeneity in the transmit and receive RF coil B1 profiles that is observable in human MR images at said high static magnetic fields B0;
  • simultaneously acquiring said two different standard images that are preferentially and respectively a conventional 3D T1-weighted MPRAGE (GRETI1) image, acquired at a time delay or inversion time TI1 from a magnetization preparation RF pulse, and a 3D Gradient Echo-Proton Density (GRETI2) image, acquired at a second time delay TI2 from a magnetization preparation RF pulse, wherein 3D refers to 3 dimensional. The acquisition of said images is notably performed according to conventional MRI techniques by means of a conventional MRI scanner;
  • combining said two different standard images in a ratio image obtained by calculating a normalized ratio between said two standard images, for example by dividing the GRETI1 magnitude image by the GRETI2 magnitude image. The normalized ratio between said two standard images provides a reduction of intensity field bias by eliminating signal variations induced by the receive RF coil B1 profiles and a removal of T2* and PD components. The ratio image is known on Siemen's MRI scanners as a "uniform" image and might be preferentially calculated using the following simple (1) or complex ratio (2):



wherein Ixl refers to the magnitude of x and x* stands for the complex conjugate of x, x being GRETI1 or GRETI2.

[0004] The MP2RAGE image acquisition technique aims thus to produce unbiased high resolution T1-weighted 3D images relying on two 3D standard images that are otherwise frequently utilized on clinical scanners. The MP2RAGE image acquisition technique is also further described by Marques et al. in the paper Neuroimage 49, 1271 (2010).

[0005] Though this approach has generated considerable interest in the MR community, it has unfortunately the disadvantage that combining said two different standard images in a ratio image amplifies the background noise of said ratio image. Besides being visually displeasing, the high background noise and the increased noise level in the meninges (tissues surrounding the brain) are problematic for registration and automatic segmentation algorithms. In particular at 7T certain regions of the brain (e.g. the cerebellum) and the neck where the transmit B1 field is poorly distributed resulting in poor preparation of the magnetization, and which have low Signal Noise Ratio (SNR) in just one of said two different standard images, result in bright signal intensities in the ratio image. These regions lose contrast and would also interfere with image analysis algorithms.

[0006] The background noise is a result of a numerical instability when dividing voxels, i.e. a volumetric pixel in a 3D image, with a very low SNR. If the intensity field biases are removed using the simple ratio given by Eq. 1, the value diverges when the signal of the denominator is low or noise. The complex ratio given in Eq. 2 has the advantageous property that it limits the intensity values of the image between - 0.5 and 0.5 but remains numerically unstable when both modules are small. Moreover, since the phase points in any arbitrary direction when the SNR is low, then the background noise takes on a "salt and pepper" noise characteristic, spreading across the range -0.5 and 0.5 for intensity values.

[0007] Two solutions have been proposed to remove the background noise. The first solution is to generate a mask on one of the image data sets typically GRETI2 and to apply it to the ratio image (Van de Moortele et al., Neuroimage 46, 432 (2009)). Though straightforward, this approach risks a thresholding out data from inside the brain and requires therefore manual intervention by the user. The second solution multiplies the ratio image by the magnitude image of one of the data sets typically GRETI2, followed by a global rescaling or equalization of the image histogram (Fujimoto et al., Proc. 19th ISMRM #130). This approach unfortunately reintroduces the intensity field biases that the image ratio intended to remove thus defeating the features of the MP2RAGE image acquisition technique.

[0008] In addition, poor inversion in the cerebellum and neck results in a low SNR in either the GRETI1 or GRETI2 images. Consequently, regions of bright signal intensity appear in the ratio image, which cannot be removed by a simple threshold on GRETI2 or would intensify using a multiplication and thus would still be present.

[0009] An objective of the present invention is to propose a method for improving the image quality of said ratio images, and preferentially for automatically removing the background noise generated in a ratio image obtained by means of the MP2RAGE image acquisition technique, overcoming thus the previously mentioned problems.

[0010] This objective is achieved by the solution proposed on the basis of the method described in independent claim 1.

[0011] According to the present invention, the claimed method comprises the steps of:
  • generating a Magnetic Resonance Imaging sequence designed for acquiring data from an object to be imaged, wherein said Magnetic Resonance Imaging sequence is configured for acquiring data for generating at least two different images, respectively a first image P and a second image NP, wherein said at least two different images have different contrasts but share the same spatial inhomogeneity in the transmit and receive RF coil B1 profiles used for acquiring said two different images;
  • acquiring said two different images, notably by means of a conventional MRI technique and using a conventional MRI scanner;
  • combining said two different images in a ratio image S, notably by means of a computer machine;
    whereby said ratio image S is obtained by calculating a ratio of the first image P and the second image NP that is tunable by a parameter γ wherein the parameter γ is optimized for maximizing the negentropy of the ratio image S and thus removing noise in the ratio image S and improve the contrast characteristics in regions of low SNR of the ratio image S.


[0012] Preferentially, said MRI sequence uses a MP2RAGE image acquisition technique for acquiring said two different images and comprises therefore the generation of a double-echo sequence, namely a first rapid gradient echo block at a first inversion time TI1 separated by a delay TB from a second rapid gradient echo block at a second inversion time TI2 for generating said two different images at different inversion times and in particular with different contrasts. Preferentially, the claimed method comprises a simultaneous acquisition of said two different images, wherein said first image P and a second image NP are preferentially and respectively a conventional 3D T1-weighted MPRAGE (GRETI1) image and a 3D Gradient Echo-Proton Density (GRETI2) image.

[0013] Negentropy, a concept which originated in the context of independent component analysis (A. Hyvärinen and E. Oja, Neural Networks 13, 411 (2000)), is a measure of how much the image histogram differs from a Gaussian distribution. An empirical observation is that the noisier an image, the closer its histogram is to a Gaussian distribution. Therefore, negentropy may be viewed as a measure of the amount of noise removed from the conventional ratio image. The additional penalty on large offsets ensures that noise is not removed at the expense of reintroducing a significant bias field back into the image. The present invention proposes thus a tunable modification of the standard image ratio calculations given by Eq. (1) or (2) that can be used to suppress the noise outside and inside the tissue and recover contrast in regions of low SNR. The present invention proposes to introduce the parameter γ into each ratio of the first image P and the second image NP as follows:





[0014] In the case of the MP2RAGE calculations, according to Eq. (1) or (2), Eq. (3) and (4) becomes preferentially





[0015] Advantageously and preferentially, the possible values of the image intensities can be forced towards either 0 in the case of Eq. (5), or -0.5 in the case of Eq. (6), in particular when the voxel SNR is low or noise. After optimization of the parameter y by means of its automatic tuning for maximize the negentropy, most of the noise is removed from the ratio image improving therefore contrast characteristics in regions of low SNR without introducing a large bias into the ratio image.

[0016] In particular, in order to tune the values of the parameter y so that it dominates noise without significantly impacting voxels with signal from the imaged object (i.e. it should not (re)introduce a large bias into the ratio image), the present method proposes to individually customize the value of the parameter y by maximizing the image negentropy for each image data set acquired preferentially by means of said conventional MRI technique.

[0017] To remove noise whilst minimizing the bias introduced back into the ratio image, the parameter γ is preferentially tuned by maximizing the ratio image negentropy whilst subject to a penalty factor, e.g. maximizing the ratio image negentropy minus a penalty factor. For example, in the case of a linear penalty factor which may favor small offset values, γ is tuned by maximizing the image negentropy minus a positive constant times the parameter γ, said positive constant times the parameter y being the linear penalty factor.

[0018] Preferentially, the tuning or numerical optimization of the parameter y can be carried out using a standard line minimization algorithm, such as Brent's method, Newton's method or a quasi-exhaustive search. Advantageously, Brent's method allows a fast a fast processing of the ratio image calculation, notably by means of a standard single processor PC.

Brief description of the drawings



[0019] 
Figure 1
schematic illustration of a preferred embodiment of the method according to the invention.

Detailed description



[0020] Figure 1 illustrates a preferred embodiment of the method according to the invention wherein a MP2RAGE image acquisition technique is used for the MRI sequence. The method comprises:
  • generating a MRI sequence comprising a double-echo sequence, namely a first rapid gradient echo block at a first inversion time TI1 separated by a delay TB from a second rapid gradient echo block at a second inversion time TI2 for generating two different images at said first and second inversion times and in particular with different contrasts, wherein the following parameters are for example used: TR/TE 6s/2.89ms, TI1/TI2 0.8s/2.7s, Matrix 256 x 240 x 176, and voxel 1.0 x 1.0 x 1.2mm, wherein TR, TE, TI1 and TI2 refer respectively, to the repetition time, the echo time, the first inversion time and the second inversion time;
  • simultaneously acquiring said two different images, respectively a conventional GRETI1 and GRETI2 magnitude and phase images;
  • combining said two different standard images in a ratio image S by means of a computer machine C;
    whereby said ratio image S is obtained by calculating a ratio of the first image GRETI1 and the second image GRETI2 using Eq. (6) wherein the parameter γ is automatically chosen for maximizing the negentropy of the ratio image S over positive real numbers subject to a linear penalty to favor small offset values, and wherein numerical optimization is carried out using Brent's method. Additionally, the obtained ratio image S may be linearly transformed to scale between 0 and 4095.


[0021] Finally, in contrast to state of the art techniques, the present invention features the parallel acquisition of two images, namely P and NP, notably by means of a MP2RAGE image acquisition technique, wherein said images have different contrasts but share the same spatial inhomogeneity in the transmit and receive RF coil B1 profiles, and calculates a ratio of said standard image volumes that is tunable by means of a parameter γ, which value is automatically chosen free of any user intervention, so that
  1. i) the salt and pepper noise outside and inside the skull is removed;
  2. ii) in regions of low signal intensity, the contrast is recovered (notably for the neck);
  3. iii) structures become better delineated (notably the cerebellum), which would aid morphometry packages.


[0022] The present invention is thus based on a simple modification and optimization procedure to suppress the background noise that inherently occurs when taking the ratio of MR images. In the case of MP2RAGE scans the result returns familiar T1-weighted images to the radiologist that have all the information intact i.e with no direct noise removal that could potentially alter the diagnostic capabilities of the image.

[0023] Advantageously, the present method is appropriate to any image ratio. Indeed, appropriate modifications of the image ratio can in particular be achieved for numerically stabilizing the image ratio calculation. For example, numerical instabilties such as a division by zero and/or small numbers in said image ratio can be avoided. In particular, the choice of the tunable parameter γ can be optimized for numerically stabilizing the image ratio calculation. In particular, the calculation of the image ratio while tuning the parameter γ might be subject to different penalty factors, like a linear or quadratic penalty factor, while using different optimization method for said calculation, such as said brent's method. In particular, different penalty factors could be applied depending on the preference for noise suppression or maintaining bias free image calculations. Therefore, the claimed method is opened to the use of different types of objective functions or mathematical optimization methods for tuning/optimizing the parameter γ when maximizing said negentropy.

[0024] The proposed combination of the standard image volumes is a tradeoff between the self-bias correcting properties of the ratio and numerical stability. The bias that is introduced in the brain is small (µ= -5.8 % ± 5.5%) and is slowly varying across the brain, which is advantageously easy to correct in a later image segmentation step. The images visually improve; they appear like a customary T1-weighted image. The recovery of contrast in regions of low SNR and better delineation of structures with the proposed optimized calculation of the ratio image will aid morphometry packages, in particular the segmentation of the total intracranial volume which is a key step in brain tissue classification.

[0025] Indeed, in other fields, such as in digital image processing, the ratio of images is often performed in change detection i.e. to see if an object in the image has moved or in microscopy to remove illumination artifacts. However in the latter the image ratio does not consist of pixels with pure noise or of pixels with very low SNR which give rise to the numerical instabilities seen when MRI images are combined such as in MP2RAGE. These approaches do suffer from a loss of dynamic range due to the ratio and consequently utilize a technique known as histogram equalization to recover the dynamic range. However such techniques can not deal with the numerical instabilities induced by the noise seen in MP2RAGE, unlike the proposed approach.


Claims

1. Magnetic Resonance Imaging method for generating a ratio image with improved contrast characteristics in regions of low SNR, the method comprising the steps of:

- generating a Magnetic Resonance Imaging sequence designed for acquiring data from an object to be imaged, wherein said Magnetic Resonance Imaging sequence is configured for acquiring data for generating at least two different images, respectively a first image P and a second image NP, wherein said at least two different images have different contrasts but share the same spatial inhomogeneity in the transmit and receive RF coil B1 profiles used for acquiring said two different images;

- acquiring said two different images;

- combining said two different images in a ratio image S;
characterized in that said ratio image S is obtained by calculating a ratio of the first image P and the second image NP that is tunable by a parameter γ wherein the parameter γ is optimized for maximizing the negentropy of the ratio image S and thus removing noise in the ratio image S and improve the contrast characteristics in regions of low SNR of the ratio image S, wherein said ratio image S is given by one of the following equation:

or


 
2. Method according to claim 1, characterized in that said Magnetic Resonance Imaging sequence is a MP2RAGE image acquisition sequence comprising a double-echo sequence, namely a first rapid gradient echo block at a first inversion time (TI1) separated by a delay (TB) from a second rapid gradient echo block at a second inversion time (TI2) for generating said two different images at the first and second inversion times.
 
3. Method according to claim 2, characterized in that the first image P is a conventional 3D T1-weighted MPRAGE image acquired at the first inversion time (TI1) of the MP2RAGE image acquisition sequence and the second image NP is a 3D Gradient Echo-Proton Density image acquired at the second inversion time (TI2) of the MP2RAGE image acquisition sequence.
 
4. Method according to one of the claims 1-3, characterized in that the mathematical optimization method is a line minimization algorithm.
 
5. Method according to claim 4, characterized in that said line minimization algorithm is a Brent's method, or Newton's method or a quasi-exhaustive search.
 


Ansprüche

1. Magnetresonanzabbildungsverfahren zum Erzeugen eines Verhältnisbildes mit verbesserten Kontrastmerkmalen in Regionen mit geringem SNR wobei das Verfahren die folgenden Schritte umfasst:

- Erzeugen einer Magnetresonanzabbildungssequenz, die gestaltet ist zum Erfassen von Daten von einem abzubildenden Objekt, wobei die Magnetresonanzabbildungssequenz ausgelegt ist zum Erfassen von Daten zum Erzeugen von wenigstens zwei unterschiedlichen Bildern, jeweils ein erstes Bild P und ein zweites Bild NP, wobei die wenigstens zwei unterschiedlichen Bilder unterschiedliche Kontraste haben, aber die gleiche räumliche Inhomogenität in den B1-Profilen der HF-Sende- und -Empfangsspulen teilen, die zum Erfassen der beiden unterschiedlichen Bilder verwendet werden,

- Erfassen der beiden unterschiedlichen Bilder,

- Kombinieren der beiden unterschiedlichen Bilder in einem Verhältnisbild S,
dadurch gekennzeichnet, dass das Verhältnisbild S erhalten wird durch Berechnen eines Verhältnisses des ersten Bildes P und des zweiten Bildes NP, das durch einen Parameter y einstellbar ist, wobei der Parameter y optimiert ist zum Maximieren der Negentropie des Verhältnisbildes S und somit zum Entfernen des Rauschens in dem Verhältnisbild S und Verbessern der Kontrastmerkmale in Regionen mit geringem SNR des Verhältnisbildes S, wobei das Verhältnisbild S durch die folgende Gleichung vorgegeben ist:

oder


 
2. Verfahren nach Anspruch 1, dadurch gekennzeichnet, dass die Magnetresonanzabbildungssequenz eine MP2RAGE-Bilderfassungssequenz ist, die eine Doppelechosequenz, nämlich einen ersten Rapid-Gradientenecho-Block zu einer ersten Inversionszeit (TI1), der durch eine Verzögerung (TB) von einem zweiten Rapid-Gradientenecho-Block zu einer zweiten Inversionszeit (TI2) getrennt ist, zum Erzeugen der zwei unterschiedlichen Bilder zu den ersten und zweiten Inversionszeiten umfasst.
 
3. Verfahren nach Anspruch 2, dadurch gekennzeichnet, dass das erste Bild P ein zu der ersten Inversionszeit (TI1) der MP2RAGE-Bilderfassungssequenz erfasstes konventionelles dreidimensionales, mit T1 gewichtetes MPRAGE-Bild ist und das zweite Bild NP ein zu der zweiten Inversionszeit (TI2) der MP2RAGE-Bilderfassungssequenz erfasstes dreidimensionales Gradientenecho-Protonendichtebild ist.
 
4. Verfahren nach einem der Ansprüche 1 bis 3, dadurch gekennzeichnet, dass das mathematische Optimierungsverfahren ein Zeilenminimierungsalgorithmus ist.
 
5. Verfahren nach Anspruch 4, dadurch gekennzeichnet, dass der Zeilenminimierungsalgorithmus ein Brent-Verfahren oder Newton-Verfahren oder eine quasi-erschöpfende Suche ist.
 


Revendications

1. Procédé d'imagerie par résonance magnétique pour générer une image-ratio avec des caractéristiques de contraste améliorées dans des régions de bas RSB, le procédé comprenant les étapes de :

- génération d'une séquence d'imagerie par résonance magnétique conçue pour acquérir des données depuis un objet à imager, dans lequel ladite séquence d'imagerie par résonance magnétique est configurée pour acquérir des données pour générer au moins deux images différentes, respectivement une première image P et une deuxième image NP, dans lequel lesdites au moins deux images différentes ont des contrastes différents mais partagent la même inhomogénéité spatiale dans les profils B1 de bobines RF de transmission et de réception utilisés pour acquérir lesdites deux images différentes ;

- acquisition desdites deux images différentes ;

- combinaison desdites deux images différentes en une image-ratio S ;
caractérisé en ce que ladite image-ratio S est obtenue en calculant un ratio de la première image P et de la deuxième image NP qui est réglable par un paramètre y dans lequel le paramètre y est optimisé pour maximiser la néguentropie de l'image-ratio S et ainsi éliminer un bruit dans l'image-ratio S et améliorer les caractéristiques de contraste dans des régions de bas RSB de l'image-ratio S, dans lequel ladite image-ratio S est donnée par une des équations suivantes :

ou




 
2. Procédé selon la revendication 1, caractérisé en ce que ladite séquence d'imagerie par résonance magnétique est une séquence d'acquisition d'images MP2RAGE comprenant une séquence de double écho, notamment un premier bloc d'écho de gradient rapide à un premier temps d'inversion (TI1) séparé par un délai (TB) d'un deuxième bloc d'écho de gradient rapide à un deuxième temps d'inversion (TI2) pour générer lesdites deux images différentes aux premier et deuxième temps d'inversion.
 
3. Procédé selon la revendication 2, caractérisé en ce que la première image P est une image MPRAGE pondérée T1 3D classique acquise au premier temps d'inversion (TI1) de la séquence d'acquisition d'images MP2RAGE et la deuxième image NP est une image de densité de gradient d'écho-proton 3D acquise au deuxième temps d'inversion (TI2) de la séquence d'acquisition d'images MP2RAGE.
 
4. Procédé selon l'une des revendications 1-3, caractérisé en ce que le procédé d'optimisation mathématique est un algorithme de minimisation linéaire.
 
5. Procédé selon la revendication 4, caractérisé en ce que ledit algorithme de minimisation linéaire est un procédé de Brent, ou un procédé de Newton ou une recherche quasi-exhaustive.
 




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