Method for noise reduction in images from multiple MRI acquisitions

Description

[0001] The invention relates to a method for attenuating noise (or "denoising") from images obtained by multiple acquisitions via a magnetic resonance imaging device.

[0002] Magnetic resonance imaging, also known by the French acronym IRM or Anglo-Saxon MRI (for "Magnetic Resonance Imaging"), is based on an analysis of the response of a proton of a water molecule when it is excited in a magnetic field. This response depends on the environment of such a proton and makes it possible to differentiate several types of tissue. A nuclear magnetic resonance imaging device 1 of an imaging analysis system S, as illustrated by way of non-limiting example by the figures 1 et 2 , is generally used. This delivers a plurality of digital image sequences 12 of one or more parts of a patient's body, by way of non-limiting examples, the brain, the heart, the lungs. Said apparatus 1 applies for this purpose a combination of high-frequency electromagnetic waves to the part of the body considered and measures the signal re-emitted by certain atoms, such as by way of non-limiting example, hydrogen for nuclear magnetic resonance imaging. The apparatus thus makes it possible to determine the magnetic properties and, consequently, the chemical composition of biological tissues and therefore their natures, in each elementary volume, which is commonly called a voxel, of the imaged volume. The magnetic resonance imaging apparatus 1 is controlled using a console 2.A user 6, for example an operator, practitioner or researcher, can thus choose commands 11 to control the apparatus 1, from parameters or instructions 16 entered via a human-machine input interface 8 of the analysis system. Such a human-machine interface 8 may consist for example of a computer keyboard, a pointing device, a touch screen, a microphone or, more generally, any interface arranged to translate a gesture or an instruction issued by a human 6 into control or parameter data. From information 10 produced by said apparatus 1, a plurality of sequences of digital images 12 of a part of a human or animal body are obtained. We will also call such information 10 or images 12 “experimental data”.

[0003] The image sequences 12 may optionally be stored within a server 3, i.e. a computer equipped with its own storage means, and constitute a medical file 13 of a patient. Such a file 13 may comprise images of different types, such as functional images highlighting the activity of the tissues or anatomical images reflecting the properties of the tissues. The image sequences 12 or, more generally, the experimental data, are analyzed by a processing unit 4 arranged for this purpose. Such a processing unit 4 may, for example, consist of one or more microprocessors or microcontrollers implementing suitable application program instructions loaded into storage means of the imaging analysis system S.

[0004] "Methods of storage" means any volatile or, advantageously, non-volatile computer memory. Non-volatile memory is a computer memory whose technology allows its data to be retained in the absence of an electrical power supply. It can contain data resulting from inputs, calculations, measurements and / or program instructions. The main non-volatile memories currently available are electrically writable, such as EPROM ("Erasable Programmable Read-Only Memory") or electrically writable and erasable, such as EEPROM ("Electrically-Erasable Programmable Read-Only Memory"), flash, SSD ("Solid-State Drive"), etc. Non-volatile memories are distinguished from so-called "volatile" memories, the data of which is lost in the absence of an electrical power supply.The main volatile memories currently available are RAM (Random Access Memory), DRAM (dynamic random access memory, requiring regular updating), SRAM (static random access memory requiring such updating during a power shortage), DPRAM or VRAM (particularly suitable for video), etc. A "data memory", in the rest of the document, can be volatile or non-volatile depending on the intended application.

[0005] The processing unit 4 comprises means of communication with the outside world in order to collect the images. Said means of communication further enable the processing unit 4 to deliver, in fine, a rendering, for example graphic and / or sound, of an estimation or quantification of a biomarker developed by said processing unit 4 from the experimental data 10 and / or 12 obtained by magnetic resonance imaging, to a user 6 of the imaging analysis system S via an output human-machine interface 5. Throughout the document, the term “output human-machine interface” means any device, used alone or in combination, making it possible to output or deliver a graphic, haptic, sound or, more generally, human-perceptible representation of a reconstructed physiological signal, in this case a biomarker, to a user 6 of a magnetic resonance imaging analysis system S. Such an output human-machine interface 5 may consist, in a non-exhaustive manner, of one or more screens, speakers or other suitable alternative means.Said user 6 of the imaging analysis system S can thus confirm or deny a diagnosis, decide on a therapeutic action that he deems appropriate, deepen research work, refine adjustment parameters of measuring equipment, etc. Optionally, this user 6 can also configure the operation of the processing unit 4 or the output human-machine interface 5 by means of operating and / or acquisition parameters 16. For example, he can thus define display thresholds or choose the biomarkers, indicators or estimated or quantified parameters for which he wishes to have a representation. The user uses for this the input human-machine interface 8 previously mentioned or a second input interface provided for this. Advantageously, the input 8 and output 5 human-machine interfaces can constitute a single physical entity.Said input 8 and output 5 human-machine interfaces of the imaging analysis system can also be integrated into the acquisition console 2. There is a variant, described in connection with the . figure 2 , for which an imaging system, as described previously, further comprises a preprocessing unit 7 for analyzing the image sequences 12, deducing experimental signals 15 therefrom and delivering the latter to the processing unit 4 which is thus relieved of this task.

[0006] Multiple two-dimensional images or three-dimensional volumes can be acquired with or without changing the settings and parameters of the medical imaging system S. For example, to study the temporal response to the arrival of a contrast agent, one can simply repeat the acquisition without changing the settings and parameters of said system. In contrast, other imaging techniques require changing said settings and / or parameters to sample the desired parameter space.

[0007] Among the techniques or modalities based on magnetic resonance imaging, we distinguish chemical exchange saturation transfer imaging or CEST for Chemical Exchange Saturation Transfer, according to English terminology. Such a technique consists of applying a radiofrequency pulse according to different determined resonance frequencies ω, so that chemical species of interest reach a state of saturation. Said determined resonance frequencies ω are in fact respectively associated with different chemical species of interest, such as, but not limited to, the amide group (-NH), the amine group (-NH2) or the hydroxyl group (-OH). Their labile hydrogen protons thus excited are exchanged with unexcited hydrogen protons of water. Such an application of a radiofrequency pulse, at a determined resonance frequency ω other than that ω0 associated with water, continuously repeated for a determined duration, for example a few seconds, leads to an accumulation of water saturation of a chemical species of interest.A concentration of the latter can be indirectly measured by the decrease in the water signal, a phenomenon called the "CEST effect" which is analyzed in the form of a Z-spectrum. Such a decrease can be easily detected by fast magnetic resonance imaging acquisition sequences such as, but not limited to, single shot echo-planar imaging. The . figure 3 presents, for a voxel V of interest, an example of a spectrum Z in the form of a set of samples Zi(Δω) (or “data set” according to Anglo-Saxon terminology) according to relative frequency shifts Δω with respect to the water frequency, such shifts Δω being expressed in ppm. More precisely, said figure 3 illustrates different images M(-6ppm), M(-3.5ppm), M(0ppm), M(3.5ppm), M(6ppm) corresponding to frequency shifts Δω respectively equal to -6 ppm, 3.5 ppm, 0 ppm, 3.5 ppm and 6 ppm. It also presents an image M0 acquired for frequencies far from the water frequency. The spectrum Z(Δω), which can be described as "normalized", corresponds, for a given voxel, to the ratio M(Δω) on M0, for frequency shifts between -6 ppm and 6ppm. When the samples Zi(Δω) are ordered according to increasing Δω, as indicated by figure 3 , said spectrum Z can be described in the form of a discrete signal describing m measurements, advantageously normalized by a signal measured without saturation by radiofrequency and expressed as percentages, of the magnitude of an experimental signal delivered by a measuring device 1 for an elementary volume of an organ. According to the figure 3 , this set of m samples, when the latter are ordered according to increasing frequency shifts Δω, constitutes a discrete signal Z substantially describing a 'V' whose minimum Zi(Δω)m is associated with the frequency shift Δω =0 ppm, when the device is perfectly adjusted and / or when the static magnetic field was homogeneous.

[0008] The CEST technique improves the detection of certain metabolites in the human body whose concentration is insufficient to be detected by traditional magnetic resonance imaging sequences. This CEST technique thus provides valuable information for a practitioner seeking to establish a diagnosis and make a therapeutic decision in the treatment of pathologies.

[0009] We can also mention diffusion tensor imaging, also known by the acronym DTI for "Diffusion Tensor Imaging" for which data are acquired for different intensities (b values) and directions of the magnetic diffusion gradient. In this category we find perfusion imaging, also known by the acronym PWI for "Perfusion Weighted Imaging" and two of its main techniques: dynamic susceptibility contrast, known by the acronym DSC for "Dynamic Susceptibility Contrast" and enhanced dynamic contrast, known by the acronym DCE for "Dynamic Contrast Enhanced". There is also the technique of high angular resolution diffusion imaging also known by the acronym HARDI for "High Angular Resolution Diffusion Imaging".

[0010] In magnetic resonance imaging, the signal induced in a receiving coil is a continuous complex signal, i.e. it has a real part and an imaginary part, acquired in a frequency domain commonly called "k-space" or "k-space" according to Anglo-Saxon terminology. The noise on such a signal can be described as a complex, additive and uncorrelated contribution to said pure signal. Such noise therefore has a real component and an imaginary component independent of each other and distributed identically according to a Gaussian distribution with a zero mean and a standard deviation (cf. Cárdenas-Blanco et al. 2008; Aja-Fernández and Vegas-Sánchez-Ferrero 2016 - technical teaching accessible for example via the hyperlink "https: / / doi.org / 10.1007 / 978-3-319-39934-8"). However, it is common in this field to consider magnitude images rather than the said complex signal.Since the transformation from the complex signal to the magnitude signal recorded in magnitude images is nonlinear, the distribution of pixel intensities in the resulting images is generally not Gaussian and, under certain conditions, can be modeled by a Rice distribution, as detailed in Cárdenas-Blanco, Arturo, Cristian Tejos, Pablo Irarrazaval, and Ian Cameron - 2008, “Noise in Magnitude Magnetic Resonance Images.” Concepts in Magnetic Resonance Part A 32A (6): 409-16, available at https: / / doi.org / 10.1002 / cmr.a.20124.

[0011] In the field of magnetic resonance imaging, it is known to reduce or remove noise from experimental data (i.e., measured data). The terms "denoising" or "denoising" are used to describe this operation, which consists of rejecting or filtering the noise to consider only the important information within the experimental data. For this, many methods use principal component analysis, also known by the acronym PCA, or even the English acronym PCA for "Principal Component Analysis".

[0012] Principal component analysis is a non-parametric technique (i.e., one that does not use any model) originally designed to reduce the dimensions of a data set, while preserving the essence of the data. Such an approach is disclosed, for example, in Jolliffe, I. 2002. “Principal Component Analysis.” Second. Springer-Verlag, accessible via the link “https: / / doi.org / 10.1007 / b98835.” Such a method identifies and exploits linear correlations in the data under consideration and extracts a new, smaller, uncorrelated data set called “principal components.” These extracted principal components are ordered according to the variability (or variances) they express. Thus, the first principal components describe the most informative part of the data, concentrating most of the variance in the data.

[0013] In simple terms, applied to imaging, the goal of principal component analysis is to identify the most meaningful basis for "re-expressing" noisy experimental data and constituting a set of denoised experimental data, i.e. for which the noise has been filtered in order to reveal one or more structures of interest possibly masked by said noise.

[0014] From a geometric point of view, the PCA problem can be simplified as follows. For a given collection of (real) points in an m-dimensional space, find q vectors for which: a jth< vector represents a line that best fits the points; said jth< vector being orthogonal to the j-1 vectors calculated previously.

[0015] A "best fit" is defined as finding a line or axis that minimizes the mean square distance of the points from the line. Such a fit is illustrated by the figure 4 . According to this figure, we can see, in the left part, a cloud of points falling within a two-dimensional reference frame or basis illustrated by the axes x1 and x2. We can see that, on the right part, the axis z1 is the line that best fits said points, the data or points being able to be projected into a new reference frame or basis described by the axis z1 and the axis z2 normal to z1. The principal component z1 is characterized by the largest variance λ1 which is also called "eigenvalue" or "eigenvalue" according to Anglo-Saxon terminology. Thus, the set Λ of the respective eigenvalues ​​λ1 and λ2 of the principal components z1 and z2 represents the quantities of information collected by said principal components or their respective informative capacities. We call "eigenvectors" Φ={φ1, φ2} or "eigenvectors" according to Anglo-Saxon terminology, the principal components z1 and z2 on the figure 4 .

[0016] Thus, a principal component analysis implies that: experimental data of dimension m can be expressed as linear combinations of vectors; the variance λ of each principal component indicates significant information; the q principal components, forming the new basis or the new reference frame, are orthogonal.

[0017] There figure 5 illustrates a known method for attenuating noise or “denoising” experimental data in the form of normalized Z-spectra (of m phases or frequencies each) for a collection of n spatial locations (also called “voxels)” of a human brain. Thus, for a given voxel, the experimental data is a vector of m acquisitions or samples. Such a method 100 comprises: a first step 110 of acquiring or collecting experimental data Z1 to Zn for a set of n voxels V1 to Vn at m different phases (or samples) of said voxel data in the form of a Casorati matrix C (matrix of n rows by m columns, for which each of the n rows contains m values ​​acquired for a given voxel. The n rows are dedicated respectively to the different voxels of a segment or volume considered; a step 120 for extracting or calculating the q = m principal components best describing this set of noisy experimental data, in this case for a given voxel, the spectrum Zi such that: Z i = Za + ∑ j = 1 q Zi − Za φ j T φ j

[0018] Za being the average of the spectrum Zi; for this, is calculated: cov C = Φ T ΛΦ

[0019] Φ describing the eigenvectors of the principal components and Λ their eigenvalues; the q components, therefore their eigenvectors φ1 to φq, extracted or calculated, are generally classified or ordered according to their respective variances or eigenvalues. The variance of a principal component of eigenvector φj is described by a scalar value λj, which is called an "eigenvalue"; a step 130 of determining the optimal number k figure 5 , a Zi' spectrum for a voxel of interest, such as: Zi ′ = Za + ∑ j = 1 k Zi − Za φ j T φ j a step 150 of reorganizing the denoised experimental data Z1' to Zn' in the original space of the voxels V1 to Vn and of exploiting all or part of said denoised experimental data Z1', ..., Zi', ..., to Zn' according to the chosen application.

[0020] For example, as indicated in the​ figure 5 , the noisy spectrum Zi of a voxel considered appears, after the implementation of such a method 100, in the form of a curve Zi' which is “softer” than that described by the “raw” experimental data Zi.

[0021] Step 120 has been described according to a first embodiment from the covariance matrix. Alternatively, such a production of q principal components may consist, according to a second embodiment, in the decomposition into singular values ​​(“Singular Value Decomposition” or SVD according to English terminology) of a Casorati matrix C as mentioned previously. According to this second embodiment, experimental data, for example in the form of a Zi spectrum, for a given voxel may be expressed in the form: Zi = Za + ∑ j = 1 q t ij v j = Za + t i V t where Za is the mean spectrum, i.e. the average of the values ​​of the i th< row of the Casorati matrix C, ti is a vector of q coefficients, hereinafter called "scores" to describe the respective contributions of the principal components and V t< is a matrix of q rows and q columns of said principal components. Thus, we can rewrite the matrix C such that C = TV t< + Za which is calculated from a singular value decomposition of the matrix C such that C - Za =USV t< where U is a matrix of n rows and n columns describing the scores (or coefficients or weights) of the principal components uj describes the principal component scores of rank j), S is a matrix of n rows and q columns describing q non-zero values ​​S 1 to S q called "singular values" and V t< is a matrix of m rows and m columns describing the new basis of the principal components.

[0022] The singular values ​​of the matrix S are ranked in descending order. Note that if the matrix C is transposed, its rows describing m variables and its columns n samples, then the respective interpretations of the matrices U and V are interchanged.

[0023] By identifying the terms in the previous equations, it can be deduced that T = US is a matrix of principal component scores multiplied by singular values, with V t< being the principal component matrix. Such a second embodiment of step 120 is mainly focused on analyzing the spatial information contained in each column of the matrix U or, similarly, of the matrix T. As discussed previously, such a step 120 may be based on the covariance matrix cov(C) = Φ T< ΛΦ for which Φ describes the eigenvectors of the principal components and Λ their eigenvalues; the q components, hence their eigenvectors φ1 to φq, extracted or calculated, are classified or ordered generally according to their respective variances or eigenvalues ​​λj.There is a direct relationship between the singular values ​​according to the second embodiment and the eigenvalues ​​according to the first embodiment obtained (i.e. from the covariance matrix) such that . λj = s j 2 n − 1 .

[0024] To attenuate the noise of experimental data, in this case a noisy spectrum Zi, said “denoised” experimental data is obtained by implementing step 140 of projecting the noisy experimental data onto the remaining k components to constitute a new set of denoised experimental data, in this case using the example of figure 5 , a Zi' spectrum for a voxel of interest, such as: Zi ′ = Za + ∑ j = 1 k t ij v j = Za + t i ′ V t ′ for which ti' is a vector of k scores and V t< ' is a matrix of k rows and q columns of the principal components.

[0025] Among the steps of denoising by principal component analysis, the component selection step 130 is the most delicate. To obtain the spatial information hidden in the principal components, each principal component (i.e., the actual voxel values ​​projected onto an eigenvector) can be reshaped into its original dimension (i.e., into a volume or slice), as shown in figure 6 . The images PC1 to PC29 associated respectively with said principal components of ranks 1 to 29 describe the scores associated with the eigenvectors φ1 to φ29 for all voxels V1 to Vn considered according to the mode of production 120 of the principal components retained. Such scores correspond to the projection of the experimental data onto said twenty-nine principal components. For simplification purposes, we will use the term "score" to describe a contribution of a principal component, regardless of the mode of production or extraction 120 of such a principal component.The principle of a principal component analysis, according to the state of the art, amounts in some way to seeking to best express the experimental signals or data of a set of voxels resulting from a multiple acquisition, in the form of linear combinations of principal components according to respective scores, said principal components being classified according to an order of preponderance resulting directly from eigenvalues ​​of said principal components or from singular values ​​associated with them. Thus, the PC1 image of the . figure 6 describes, like an image of a slice of a human brain, the scores of the first principal component, to reconstruct the experimental curves (Zi spectra for i between 1 and n) for n voxels considered. In the same way the PC2 image of the figure 6 describes, like an image of a slice of a human brain, the scores of the second principal component to reconstruct the experimental curves of the voxels considered. In other words, the PC2 image of the figure 6 describes the said experimental data Z1 to Zn projected onto the second principal component. The same applies to images PC3 to PC29 describing respectively the scores of the third to the twenty-ninth principal component for the voxels considered. This figure 6 illustrates the result of the implementation of step 120 previously mentioned. As we can see with the naked eye, the first nine images PC1 to PC9 respectively associated with the principal components of ranks 1 to 9 seem, to a decreasing degree, to express relevant spatial information as opposed to the other images PC10 to PC29 respectively associated with the principal components of higher ranks which seem to no longer indicate anything at all. Obviously, the selection 130 of k significant principal components is not done with the naked eye but according to an automatic process.

[0026] The state of the art is full of processes or methods using different mathematical or statistical models and assumptions, combinations with other techniques and various criteria to determine an "optimal" subset of k components in many fields and applications. Among these, some have been successfully used for denoising or filtering purposes in magnetic resonance imaging, generally reconstructing the experimental dataset using the informative components and rejecting the components mainly related to noise. To this end, many data-driven approaches have been proposed to determine the k principal components among q extracted principal components that are relevant to preserve and qk principal components to exclude when reconstructing the experimental data to obtain denoised experimental data.

[0027] However, most selection criteria are based on the set Λ of eigenvalues ​​associated with the q extracted principal components, i.e. based on variance measures expressed by the original data set, component by component, or on the residuals, i.e. on the respective mathematical differences between the original data and the reconstructed data. We can cite in this regard, and in a non-exhaustive manner, the document Breitling, Johannes, Anagha Deshmane, Steffen Goerke, Andreas Korzowski, Kai Herz, Mark E. Ladd, Klaus Scheffler, Peter Bachert, and Moritz Zaiss - 2019 - “Adaptive Denoising for Chemical Exchange Saturation Transfer MR Imaging.” NMR in Biomedicine, no. March 2019: 1-14, accessible by the link “https: / / doi.org / 10.1002 / nbm.4133”, the document Kaiser, Henry F. 1958. “The Varimax Criterion for Analytic Rotation in Factor Analysis. » Psychometrika 23 (3): 187-200, accessible via the link https: / / doi.org / 10.1007 / BF02289233 or also the document Valle, Sergio, Weihua Li, and S. Joe Qin. 1999. “Selection of the Number of Principal Components: The Variance of the Reconstruction Error Criterion with a Comparison to Other Methods.” Industrial and Engineering Chemistry Research 38 (11): 4389-4401, accessible via the link https: / / doi.org / 10.1021 / ie990110i). These publications present methods based on the eigenvalues ​​of the principal components based on Malinowski, Nelson and Median criteria and applied to the denoising of in vivo CEST magnetic resonance imaging data.

[0028] Other methods have been proposed for denoising experimental diffusion MRI data. Some of these methods are based on random matrix theory as disclosed for example in the paper Veraart, Jelle, Dmitry S. Novikov, Daan Christiaens, Benjamin Ades-aron, Jan Sijbers, and Els Fieremans - 2016 - "Denoising of Diffusion MRI Using Random Matrix Theory." Neurolmage 142: 394-406, accessible via the link "https: / / doi.org / 10.1016 / j.neuroimage.2016.08.016". Other methods are based on singular value shrinkage as described in the paper Ma, Xiaodong, Kâmil U urbil, and Xiaoping Wu - 2020 - « Denoise Magnitude Diffusion Magnetic Resonance Images via Variance-Stabilizing Transformation and Optimal Singular-Value Manipulation. » Neurolmage 215 (July): 116852, accessible via the link « https: / / doi.org / 10.1016 / j.neuroimage.2020.116852 » whose mathematical principles are explained in the document Gavish, Matan, and David L. Donoho - 2017 - « Optimal Shrinkage of Singular Values. » IEEE Transactions on Information Theory 63 (4): 2137-52, accessible via the link « https: / / doi.org / 10.1109 / TIT.2017.2653801 »).

[0029] Like the known process illustrated by the figure 5 , most of the methods proposed in the literature are based on the eigenvalues ​​Λ to determine the k components to preserve and those to exclude during data reconstruction. Thus, as suggested by the figure 5 , a step 130 consists of comparing, in this case via a logarithmic scale, the eigenvalues ​​λj of the q principal components, j being between 1 and q. Below a certain predetermined variance threshold, the principal components are considered as expressing little or no relevant information and are rejected. Only the first k principal components, i.e. those whose variances or eigenvalues ​​are greater than said threshold, are retained for use in step 140. These methods have a major drawback because they generally lead to ignoring anatomical structures and pathological information hidden in certain discarded principal components. This drawback arises directly from the use, as the main selector, of the eigenvalues ​​which are measures of variance of the different principal components, blindly measuring the noise and the relevant information.Principal component analysis generally assumes that variance indicates the presence or absence of important information, which assumes that the experimental data being analyzed have a high signal-to-noise ratio, i.e., that the power of the signal of interest is stronger than the power of the noise. This is a strong and sometimes incorrect assumption, especially for particularly noisy experimental magnetic resonance imaging data. Such a situation arises, for example, when the noise level in the experimental data delivered by a magnetic resonance imaging device is high or when the signals of interest concern a small anatomical or pathological area compared to the size of the images considered.Most of the methods proposed in the literature, based on eigenvalues, thus exclude principal components containing hidden anatomical structures or pathological information of interest because the contributions of these areas to the eigenvalues ​​of the associated components are low. Thus, a significant amount of relevant information is lost because it is amalgamated with noise according to the state of the art. The selection of principal components cannot be done according to such criteria that we could qualify as simplistic or trivial to guarantee a relevant exploitation of experimental data and deliver reliable indicators or biomarkers from denoised experimental data, as we can see in connection with the . figure 12 for example.

[0030] The invention addresses the drawbacks raised by the state of the art. Among the numerous advantages provided by the invention, we can mention more particularly that a denoising method according to the invention comprises a step of selecting the principal components making it possible to obtain a better compromise between denoising efficiency and conservation of the relevant spatial information of the experimental data considered during their reconstruction to produce denoised experimental data. As an example, a proposed selection criterion is based on the reduction in the variance of a score image associated with a principal component after the application of a smoothing filter thereon.This technique makes it possible to exploit hidden spatial information, in particular anatomical structures and pathological information, in all the extracted or calculated principal components and makes it possible to select those making it possible to express important information, independently of each other and of their ordering resulting from their extraction or production, thus minimizing the risk of rejecting relevant principal components. The invention also makes it possible to have a filtering step that we can describe as adaptive, that is to say making it possible to further denoise the experimental data considered, by rejecting the persistent spatial noise linked to each extracted principal component, while adapting the degree of filtering of the principal components to the quantity of relevant information that each of them makes it possible to express.

[0031] To this end, the invention provides a method as defined by claim 1 for attenuating the noise of experimental data resulting from multiple acquisitions by a magnetic resonance imaging device in relation to an elementary volume of interest, hereinafter referred to as a "voxel" of interest, among a plurality of voxels, said method being implemented by a processing unit of a medical imaging analysis system. Such a method comprises: a step of collecting said experimental data in relation to the plurality of voxels; a step of producing an ordered set, according to their respective eigenvalues, of q principal components as well as their respective scores for the plurality of voxels, said scores corresponding to the projection of said experimental data onto said q principal components; a step of selecting k first principal components from among the q principal components produced; a step of projecting said experimental data onto the k selected principal components and producing noise-attenuated experimental data in relation to the elementary volume of interest.

[0032] To optimize the denoising of experimental data without losing spatial information, the step of selecting k principal components of such a process includes: a sub-step of producing images of the scores of the principal components for the plurality of voxels and informative indicators quantifying the spatial information contained in said images of the scores associated with said principal components in the form of a rate of decrease in the values ​​of a determined characteristic of said scores before and after the application of a smoothing filter on said scores, said determined characteristic of the scores for a plurality of voxels belonging to a set of characteristics, taken alone or in combination, comprising the variance, the standard deviation, the median and the mean; a sub-step of determining the rank k for selecting the first k principal components from said informative indicators produced.

[0033] According to a first advantageous embodiment, the sub-step of determining the rank k can consist of calculating the mathematical difference between values ​​of said informative indicators when the latter are ordered in accordance with the principal components produced, the rank k being that of the principal component for which said mathematical difference becomes less than or equal to a determined threshold.

[0034] Alternatively, said sub-step of determining rank k may consist of detecting a plateau described by the values ​​of the informative indicators, from the informative indicator of the main component of rank q to the informative indicator of the main component of rank 1, rank k being that of the first main component whose value of the informative indicator deviates from said plateau.

[0035] To reinforce the weight of the predominant principal components during the reconstruction of the experimental data thus denoised, the step of selecting k principal components of a method according to the invention may comprise a sub-step of applying a filtering operation of the k selected principal components, the filtering intensity of which is specific to each of the k selected principal components and proportional to the informative indicator of the latter.

[0036] In this case, the filtering operation can consist of the single or combined use of a Gaussian filter, an average filter, a median filter, an anisotropic diffusion filter.

[0037] According to a second object, the invention relates to a computer program product as defined by claim 7 comprising one or more program instructions interpretable by the processing unit of a medical imaging analysis system, said program instructions being loadable into a non-volatile memory of said medical imaging analysis system, the execution of said instructions by said processing unit of which causes the implementation of a method for attenuating the noise of experimental data according to the invention.

[0038] According to a third object, the invention relates to a computer-readable storage medium comprising the instructions of such a computer program product according to said invention.

[0039] According to a fourth object, the invention relates to a medical imaging analysis system as defined by claim 9 comprising a processing unit arranged to attenuate the noise of experimental data resulting from multiple acquisitions by a magnetic resonance imaging device in relation to an elementary volume of interest, hereinafter called "voxel" of interest, among a plurality of voxels, said processing unit being configured to: collecting said experimental data in relation to the plurality of voxels; producing an ordered set, according to their respective eigenvalues, of q principal components as well as their respective scores for the plurality of voxels, said scores corresponding to the projection of said experimental data onto said q principal components; selecting k first principal components from among the q principal components produced; projecting said experimental data onto the k selected principal components and producing noise-attenuated experimental data in relation to the elementary volume of interest.

[0040] To optimize the denoising of experimental data without losing spatial information, the processing unit of such a system is configured to: producing images of the scores of the principal components for the plurality of voxels and informative indicators quantifying the spatial information contained in said images of the scores associated with said principal components in the form of a rate of decrease in the values ​​of a determined characteristic of said scores before and after the application of a smoothing filter on said scores, said determined characteristic of the scores for a plurality of voxels belonging to a set of characteristics, taken alone or in combination, comprising the variance, the standard deviation, the median and the mean; selecting the first k principal components from said produced informative indicators.

[0041] According to an advantageous embodiment, such a medical imaging analysis system may comprise a program memory comprising the program instructions of a computer program product in accordance with the invention.

[0042] Furthermore, the invention relates to a method as defined by claim 2 and a medical imaging analysis system as defined by claim 10, wherein the principal component analysis is replaced by a singular value decomposition, as an alternative to the method of claim 1 and the system of claim 9, respectively.

[0043] Other features and advantages will become more apparent upon reading the following description and examining the accompanying figures, including: there figure 1 , already described, illustrates a simplified description of a system for analyzing images obtained by magnetic resonance; figure 2 , already described, illustrates a simplified description of a variant of a system for analyzing images obtained by magnetic means; figure 3 , already described, presents an example of a Z spectrum in the form of a set of samples Zi(Δω) according to relative frequency shifts Δω with respect to the water frequency, such shifts Δω being expressed in ppm; figure 4 , already described, illustrates the principle of principal component analysis consisting of transforming variables or data linked to each other (also called "correlated" in statistics) into new variables or data decorrelated from each other; figure 5 , already described, illustrates a known method for denoising a noisy Zi spectrum resulting from a CEST acquisition according to the state of the art; figure 6 , already described, illustrate the experimental data projected onto different principal components extracted from a CEST data set after reshaping into their original space; figure 7 illustrates an example of a functional algorithm of a denoising method according to the invention; figure 8 illustrates an example of calculating an informative indicator of a principal component specific to the invention; figure 9 illustrates an example of selection of principal components according to the invention; figure 10 illustrates an example of noise attenuation of HARDI experimental scattering data for different phases; figure 11 illustrates an example of noise removed by implementing the invention from such HARDI data for different phases; figure 12 illustrates an example of exploitation and the benefit resulting from such exploitation of denoised data using the invention in the field of tractography; figure 13 illustrates an example of noise attenuation of experimental perfusion data for different phases; figure 14 illustrates an example of noise removed by implementing the invention from such perfusion data for different phases; figure 15 illustrates an advantageous embodiment of a method according to the invention allowing an adjustment of the eigenweights to the different principal components selected prior to the projection of the noisy experimental data onto them.

[0044] A method 100 for attenuating noise in medical images, in accordance with the invention and illustrated by the figure 7 , is advantageously translated into the form of a computer program product whose program instructions are intended to be implanted in the program memory of an element of a medical imaging system, such as the system S according to the figures 1 et 2 , for example a computer or a computer server or, more generally, any electronic object with sufficient computing power.

[0045] There figure 7 thus illustrates a method 100, in accordance with the invention, for attenuating the noise of experimental data resulting from multiple acquisitions by a magnetic resonance imaging device, such as the device 1 of the medical analysis system S illustrated by the figures 1 et 2 , in connection with a plurality of n voxels. Such a method comprises certain steps common with the method 100 according to the figure 5 . Thus, a “noise reduction” method 100 according to the invention may comprise, if necessary, a step 110 of collecting or acquiring noisy experimental data linked to or concerning said plurality of voxels V1 to Vn. Such experimental data may consist of curves such as spectra Z1 to Zn following a CEST type acquisition like the example linked to figures 3 And 5or, more generally, any experimental data for a set of voxels or pixels related to a volume or slice of an organ concerned by a plurality of m acquisitions according to different instants or different phases for example. Such experimental data could thus, in a non-exhaustive manner, result from a multi-phase acquisition of HARDI high angular resolution diffusion imaging, or even concern the field of perfusion imaging.

[0046] As a preferred but non-limiting example, the figure 7 described, like the figure 5 , experimental data Zi linked to a voxel of interest among the n acquired voxels. The figure 7 illustrates experimental data Zi, in the form of a noisy spectrum identical to that illustrated in figure 5 . Such noise results in discontinuities or oscillations that are more or less marked depending on the frequency shifts concerned, as evidenced by the partial enlargement of said spectrum Zi. To attenuate this phenomenon, like the process 100 known and illustrated by the figure 5 , a method 100 according to the invention comprises a step 120 of implementing a principal component analysis to produce or calculate q principal components, said step being similar to step 120 of the known method 100 according to the figure 5 . Such a step 120 consists of determining q principal components symbolized or associated respectively with eigenvalues ​​λ1, ..., λq or with singular values ​​s 1 to sq forming an ordered set of principal components whose rank is determined according to a set of scalars of decreasing values, such as the eigenvalues ​​Λ={λ1, ..., λq} or the singular values ​​S. Such a method 100 according to the invention further comprises a step 130 of determining or selecting k principal components from among the q produced, i.e. the principal components of ranks 1 to k to determine a more restricted set of principal components. The selection 130, in accordance with the invention and illustrated by the figure 7 , is clearly distinguished from the approach described very largely in the state of the art based on the eigenvalues ​​Λ. Such a step 130 specific to the invention will be detailed later. A method 100 according to the invention further comprises, like the known methods 100 such as illustrated by the figure 5 , a step 140 of projecting the noisy experimental data Zi onto the k principal components selected in step 130 and producing experimental data attenuated in noise Zi' or "denoised" in relation to an elementary volume of interest. We will be able to observe the increased efficiency of a method 100 according to the invention thanks to a better selection, or even a modification of the k principal components, prior to said projection 140 by examining the denoised spectrum Zi' describing a particularly smooth curve compared to that described by the original spectrum Zi. Like this illustrated by the figure 5 , step 140 of the method 100 according to the figure consists of projecting all of the noisy experimental data Z1, ..., Zi, ..., Zn onto the k selected principal components and producing noise-attenuated experimental data Z1', ..., Zi', ..., Zn' for the voxels of interest V1 to Vn. Thus, optionally and advantageously, such a method 100 according to the invention may comprise a step 150 of joint exploitation of the denoised experimental data Z1', ..., Zi', ..., Zn' for all or part of a plurality of said voxels V1 to Vn of interest. Such exploitation may for example consist of counting fibers in tractography from HARDI diffusion experimental data as we will study in connection with the figures 10 à 12 .

[0047] The invention is therefore distinguished mainly by the implementation of step 130 making it possible to select the relevant principal components, that is to say, making it possible to optimize the attenuation of the noise without losing relevant spatial information. Thus, as illustrated by figure 7 , such a step 130 comprises a sub-step 131 of producing informative indicators, referenced σ1 r< , ..., σq r< on the figure 7 , respectively calculated for the q principal components produced of ranks 1 to q. Such a set Σ r< of informative indicators σ1 r< to σq r< is an alternative to the set Λ of eigenvalues ​​λ1 to λq exploited directly by the known methods for selecting k principal components from among the q extracted. Such an informative indicator σj r< characterizes the capacity of a principal component of rank j to express relevant spatial information. It can advantageously consist of a rate of decrease in the values ​​of a determined characteristic of the scores of said principal component in relation to the plurality of voxels V1 to Vn considered or of interest, before and after the application of a smoothing filter on said scores. Such an operation 131a is illustrated by the figure 8 . This describes for the principal component of rank 3 two images PC3a and PC3b illustrating the scores of said principal component of rank 3 for all the voxels considered. The image PC3a describes said scores before the application 131a of a smoothing filter and the image PC3b describes these same scores after said application 131a of the smoothing filter. The latter thus appears more blurred compared to the image PC3a on the figure 8 Such an operation 131a can advantageously consist of the application of a Gaussian filter with a factor F of predetermined value, possibly configurable.

[0048] As a preferred but non-limiting example, said determined characteristic may consist of the standard deviation of said scores. The informative indicator σj r< of a principal component of rank j can then be calculated in an operation 131b such that: σj r = σj a − σj b σj a where σj a< is the standard deviation of the scores of the principal component of rank j before the application 131a of the smoothing filter and σj b< is the standard deviation of said scores after the application 131a of said smoothing filter.

[0049] In this case on the figure 8 , the informative indicator σ3 r< of the principal component of rank 3 is such that: σ 3 r = σ 3 a − σ 3 b σ 3 a

[0050] In the same way, the figure 8 illustrates the calculation of the informative indicator σ9 r< of the principal component of rank 9 is such that: σ 9 r = σ 9 a − σ 9 b σ 9 a or even the informative indicator σ29 r< of the principal component of rank 29 is such that: σ 29 r = σ 29 a − σ 29 b σ 29 a

[0051] In this way, the set Σr of informative indicators of the q principal components of ranks 1 to q produced in step 120 and ordered according to their respective eigenvalues ​​Λ or singular values ​​S, can be constituted at the end of the implementation of sub-step 131. As a variant, this set Σr of informative indicators of the q principal components can be ordered according to the numerical values ​​of said informative indicators.

[0052] Alternatively or in addition, such a determined characteristic σjr could exploit, instead of the standard deviation, the variance, the entropy, the median or even the average of said scores of the principal component of rank j, or even result from a combination of all or part of these.

[0053] Once the set Σ r< of informative indicators has been calculated, step 131 of a method 100 according to the invention comprises a sub-step 132 of determining the rank k to select the first k principal components (symbolized by the set of eigenvectors Φ'={φ1, ..., φk} or the matrix U'={u 1 , ..., uk} on the figure 7 ) from said set Σ r< of said informative indicators. The implementation of such a sub-step 132 is illustrated by the figure 9 . The latter presents, in connection with the q=29 principal components of ranks 1 to 29 whose scores PC1 to PC29 for a plurality of voxels of interest are illustrated by the figure 6 , the set Σ r< of informative indicators calculated in 131b and ordered according to the respective ranks of the q=29 principal components produced in step 120. The values ​​of said informative indicators describe a curve “σj r<” which increases substantially when the rank of the principal components increases to reach a plateau σp r< of the order of ninety percent in the example illustrated by the figure 9 from rank 12. The value of convergence or plateau σp r< depends on the strength of the smoothing that has been applied to the score images of the figure 8 and the characteristic used to produce the informative indicators. As previously observed in connection with the figure 6 , the component of rank j=1 seems to express more spatial information than the components of ranks j greater than 10. Step 132 thus makes it possible, according to different embodiments, to objectify the selection of the first k principal components to obtain the desired compromise. Thus, according to a first embodiment, such a sub-step 132 may consist of calculating the mathematical difference between values ​​of said informative indicators Σ r< ={σ1 r< , ..., σq r<} when the latter are ordered according to the ranks j of the q principal components produced. The rank k can be determined from such a difference between two informative indicators of consecutive ranks.Thus, as soon as said difference becomes less than or equal (in absolute value) to a determined threshold, for example a threshold of a value less than three hundredths, the rank k sought is that of the principal component of rank immediately lower than that of the principal component for which the informative indicator is substantially equal (difference substantially zero or lower than said threshold) to that of the principal component of rank immediately higher than it. Such an approach is relevant when the informative indicators, ordered according to the ranks of the principal components respectively associated with them, describe a curve "σj r<" which is substantially increasing until reaching a plateau σp r<.As an alternative or in addition to the calculation of a mathematical difference between two informative indicators of consecutive ranks, the invention provides that it is possible to calculate an average mathematical difference between an increased collection of consecutive indicators, a difference between the informative indicators of ranks lower and / or higher than a given disconfirming indicator or any other equivalent technique. However, as indicated by the curve "σj r<" illustrated in . figure 9 , this may have one or more levels. In this case, a first level is described by the values ​​of the informative indicators σ7 r< and σ8 r< of the principal components of ranks 7 and 8. A second level is described by the values ​​of the informative indicators σ12 r< to σ29 r< of the principal components of ranks greater than or equal to 12. In order not to determine a rank k that is too low linked to the existence of an intermediate level for which the derivative of the curve “σj r<” is substantially zero, such a calculation of the derivative may be accompanied or combined with the calculation of a deviation (described by the curve “σp r< - σj r<” on the figure 9 ) with the informative indicator associated with the principal component of the highest rank in this case the rank q=29, or even the informative indicator of maximum value. If said deviation “σp r< - σj r<” is very small (for example less than three hundredths), then step 132 considers that the curve has reached the asymptote or the plateau σp r< . The rank k is thus determined as being that of the principal component of the immediately lower rank. This is for example the case of the component of rank k=11 which is the last presenting an informative indicator σ11 r< “escaping” the plateau σp r< or, more precisely, whose value deviates substantially from the value σp r< of said plateau.On the other hand, if the said deviation "σp r< - σj r<" is significant, for example greater than ten percent of the value σp r< , (case of the principal component of rank 7 whose informative indicator σ7 r< is substantially equal to that σ8 r< of the principal component of the immediately higher rank in the example illustrated by the . figure 9 ), then such a level of the curve “σj r<” is ignored.

[0054] Alternatively, such a sub-step 132 may consist of detecting a plateau σpr, as mentioned previously, described by the values ​​of the informative indicators σjr (j being between 1 and q) by traversing the latter from the highest rank of the principal components, in this case rank q=29, towards rank 1. By hypothesis, such a plateau σpr exists for the principal components of the highest ranks, in this case for the figure 9 , the plateau σpr is of the order of ninety percent. The rank k sought is determined as being that of the first principal component whose value of the informative indicator σkr deviates from said plateau σpr in a significant manner (i.e. beyond a predetermined threshold, for example five percent of the value of said plateau). In this case, the informative indicator σ11r is the first to describe such a value sufficiently distinct from the value of said plateau σpr. The rank k sought is therefore that of said principal component of k equal to eleven.

[0055] The invention cannot be limited to such operations implemented within the framework of sub-step 132 to detect the value of rank k from the ordered set of informative indicators Σ r< .

[0056] The invention further provides for being able to modify the k principal components thus selected in step 130 (symbolized by the vector Φ' in figure 7 ), prior to the implementation of step 140. The objective is to increase the contributions of the most “informative” principal components (whose ranks are the lowest) with respect to those whose respective capacities to express spatial information are lower, those of higher ranks. For this, such a step 130 may further comprise a sub-step 133 of applying a filtering operation of the k selected principal components, the filtering intensity of which is specific to each of the k principal components of ranks 1 to k less than or equal to q) selected and proportional to the informative indicator σ1 r< to σk r< of the latter.Such a filtering operation 133 may consist of the use alone or in combination of a Gaussian filter, an average filter, a median filter, an anisotropic diffusion filter, a wavelet filter or any filter, spatial or otherwise, that can be used in this context. As a preferred but non-limiting example, such a sub-step 133 may consist of the adjustment of weights wj r< specific to the principal components of ranks j carried out from the values ​​σj r< . The weights, in this example, may be multiplied linearly by the strength A of a Gaussian filter. Each selected component, i.e. of rank j, j being less than or equal to k, is filtered by a factor of wj r< x A. As a variant, or complement to linear multiplication, the weights can be previously expressed in the form of inputs to a logarithmic, exponential or polynomial function f.In this way, each selected score image PCj, i.e. of rank j, j being less than or equal to k, can be filtered by a factor defined by f(wj r< ) x A. Finally, the "filtered" score images are used to reproduce the score matrix U. As illustrated in . figure 15 , such a weight wj r< can be calculated from the curve “σj r<” illustrated in figure 9 , as follows: wj r = σj r − σ 1 r σp r − σ 1 r where σj r< is the value of the informative indicator of the principal component of rank j, σ1 r< is the value of the informative indicator of the principal component of rank 1 and σp r< the value of the plateau of the curve “σj r<” described by said informative indicators Σ r< . Thus, such a weight wj r< is between w1 r< =0 and wp r< =1 (i.e. when the curve “σj r<” has reached a plateau σp r< ).

[0057] The invention cannot be limited to these sole examples of calculations of the weights wj r< of sub-step 133.

[0058] THE figures 10 à 12 illustrate the benefit of implementing a method for attenuating the noise of experimental data resulting from multiple acquisitions by a magnetic resonance imaging device in the field of HARDI diffusion imaging. Such acquisition of experimental data can thus relate to a plurality of slices, in this case a human brain, for a plurality of phases. figure 10 thus illustrates the raw data, for a given tranche, of phases two, twelve, twenty-two, thirty-two, forty-two, fifty-two and sixty-two, respectively referenced P2, P12, P22, P32, P42, P52 and P62 on the figure 10 . On the first series of images referenced L1, the raw experimental data appear, therefore noisy for a plurality of voxels and respectively for the different phases P2 to P62 mentioned previously. The second and third series of images L2 and L3 illustrate said experimental data denoised by the implementation of the invention, depending on whether the k selected principal components have been filtered (step 133 of a method according to the figure 7 ) prior (L3) or not (L2) to the projection of the experimental data on said k principal components. We can visually observe the increasing gain brought by the invention from the L1 series to the L3 series.

[0059] Furthermore, the performance of a method 100 according to the invention can be illustrated by the figure 11 which describes the noise removed in step 140 of a method according to the invention, for the different phases referenced P2, P12, P22, P32, P42, P52 and P62 on the figure 10 , in connection with the denoised experimental data of the L2 and L3 series of the figure 10 . It is clear that such noise does not reveal any spatial information of interest.

[0060] There figure 12 describes the exploitation of HARDI diffusion data described in connection with the figure 10 as part of a tractography. The figure 12 presents three images I1, I2, I3 in color, translated into gray level, for a given slice, as well as a partial enlargement PI1, PI2 and PI3 of each image I1, I2, I3. The latter were produced in one step 150 from the respectively raw experimental data (series L1 on the figure 10 ), of data denoised by the implementation of the invention and as illustrated by the L2 series on the figure 10 and by the L3 series on the said figure 10 . The said enlargements PI1, PI2 and PI3 allow to see, with the naked eye, an increasing highlighting of fibers. On the enlargement PI1, an area (circled in white) of the organ could be considered as dead, which is not the case on the enlargements PI2 and PI3. In addition, such an exploitation allows to count such fibers. From the raw data of the L1 series of the figure 10 , it is possible to count around twenty-seven thousand fibers. They are around thirty-four thousand if denoised data forming the L2 series of the figure 10 are exploited and greater than thirty-seven thousand if the denoised data forming the L3 series of the figure 10 are exploited. Denoising of experimental data according to the invention thus provides a particularly significant gain.

[0061] THE figures 13 And 14 illustrate, like the figures 10 And 11, the gain provided by the invention from perfusion data. Thus, the acquisition of experimental data relates to a plurality of slices, in this case a human brain, for a plurality of phases. The figure 13 thus illustrates the raw data, for a given tranche, of phases one, five, ten, fifteen, twenty, twenty-five, thirty, thirty-five and forty, respectively referenced P1, P5, P10, P15, P20, P25, P30, P35 and P40 on the figure 13 . On the first series of images referenced L1, the raw experimental data appear, therefore noisy for a plurality of voxels and respectively for the different phases P1 to P40 mentioned previously. The second and third series of images L2 and L3 illustrate said experimental data denoised by the implementation of the invention, depending on whether the k selected principal components have been filtered (step 133 of a method according to the figure 7 ) prior (L3) or not (L2) to the projection of the experimental data on said k principal components. We can visually see the increasing gain brought by the invention from the series L1 towards the series L3. This gain is all the more notable, when we examine on said figure 13 , the spectrum of the experimental data, for a voxel V of interest, respectively raw (A), denoised by the implementation of the invention when the k selected principal components have not been filtered (B) and when said filtering has been carried out (C). We can notice that the curves described respectively by said spectra are increasingly smooth.

[0062] Furthermore, the performance of a method 100 according to the invention can be illustrated by the figure 14 which describes the noise removed in connection with the L2 and L3 series of the figure 13 . It is clear that such noise does not reveal any spatial information of interest.

Claims

1. Method (100) for attenuating the noise in experimental data (Zi) resulting from multiple acquisitions by a magnetic resonance imaging device (1) in relation to an elementary volume of interest, hereinafter referred to as a "voxel" of interest, among a plurality of voxels, said method (100) being implemented by a processing unit (4) of a medical imaging analysis system (S), said method (100) comprising: - a step (110) of collecting said experimental data (Z1,... , Zi, ... , Zn) in relation to the plurality of voxels (V1, ... , Vn); - a step (120) of producing an ordered set of q main components, according to their respective eigenvalues, and their respective scores for the plurality of voxels (V1, ..., Vn), said scores corresponding to the projection of said experimental data onto said q main components; - a step (130) of selecting the k first main components from among the q main components produced; - a step (140) of projecting said experimental data (Zi) onto the k selected main components and producing noise-attenuated experimental data (Zi') in relation to the elementary volume of interest; said method being characterized in that the step (130) of selecting k main components (130) of said method (100) comprises: - a sub-step (131) of producing main component score images for the plurality of voxels (V1, ... , Vn) and informative indicators (σ1r, ... , σqr) quantifying the spatial information contained in said images of the scores associated with said main components in the form of a rate of decrease of the values of a determined characteristic of said scores before and after the application of a smoothing filter to said scores, said determined characteristic of the scores for a plurality of voxels (V1, ... , Vn) belonging to a set of characteristics, taken alone or in combination, including variance, standard deviation, median and mean; - a sub-step (132) of determining the rank k in order to select the k first main components from said informative indicators (σ1r, ... , σqr) produced.

2. Method (100) for attenuating the noise in experimental data (Zi) resulting from multiple acquisitions by a magnetic resonance imaging device (1) in relation to an elementary volume of interest, hereinafter referred to as a "voxel" of interest, among a plurality of voxels, said method (100) being implemented by a processing unit (4) of a medical imaging analysis system (S), said method (100) comprising: - a step (110) of collecting said experimental data (Z1, ... , Zi, ... , Zn) in relation to the plurality of voxels (V1, ... , Vn); - a step (120) of producing an ordered set of q main components, according to their respective singular values, and their respective scores for the plurality of voxels (V1, ... , Vn), said scores corresponding to the projection of said experimental data onto said q main components; - a step (130) of selecting the k first main components from among the q main components produced; - a step (140) of projecting said experimental data (Zi) onto the k selected main components and producing noise-attenuated experimental data (Zi') in relation to the elementary volume of interest; said method being characterized in that the step (130) of selecting k main components (130) of said method (100) comprises: - a sub-step (131) of producing main component score images for the plurality of voxels (V1, ... , Vn) and informative indicators (σ1r, ..., σqr) quantifying the spatial information contained in said images of the scores associated with said main components in the form of a rate of decrease of the values of a determined characteristic of said scores before and after the application of a smoothing filter to said scores, said determined characteristic of the scores for a plurality of voxels (V1, ... , Vn) belonging to a set of characteristics, taken alone or in combination, including variance, standard deviation, median and mean; - a sub-step (132) of determining the rank k in order to select the k first main components from said informative indicators (σ1r, ... , σqr) produced.

3. Method (100) according to claim 1 or 2, for which the sub-step (132) of determining the rank k consists in calculating the mathematical difference between values of said informative indicators (σ1r, ... , σqr) when the indicators are ordered according to the q main components produced, with rank k being that of the main component for which said mathematical difference becomes less than or equal to a given threshold.

4. Method (100) according to claim 1 or 2, for which the sub-step (132) of determining the rank k consists in detecting a plateau (σpr) described by the values of the informative indicators (σ1r, ... , σqr) when the indicators are ordered according to the q main components produced, from the informative indicator (σqr) of the main component of rank q to the informative indicator (σ1r) of the main component of rank 1, with rank k being that of the first main component of which the informative indicator value (σkr) deviates from said plateau (σpr).

5. Method (100) according to any one of the preceding claims, comprising a sub-step (133) of applying a filtering operation to the k selected main components, the filtering intensity of which is specific to each of the k selected main components and proportional to the informative indicator (σ1r, ... , σ11r) of said component.

6. Method (100) according to the preceding claim, for which the filtering operation (133) consists in the single or combined use of a Gaussian filter, a mean filter, a median filter or an anisotropic diffusion filter.

7. Computer program product comprising one or more program instructions which can be interpreted by the processing unit (4) of a medical imaging analysis system (S), said program instructions being loadable into a non-volatile memory of said medical imaging analysis system (S), characterized in that the execution of said instructions by said processing unit (4) causes the implementation of a method (100) according to any one of the preceding claims.

8. Computer-readable storage medium containing instructions for a computer program product according to the preceding claim.

9. Medical imaging analysis system (S) comprising a processing unit (4) arranged for attenuating the noise in experimental data (Zi) resulting from multiple acquisitions by a magnetic resonance imaging device (1) in relation to an elementary volume of interest, hereinafter referred to as a "voxel" of interest, among a plurality of voxels, said processing unit (4) being configured to: - collect said experimental data (Z1, ... , Zi, ... , Zn) in relation to the plurality of voxels (V1, ... , Vn); - produce an ordered set of q main components, according to their respective eigenvalues, and their respective scores for the plurality of voxels (V1, ... , Vn), said scores corresponding to the projection of said experimental data onto said q main components; - select k first main components from among the q main components produced; - project said experimental data (Z1, ... , Zi, ... , Zn) onto the k selected main components and produce noise-attenuated experimental data (Z1', ... , Zi', ... , Zn') in relation to the plurality of voxels (V1, ... , Vn). said system (S) being characterized in that the processing unit (4) is configured to: - produce main component score images for the plurality of voxels (V1, ... , Vn) and informative indicators (σ1r, ... , σqr) quantifying the spatial information contained in said images of the scores associated with said main components in the form of a rate of decrease of the values of a determined characteristic of said scores before and after the application of a smoothing filter to said scores, said determined characteristic of the scores for a plurality of voxels (V1, ... , Vn) belonging to a set of characteristics, taken alone or in combination, including variance, standard deviation, median and mean; - select the k first main components from said informative indicators produced.

10. Medical imaging analysis system (S) comprising a processing unit (4) arranged for attenuating the noise in experimental data (Zi) resulting from multiple acquisitions by a magnetic resonance imaging device (1) in relation to an elementary volume of interest, hereinafter referred to as a "voxel" of interest, among a plurality of voxels, said processing unit (4) being configured to: - collect said experimental data (Z1, ... , Zi, ..., Zn) in relation to the plurality of voxels (V1, ..., Vn); - produce an ordered set of q main components, according to their respective singular values, and their respective scores for the plurality of voxels (V1, ... , Vn), said scores corresponding to the projection of said experimental data onto said q main components; - select k first main components from among the q main components produced; - project said experimental data (Z1, ... , Zi, ... , Zn) onto the k selected main components and produce noise-attenuated experimental data (Z1', ... , Zi', ... , Zn') in relation to the plurality of voxels (V1, ... , Vn). said system (S) being characterized in that the processing unit (4) is configured to: - produce main component score images for the plurality of voxels (V1, ... , Vn) and informative indicators (σ1r, ... , σqr) quantifying the spatial information contained in said images of the scores associated with said main components in the form of a rate of decrease of the values of a determined characteristic of said scores before and after the application of a smoothing filter to said scores, said determined characteristic of the scores for a plurality of voxels (V1, ... , Vn) belonging to a set of characteristics, taken alone or in combination, including variance, standard deviation, median and mean; - select the k first main components from said informative indicators produced.

11. Medical imaging analysis system (S) according to claim 9 or 10, comprising a program memory containing the program instructions of a computer program product according to claim 7.