Method for attenuating noise in images resulting from multiple acquisitions by magnetic resonance imaging

Abstract

The invention relates to a method (100) for denoising experimental data (Z1, ..., Zi, ..., Zn) resulting from multiple acquisitions by a magnetic resonance imaging device (1) by analysis (120) of selected (130) principal components, obtaining a better compromise between the efficiency of the denoising and preserving relevant information in the experimental data under consideration during its reconstruction (140) to generate denoised experimental data (Z1, ..., Zi, ..., Zn). The selection criterion is based on an informative index defining the amount of spatial information contained in the scored images associated with said principal components. The invention further provides the possibility of making available a step of applying an adaptive filtering (133) that filters out persistent spatial noise associated with each selected principal component.

Description

Technical Field

[0001] The present invention relates to a method for attenuating (or "denoising") the noise of images obtained by multiple acquisitions with a magnetic resonance imaging apparatus.

Background Art

[0002] Magnetic resonance imaging, also known by the abbreviation MRI, is based on the analysis of the response of protons in water molecules when the protons of water molecules are excited by a magnetic field. This response depends on the environment of such protons and makes it possible to distinguish between various types of tissues. A nuclear magnetic resonance imaging apparatus 1 of an imaging analysis system S, as shown by way of non-limiting example in FIGS. 1 and 2, is generally used. This provides a plurality of digital image sequences 12 of one or more parts of a patient's body, by way of non-limiting example, the brain, the heart, the lungs. For this purpose, the apparatus 1 applies a combination of high-frequency electromagnetic waves to the part of the body under consideration and measures the signals returned by specific atoms such as hydrogen for nuclear magnetic resonance imaging, by way of non-limiting example. Thus, the apparatus makes it possible to determine the magnetic properties, and as a result the chemical composition, and thus the nature, of the imaged volume, generally in each of its basic volumes called voxels. The magnetic resonance imaging apparatus 1 is controlled using a console 2. Thus, a user 6, for example, an operator, a doctor, or a researcher, can select a command 11 for controlling the apparatus 1 based on parameters or set points 16 entered by the input human-machine interface 8 of the analysis system. Such a human-machine interface 8 can consist, for example, of a computer keyboard, a pointing device, a touch screen, a microphone, or, more generally, any interface provided to convert gestures or commands given by a human 6 into control or setting data. Based on the information 10 generated by the apparatus 1, a plurality of digital image sequences 12 of a part of a human or animal body are obtained. Such information 10 or images 12 will be referred to as "experimental data" without distinction.

[0003] The image sequence 12 can optionally be stored in the server 3, i.e., in a computer equipped with its own storage means, and can constitute the medical file 13 of a patient. Such a file 13 can include different types of images, such as functional images that emphasize the activity of tissues or anatomical images that show the characteristics of tissues. The image sequence 12 or more generally the experimental data is analyzed by the processing unit 4, which is provided for this purpose. Such a processing unit 4 can consist of, for example, one or more microprocessors or microcontrollers that execute appropriate application program instructions placed in the storage means of the imaging analysis system S.

[0004] "Storage means" means any volatile or preferably non-volatile, computerized storage device. A non-volatile memory mechanism is a computerized storage device that, by its technology, can retain its data without a power supply. It can include data resulting from inputs, calculations, measurements, and / or program instructions. The main currently available non-volatile memory mechanisms are of the electrically writable type such as EPROM (erasable programmable read-only memory), or of the more electrically writable and erasable type such as EEPROM (electrically erasable programmable read-only memory), flash, SSD (solid state drive), etc. A non-volatile memory mechanism is distinguished from a memory mechanism known as "volatile" in which data is lost in the absence of a power supply. The main currently available types of volatile memory mechanisms are RAM (random access memory, also called "read-write memory"), DRAM (dynamic random access memory that requires periodic refreshing), SRAM (static random access memory that requires such refreshing in the event of a power drop), DPRAM or VRAM (especially suitable for video), etc. Hereinafter, the "data storage device" can be volatile or non-volatile depending on the application in question.

[0005] The processing unit 4 includes means for communicating with the external world to collect images. Moreover, by means of said communication means, the processing unit 4 can finally provide, to the user 6 of the imaging analysis system S via the output human-machine interface 5, a description, for example a graphic and / or audible description, of the estimated value of the biomarker or the determined quantity calculated by said processing unit 4 based on the experimental data 10 and / or 12 obtained by magnetic resonance imaging. Throughout this document, "output human-machine interface" is used alone or in combination and means any device capable of outputting or providing to the user 6 of the magnetic resonance imaging analysis system S a reconstructed physiological signal, in this particular case of a biomarker, a graphic, haptic, or sound display, or more generally something perceptible to humans. Such an output human-machine interface 5 can consist, without being exhaustive, of one or more screens, loudspeakers, or other suitable alternative means. Thus, the user 6 of the imaging analysis system S can confirm that the diagnosis is correct or invalidate the diagnosis, decide on a treatment action judged appropriate, go deeper in research activities, refine the parameters for adjusting the measuring device, etc. Optionally, this user 6 can also set the operation of the processing unit 4 or the output human-machine interface 5 by means of the operation and / or acquisition parameters 16. Thus, for example, it is possible to define a display threshold or select the biomarker, indicator, or parameter for which the estimated or determined quantity is desired to be displayed. For this purpose, the user uses the aforementioned input human-machine interface 8 or a second input interface provided therefor. Advantageously, the input human-machine interface 8 and the output human-machine interface 5 can constitute just one and the same physical entity. Also, the input human-machine interface 8 and the output human-machine interface 5 of the imaging analysis system can be integrated into the acquisition console 2. There are variants described with respect to Figure 2.Regarding that, the imaging system as described above further includes a preprocessing unit 7 that analyzes the image sequence 12, then estimates the experimental signal 15, and thus sends the experimental signal 15 to the processing unit 4 that is released from this task.

[0006] With or without adjusting and changing the parameters of the medical imaging system S, several two-dimensional images or three-dimensional volumes can be acquired. For example, in order to examine the reaction timing immediately after the arrival of the contrast agent, it is possible to simply repeat the acquisition without adjusting and changing the parameters of the system. On the other hand, other imaging techniques require changing the adjustment and / or parameters in order to sample the desired parameter space.

[0007] Among the techniques or methods based on magnetic resonance imaging, chemical exchange saturation transfer (CEST) imaging is well-known. Such a technique consists of applying high-frequency pulses according to various defined resonance frequencies ω so that the chemical species of interest reaches a saturation state. The defined resonance frequencies ω are actually each related to various chemical species of interest, such as, without limitation, an amide group (-NH), an amine group (-NH2), or a hydroxyl group (-OH). Such excited unstable hydrogen protons are exchanged with the unexcited hydrogen protons of water. Such application of high-frequency pulses at a defined resonance frequency ω different from the resonance frequency ω0 related to water, continuously repeated for a defined duration, for example, several seconds, accumulates water saturation of the chemical species of interest. Its concentration can be indirectly measured by the decrease in the signal of water, a phenomenon called the "CEST effect". The CEST effect is analyzed in the form of a Z-spectrum. Such a decrease can be easily detected by a fast magnetic resonance imaging acquisition sequence such as, without limitation, "single-shot echo planar imaging". FIG. 3 shows an example of a Z-spectrum in the form of a set (or dataset) of samples Zi(Δω) according to the relative frequency shift Δω with respect to the frequency of water for a voxel V of interest, such a shift Δω being expressed in ppm. More specifically, FIG. 3 shows various images M(-6ppm), M(-3.5ppm), M(0ppm), M(3.5ppm), M(6ppm) corresponding to frequency shifts Δω equal to -6ppm, 3.5ppm, 0ppm, 3.5ppm, and 6ppm respectively. It also shows an image M0 acquired at a frequency far from the frequency of water. The Z(Δω) spectrum, which can be expressed as "normalized", corresponds to the ratio of M(Δω) to M0 for frequency shifts included between -6ppm and 6ppm for a given voxel. If the samples Zi(Δω) are ordered according to increasing Δω as shown in FIG. 3, the Z-spectrum can be advantageously shown in the form of a discrete signal showing m measured values of the magnitude of the experimental signal provided by the measuring device 1 with respect to the basic volume of the organ, normalized by the signal measured without high-frequency saturation and expressed as a percentage.According to FIG. 3, this set of m samples constitutes a discrete Z signal that substantially traces a "V" when these samples are ordered according to the increasing frequency shift Δω, and its minimum value Zi(Δω)m is associated with the frequency shift Δω = 0 ppm when the device is fully adjusted and / or the static magnetic field is uniform.

[0008] CEST technology improves the detection of specific metabolites in the human body. Their concentration is insufficient to detect them with conventional magnetic resonance imaging sequences. Therefore, this CEST technology can provide very valuable information for physicians who seek to establish a diagnosis and make treatment decisions in the treatment of lesions.

[0009] Diffusion tensor imaging (DTI) may also be mentioned. In DTI, data is acquired for various strengths (b-values) and directions of magnetic diffusion gradients. This category includes perfusion-weighted imaging (PWI), as well as two of its main techniques, dynamic susceptibility contrast (DSC) and dynamic contrast-enhanced (DCE). There is also high angular resolution diffusion imaging (HARDI) technology.

[0010] In magnetic resonance imaging, the signal induced in the receive coil is a continuous complex signal. That is, it generally includes a real part and an imaginary part acquired within a frequency domain commonly referred to as "k-space". The noise on such a signal can be said to be a complex contribution added to and uncorrelated with the pure signal. Thus, such noise includes real and imaginary components that are independent of each other and are similarly distributed according to a Gaussian distribution with zero mean and standard deviation (Cardenas-Blanco et al. 2008; Aja-Fernandez and Vegas-Sanchez-Ferrero 2016, see for example the technical teachings accessible by the hyperlink "https: / / doi.org / 10.1007 / 978-3-319-39934-8"). However, in this field, it is usually of interest to consider the absolute value image rather than the complex signal. Taking into account that the conversion from the complex signal to the absolute value signal recorded in the absolute value image is non-linear, the distribution of the intensities of the pixels in the resulting image is generally not a Gaussian distribution, and under certain conditions, it can be modeled by a Rice distribution as detailed in the document Cardenas-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, which can be examined by the link "https: / / doi.org / 10.1002 / cmr.a.20124".

[0011] In the field of magnetic resonance imaging, it is known to reduce or remove the noise of experimental data (i.e., measurement data) or the noise from experimental data. The term "denoise" is used to represent this operation that consists of excluding or removing the noise in order to take into account only the important information within the experimental data. For this purpose, there are numerous methods that use principal component analysis (PCA).

[0012] Principal component analysis is originally a non-parametric technique (i.e., not related to any model) designed to reduce the dimension of a dataset while retaining the essence of the data. Such an approach is disclosed, for example, in the document Jolliffe, I. 2002. “Principal Component Analysis” Second Edition, Springer-Verlag, which can be accessed by the link “https: / / doi.org / 10.1007 / b98835”. Such a method finds and uses the linear correlations within the data under consideration and extracts from them a new uncorrelated dataset of smaller size called “principal components”. These principal components thus extracted are ordered according to the variability (or variance) they represent. Therefore, the first principal component represents the most informative part of the data that captures most of the data variance.

[0013] Put simply, when applied to imaging, the purpose of principal component analysis is to “re-represent” the experimental data with noise and find the most meaningful basis to construct a denoised experimental dataset, i.e., one from which the noise has been removed to reveal one or more structures of interest that may have been hidden by the said noise.

[0014] From a geometric perspective, the problem of PCA can be simplified as follows. That is, for a given set of (real) points in a space with m dimensions, - the j-th vector represents the line of best fit passing through the points, - find q vectors for the j-th vector that are orthogonal to the previously calculated j - 1 vectors.

[0015] "Best fit" means finding a line or axis that minimizes the mean squared distance from points to the line. Such a fit is shown in Figure 4. In the left part of this figure, a scatter plot can be seen within a two-dimensional coordinate system or basis represented by axes x1 and x2. In the right part, it can be seen that axis z1 is the line with the best fit passing through the said points. The data or points can be projected onto a new coordinate system or basis represented by axes z1 and z2 perpendicular to z1. The principal component z1 is characterized by a larger variance λ1, also called an "eigenvalue". Thus, the set Λ of the eigenvalues λ1 and λ2 of the principal components z1 and z2 respectively represents the amount of information collected by the said principal components or their respective information-providing capabilities. The principal components z1 and z2 in Figure 4 are called "eigenvectors" Φ = {φ1, φ2}.

[0016] Therefore, principal component analysis - experimental data of dimension m can be represented in the form of a linear combination of vectors, - the variance λ of each principal component indicates meaningful information, - and it naturally follows that the q principal components that form a new basis or a new coordinate system are orthogonal.

[0017] Figure 5 shows a known method for attenuating the noise of experimental data in the form of a normalized Z-spectrum (having m phases or frequencies respectively) for a collection of n spatial positions (also called "voxels") in the human brain or "denoising" the experimental data. Thus, for a given voxel, the experimental data is a vector having m acquired values or samples. Such a method 100 includes steps 110, 120, 130, 140, and 150.

[0018] - In the first step 110, the experimental data Z1 to Zn for a set of n voxels V1 to Vn having m different phases (or samples) are obtained or collected from the voxel data in the form of a collinearity matrix C (an n-by-m matrix where each of the n rows contains m values obtained for a given voxel). The n rows are each dedicated to a different voxel of the segment or volume under consideration.

[0019] - In step 120, for this experimental data set with noise, in this case, for a given voxel, q = m principal components that best represent the Zi spectrum are

Number

Number

[0020] - In step 130, the optimal number k < q of principal components is determined, or more specifically, k components that are the most meaningful or informative are selected from among the q extracted components, i.e., k components that represent the spatial information of interest as opposed to the q - k other components that should be discarded and mainly represent noise.

[0021] - In step 140, the experimental data with noise is projected onto the remaining k components to form a new denoised experimental data set. In this case, in the example of Figure 5, for the voxels of interest, the Zi' spectrum is

Number

[0022] - In step 150, in the original space of voxels V1 to Vn, the denoised experimental data Z1’ to Zn’ are rearranged, and all or part of the denoised experimental data Z1’, …, Zi’, …, Zn’ are used according to the selected application.

[0023] As an example, as shown in FIG. 5, the Zi spectrum with noise of the voxel under consideration appears in the form of a “smoother” curve Zi’ than the curve represented by the “raw” experimental data Zi after such a method 100 is executed.

[0024] Step 120 has been described based on the covariance matrix according to the first embodiment. In a variant, the generation of such q principal components can consist of the singular value decomposition (SVD) of the above-described co-occurrence matrix C according to the second embodiment. According to this second embodiment, the experimental data in the form of, for example, the Zi spectrum for a given voxel is

Equation

[0025] The singular values of matrix S are classified in decreasing order. Note that if matrix C is transposed, where its rows represent m variables and its columns represent n samples, the interpretations of matrices U and V are swapped respectively.

[0026] By examining the terms of the aforementioned equation, T = US is a matrix of the scores of the principal components multiplied by the singular values, and V t can be reached as being the matrix of the principal components. Such a second embodiment of step 120 mainly focuses on the analysis of the spatial information contained in each column of matrix U or, similarly, matrix T. As described above, such a step 120 is based on the covariance matrix [Number] obtained therefrom. In this regard, Φ represents the eigenvectors of the principal components and Λ represents their eigenvalues. The q components extracted or calculated, and thus their respective eigenvectors φ1 to φq, are generally classified or ordered according to their respective variances or eigenvalues λj. There is a [Number] direct relationship as follows.

[0027] For experimental data, in this case, to attenuate the noise of the Zi spectrum with noise, the "denoised" experimental data projects the noisy experimental data onto k remaining components to obtain, for the voxels of interest, a new denoised experimental data set, in this case, in the example of Figure 5, the Zi' spectrum, [Number] by performing step 140 configured as follows. In this regard, ti' is a vector of k scores, and V t ' is a k-row q-column matrix of the principal components.

[0028] Among the steps of denoising by principal component analysis, step 130 of selecting components requires the most attention. In order to obtain the spatial information hidden in the principal components, as shown in FIG. 6, each principal component (i.e., the real value of the voxel projected onto the eigenvector) can be remodelled within its original range (i.e., in the volume or slice). The images PC1 to PC29 related to the principal components from rank 1 to 29 respectively represent the scores related to the eigenvectors φ1 to φ29 for the set of voxels V1 to Vn under consideration according to the method used in the generation 120 of the principal components. Such scores correspond to the projection of the experimental data onto the 29 principal components. For the sake of convenience, the term "score" will be used to represent the contribution of the principal component, regardless of the method of generation or extraction 120 of such principal components. According to the prior art, the principle of principal component analysis is, in a sense, equivalent to trying to best represent the experimental signal or data of the set of voxels resulting from multiple acquisitions in the form of a linear combination of principal components according to each score. The principal components are classified according to the order of superiority directly arising from or related to the eigenvalues of the principal components. Therefore, the image PC1 in FIG. 6 depicts the score of the first principal component, like an image of a slice of the human brain, in order to reconstruct the experimental curve (the Zi spectrum for i between 1 and n) for the n voxels under consideration. Similarly, the image PC2 in FIG. 6 depicts the score of the second principal component, like an image of a slice of the human brain, in order to reconstruct the experimental curve of the voxels under consideration. In other words, the image PC2 in FIG. 6 depicts the experimental data Z1 to Zn projected onto the second principal component. The same applies to the images PC3 to PC29 representing the scores of the third to twenty-ninth principal components for each voxel under consideration. This FIG. 6 can show the result of executing the above-described step 120. As can be seen with the naked eye, the first nine images PC1 to PC9 related to the principal components from rank 1 to 9 respectively appear to represent appropriate spatial information while decreasing in degree, in contrast to the other images PC10 to PC29 related to the principal components of higher ranks. The images PC10 to PC29 seem to show nothing anymore.Of course, the selection 130 of the k meaningful principal components is not performed by the naked eye, but according to an automated method.

[0029] In the prior art, there are various models and processes or methods using mathematical assumptions or statistics, combinations with other technologies, and various criteria for determining the "optimal" subset of k components in many fields and applications. Among these, there are some that are generally used for the purpose of denoising or filtering in magnetic resonance imaging by reconstructing an experimental data set using components that provide information and excluding components mainly related to noise, and have been successful. For this purpose, k principal components out of the q extracted principal components that are appropriate to retain, and q - k principal components to be excluded during the reconstruction of the experimental data are determined, and numerous data-focused approaches have been proposed to obtain denoised experimental data.

[0030] However, most of the selection criteria are based on the set of eigenvalues Λ associated with the q principal components extracted, i.e., for each component, on the measure of the variance represented by the original dataset, or on the residuals, i.e., the respective mathematical differences between the original data and the reconstructed data. In this regard, non-exhaustively, 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 via 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 further 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 may be mentioned.These publications propose a method based on the eigenvalues of the principal components that relies on the Marinoffski, Nelson, and Median criteria and is applied to the denoising of in vivo CEST magnetic resonance imaging data.

[0031] 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 document Veraart, Jelle, Dmitry S. Novikov, Daan Christiaens, Benjamin Ades-aron, Jan Sijbers, and Els Fieremans - 2016 - “Denoising of Diffusion MRI Using Random Matrix Theory.” NeuroImage 142: 394 - 406, accessible via the link “https: / / doi.org / 10.1016 / j.neuroimage.2016.08.016”. Other methods are based on the amount of singular value reduction, as described in the document Ma, Xiaodong, Kamil Ugurbil, and Xiaoping Wu - 2020 - “Denoise Magnitude Diffusion Magnetic Resonance Images via Variance-Stabilizing Transformation and Optimal Singular-Value Manipulation.” NeuroImage 215 (July): 116852, accessible via the link “https: / / doi.org / 10.1016 / j.neuroimage.2020.116852”. The mathematical principle is described 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”.

[0032] Similar to the known method shown in FIG. 5, most of the methods proposed in the literature are based on the eigenvalues Λ to determine the k components to be retained and those to be excluded during data reconstruction. Thus, as suggested in FIG. 5, step 130 consists, in this case, of comparing the eigenvalues λj of the q principal components on a logarithmic scale, where j is included between 1 and q. If below a certain dispersion threshold, the principal components are considered to represent little or no appropriate information and are excluded. Only the first k principal components, i.e., those with a dispersion or eigenvalue greater than the threshold, are retained for use in step 140. These methods generally have a significant drawback in that they tend to ignore anatomical and pathological information hidden in some of the discarded principal components. This drawback is directly due to using the eigenvalues, which are measures of the dispersion of the different principal components, as the main selector and groping for noise and appropriate information. Principal component analysis generally starts from the principle that dispersion indicates the presence or absence of important information. It assumes that the experimental data being analyzed has a high signal-to-noise ratio, i.e., the power of the signal of interest is greater than the power of the noise. This is a strong and sometimes incorrect assumption, especially for experimental magnetic resonance image data with noise. Such a situation occurs, for example, when the noise level of the experimental data provided by a magnetic resonance imaging device is high, or when the signal of interest relates to a small anatomical or pathological region relative to the size of the image under consideration. Thus, most of the methods proposed in the literature are eigenvalue-based but discard the principal components containing hidden anatomical or pathological information of interest. This is because the contribution of the relevant components of these regions to the eigenvalues is small. Thus, a non-negligible amount of appropriate information is lost. This is because, according to the prior art, it is merged with the noise. As can be seen by referring to FIG. 12 as an example, it is not possible to select the principal components according to such criteria, which may be considered too simple or insufficient, to ensure the proper use of the experimental data and provide reliable indicators or biomarkers based on the denoised experimental data. SUMMARY OF THE INVENTION

[0033] The present invention addresses the drawbacks posed by the prior art. Among the numerous advantages achieved by the present invention, in particular, it may be mentioned that the method of denoising according to the present invention enables a better compromise between the efficiency of denoising and the maintenance of appropriate spatial information during the reconstruction of the experimental data under consideration, in order to generate denoised experimental data, and includes a step of selecting principal components. By way of example, the proposed selection criterion is based on the amount of reduction in the variance of an image having a score related to the principal components after application of a smoothing filter. By this technique, in all the principal components extracted or calculated, hidden spatial information, in particular anatomical or pathological information, can be used, and those that can represent important information can be selected regardless of each other and regardless of their ordering resulting from their extraction or generation, thus minimizing the risk of excluding appropriate principal components. Also, the present invention enables the use of a filtering step that can be evaluated as being adaptive, i.e., while eliminating persistent spatial noise related to each of the extracted principal components, and while adapting the level of filtering of the principal components to the amount of appropriate information that can be represented by each of them, the experimental data under consideration can be further denoised.

[0034] For this purpose, the present invention provides a method for attenuating the noise of experimental data resulting from multiple acquisitions by a magnetic resonance imaging device related to a basic volume of interest, hereinafter referred to as a "voxel" of interest, the voxel of interest being among a plurality of voxels, and the method being executed by a processing unit of a medical imaging analysis system. Such a method comprises - a step of collecting the experimental data related to a plurality of voxels, and - for a plurality of voxels, generating an ordered set of q principal components ordered according to their respective eigenvalues, and their respective scores, the scores corresponding to the projection of the experimental data onto the q principal components, and - selecting the first k principal components from among the q generated principal components; - projecting the experimental data onto the selected k principal components to generate experimental data with attenuated noise related to the basic volume of interest.

[0035] To optimize the denoising of experimental data without losing spatial information, the step of selecting k principal components of such a method - generating an information-providing index that determines the amount of spatial information contained in the score image related to the principal component in the form of a score image of the principal component for a plurality of voxels and a reduction rate of a defined characteristic value of the score before and after application of a smoothing filter to the score, where the defined characteristics of the score for the plurality of voxels belong to a set of characteristics including variance, standard deviation, median, and mean, either alone or in combination; - determining a rank k to select the first k principal components based on the generated information-providing index.

[0036] According to a first advantageous embodiment, the sub-step of determining rank k may consist of calculating the mathematical difference between the values of the information-providing index when the information-providing indices are sorted according to the generated principal components, and rank k is that of the principal component for which the mathematical difference is below a defined threshold.

[0037] In a variant, the sub-step of determining rank k may consist of detecting a flat portion depicted by the value of the information-providing index from the information-providing index of the principal component of rank q to the information-providing index of the principal component of rank 1, and rank k is that of the first principal component for which the value of the information-providing index deviates from the flat portion.

[0038] During the reconstruction of the experimental data thus denoised, in order to increase the weight of the dominant principal components, the step of selecting the k principal components of the method according to the present invention may include a sub-step of applying a filtering operation to the selected k principal components, and the filtering intensity of the filtering operation is specific to each of the selected k principal components and is proportional to its information providing index.

[0039] In this case, the filtering operation may consist of using a Gaussian filter, an average filter, a median filter, an anisotropic diffusion filter, alone or in combination.

[0040] According to a second subject, the present invention relates to a computer program, the computer program comprising one or more program instructions that can be interpreted by a processing unit of a medical imaging analysis system, the program instructions being able to be placed in a non-volatile storage device of the medical imaging analysis system, and by execution of the instructions by the processing unit, a method for attenuating the noise of experimental data according to the present invention is executed.

[0041] According to a third subject, the present invention relates to a computer-readable storage medium comprising instructions of such a computer program according to the present invention.

[0042] According to a fourth subject, the present invention relates to a medical imaging analysis system, the medical imaging analysis system comprising a processing unit provided to attenuate the noise of experimental data resulting from multiple acquisitions by a magnetic resonance imaging device related to a basic volume of interest, hereinafter called a "voxel" of interest, the voxel of interest being among a plurality of voxels, the processing unit - collecting the experimental data related to the plurality of voxels, - generating, for the plurality of voxels, an ordered set of q principal components ordered according to their respective eigenvalues, and their respective scores, the scores corresponding to the projection of the experimental data onto the q principal components, - Select the first k principal components from among the q generated principal components, - Project the experimental data onto the selected k principal components to generate experimental data with attenuated noise related to the basic volume of interest.

[0043] To optimize the denoising of experimental data without losing spatial information, the processing unit of such a system - Generate a score image of the principal components for a plurality of voxels and an information-providing index that determines the amount of spatial information contained in the score image related to the principal components in the form of the reduction rate of the determined characteristics of the scores before and after application of a smoothing filter to the scores, where the determined characteristics of the scores for the plurality of voxels belong to a set of characteristics including variance, standard deviation, median, and mean, either alone or in combination, - Select the first k principal components based on the generated information-providing index.

[0044] According to an advantageous embodiment, such a medical imaging analysis system may include a program memory containing program instructions of a computer program according to the present invention.

[0045] Other features and advantages will become clearer upon reading the following description and examining the accompanying drawings.

Brief Description of the Drawings

[0046]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

[0047] In accordance with the present invention, a method 100 for attenuation of noise in a medical image shown in FIG. 7 is advantageously represented in the form of a computer program, the program instructions of which are intended to be installed in the program memory of a component of a medical imaging system, such as the system S of FIGS. 1 and 2, such as a computer or a computer system server, or more generally any electronic object enabling the use of sufficient computing power.

[0048] Accordingly, FIG. 7 shows a method 100 according to the invention for attenuation of noise in experimental data resulting from multiple acquisitions by a magnetic resonance imaging device such as device 1 of the medical analysis system S shown in FIGS. 1 and 2, relating to a plurality of n voxels. Such a method includes some steps common to the method 100 according to FIG. 5. Accordingly, the "denoising" method 100 according to the invention may include, if necessary, a step 110 of collecting or acquiring experimental data with noise relating to or associated with the plurality of voxels V1 to Vn. Such experimental data may consist of curves such as the Z1 to Zn spectra after CEST-type acquisitions similar to the examples referred to in FIGS. 3 and 5, or more generally any experimental data for a set of voxels or pixels relating to the volume or slice of an organ related by a plurality of m acquisitions depending, for example, on different instants or different phases. Accordingly, such experimental data may non-exhaustively result from multi-phase high angular resolution diffusion imaging (HARDI) acquisitions or may be related to the field of perfusion imaging.

[0049] As a preferred but non-limiting example, similar to FIG. 5, FIG. 7 depicts the experimental data Zi related to the voxels of interest among the n voxels obtained. FIG. 7 shows the experimental data Zi in the form of a spectrum with the same noise as that shown in FIG. 5. Such noise is represented by more or less significant discontinuities or oscillations depending on the relevant frequency shift, as indicated by the partial enlarged view of the Zi spectrum. To attenuate this phenomenon, similar to the known method 100, as shown in FIG. 5, the method 100 according to the present invention includes step 120 of performing principal component analysis and generating or calculating q principal components, said step being similar to step 120 of the known method 100 according to FIG. 5. Such step 120 forms an ordered set of principal components, each having an eigenvalue λ1, …, λq or singular value s1 to s qconsisting of determining q principal components represented by or related to them, the rank of the principal components being determined according to a set of decreasing scalar values such as the eigenvalues Λ = {λ1, …, λq} or the singular values S. Such a method 100 according to the present invention further includes step 130 of determining or selecting k principal components out of the q generated principal components, i.e., the principal components from rank 1 to k, and determining a smaller set of principal components. According to the present invention, the selection 130 shown in FIG. 7 is clearly distinguished from the approach described in the majority of the prior art based on the eigenvalues Λ. Such step 130 specific to the present invention will be detailed below. The method 100 according to the present invention, similar to the known method 100 shown in FIG. 5, further includes step 140 of projecting the noisy experimental data Zi onto the k principal components selected in step 130 to generate attenuated or “denoised” experimental data Zi′ related to the basic volume of interest. By examining the denoised Zi′ spectrum representing a particularly smooth curve with respect to the curve represented by the original Zi spectrum, an improved performance of the method 100 according to the present invention can be observed thanks to the improved selection or modification of the k principal components before said projection 140. Similar to that shown in FIG. 5, step 140 of the method 100 according to FIG. 7 consists of projecting a set of noisy experimental data Z1, …, Zi, …, Zn onto the selected k principal components to generate attenuated experimental data Z1′, …, Zi′, …, Zn′ for the voxels V1 to Vn of interest. Thus, optionally and advantageously, such a method 100 according to the present invention may include step 150 of combining and using the denoised experimental data Z1′, …, Zi′, …, Zn′ for all or part of the plurality of said voxels V1 to Vn of interest. Such use may consist, for example, of counting fibers in tractography based on HARDI diffusion experimental data, as taken up with reference to FIGS. 10 to 12.

[0050] Accordingly, the present invention is mainly distinguished by the execution of step 130, which enables the selection of appropriate principal components, i.e., enables the optimization of noise attenuation without losing appropriate spatial information. Accordingly, as shown in FIG. 7, such a step 130 calculates, for each of the q generated principal components of ranks 1 to q, σ1 in FIG. 7 r , …, σq r and includes a sub-step 131 that generates an information-providing index referred to in FIG. 7. Such an information-providing index σ1 r to σq r set Σ r is a substitute for the set Λ of eigenvalues λ1 to λq that is directly used by a known method for selecting k principal components from among the q extracted principal components. Such an information-providing index σj r characterizes the ability of the principal component of rank j to represent appropriate spatial information. It may advantageously consist of the rate of decrease before and after the application of a smoothing filter to the value of a defined characteristic of the scores of the principal components related to a plurality of voxels V1 to Vn under consideration or of interest. Such an operation 131a is shown in FIG. 8. This depicts two images PC3a and PC3b showing the scores of the principal component of rank 3 for the set of voxels under consideration. Image PC3a depicts the scores before the application 131a of the smoothing filter, and image PC3b depicts these same scores after the application 131a of the smoothing filter. Accordingly, this image PC3b appears more blurred than image PC3a in FIG. 8. Such an operation 131a may advantageously consist of applying a Gaussian filter of a factor F having an arbitrarily settable predetermined value.

[0051] As a preferred but non-limiting example, the defined characteristic may consist of the standard deviation of the scores. And the information-providing index σj of the principal component of rank j r is, in operation 131b,[[]]

Equation

[0052] In this case, in FIG. 8, the information providing index σ3 of the principal component of rank 3 r is

Number

Number

Number

[0053] In this way, the set Σr of the information providing indices of the q principal components of ranks 1 to q generated in step 120 and ordered according to the respective eigenvalues Λ or singular values S can be constructed at the end of the execution of sub-step 131. In a variant, this set Σr of the information providing indices of the q principal components can be ordered according to the numerical values of the information providing indices.

[0054] In a variant or in addition, such defined characteristics σjr may use, instead of the standard deviation, the variance, entropy, median, or further the mean of the scores of the principal component of rank j, or may further result from all or some combinations of these.

[0055] In step 131 of the method 100 according to the present invention, when the set Σ r of the information providing indices is calculated, the set Σ rIncluding sub-step 132 for determining rank k to select the first k principal components (represented by the set Φ' = {φ1, …, φk} of eigenvectors or the matrix U' = {u1, …, uk} in FIG. 7). The execution of such sub-step 132 is shown in FIG. 9. FIG. 9 shows a set Σ of information-providing indicators calculated at 131b and ordered according to the respective ranks of the q = 29 principal components generated at step 120 for q = 29 principal components from rank 1 to 29 r is shown. The scores PC1 to PC29 of the q = 29 principal components from rank 1 to 29 for a plurality of voxels of interest are shown in FIG. 6. The values of the information-providing indicators draw a curve “σj r ” that increases substantially as the rank of the principal component increases, and reach an approximately 90% flat part σp r starting from rank 12 in the example shown in FIG. 9. The converged value of the flat part σp r depends on the strength of the smoothing applied to the score image of FIG. 8 and the characteristics used to generate the information-providing indicators. As previously mentioned with reference to FIG. 6, the component of rank j = 1 appears to represent more spatial information than the components of rank j greater than 10. Thus, step 132 can give objectivity to the selection of the first k principal components in order to obtain the required compromise according to different embodiments. Thus, according to the first embodiment, such sub-step 132 is the information-providing indicator Σ r when ordered according to the rank j of the q principal components for which the information-providing indicator is generated r , …, σq r} may consist of calculating the mathematical difference between the values. The rank k can be determined based on such a difference between two information-providing indicators of consecutive ranks. Thus, when the difference is below a predetermined threshold value (in absolute value), for example, a threshold value of less than 1 / 300, the required rank k is that of the principal component of the rank immediately below the rank of the current principal component. For the current principal component, the information-providing indicator is substantially equal to that of the principal component of the rank immediately above it (the difference is substantially zero or less than the threshold value). The information-providing indicators are ordered according to the ranks of the principal components associated with them respectively. The information-providing indicator is the flat part σpr draw a curve “σj r ” that increases substantially until it reaches, such an approach is appropriate. In a variant of the calculation of the mathematical difference between two information-providing indicators of consecutive ranks or in addition thereto, the present invention provides for calculating the mathematical difference of the average between increasing collections of consecutive indicators, the difference between information-providing indicators of ranks lower and / or higher than a given information-providing indicator, or any other equivalent method. However, as shown by the curve “σj r ” shown in FIG. 9, the curve may have one or more flat portions. In this case, the first flat portion is depicted by the values of the information-providing indicators σ7 r and σ8 r of the principal components of ranks 7 and 8. The second flat portion is depicted by the values of the information-providing indicators σ12 r to σ29 r of the principal components of ranks 12 and above. Such a calculation of the differential coefficient is accompanied by or combined with the calculation of the deviation (depicted by the curve “σp r ”-“σj r ” in FIG. 9) from the information-providing indicator associated with the highest-rank principal component, in this case rank q = 29, or the information-providing indicator of the maximum value, so as not to determine rank k as being too low in relation to the existence of an intermediate flat portion where the differential coefficient of the curve “σj r ” is substantially zero. If the said deviation “σp r -“σj r ” is very small (for example less than 1 / 300), in step 132, it is considered that the curve has reached an asymptote or a flat portion σp r . Thus, rank k is determined as that of the principal component of the immediately lower rank. This is the case, for example, for the component of rank k = 11. It is the last one having the information-providing indicator σ11 r that “escapes” the flat portion σp r , or, more specifically, its value substantially deviates from the value σp r of the said flat portion. On the other hand, if the said deviation “σp r -“σj r ” is significant, for example a value σp rwhen it is greater than 10 percent (in the example shown in FIG. 9, information providing index σ7 r is substantially equal to that of the principal component of the immediately higher rank, σ8 r and the principal component of rank 7), such a flat portion of the curve “σj r ” is ignored.

[0056] In a modified example, such a sub-step 132, as described above, may consist of detecting a flat portion σp r depicted by the value of the information providing index σj r (where j is included between 1 and q), from the highest rank of the principal components, in this case rank q = 29 to rank 1, by looking through these information providing indices. By assumption, such a flat portion σp r exists for the principal component of the highest rank, in this case, in FIG. 9, the flat portion σp r is approximately 90%. The required rank k is such that the value of the information providing index σk r is determined to be that of the first principal component that clearly deviates (i.e., exceeds a predetermined threshold, for example 5 percent of the value of the flat portion) from said flat portion σp r . In this case, the information providing index σ11 r is the first one that represents a value sufficiently different from the value of such said flat portion σp r . Therefore, the required rank k is that of the principal component of k equal to 11.

[0057] The present invention should not be considered to be limited to such operations performed within the scope of sub-step 132 for detecting the value of rank k based on an ordered set Σ r of information providing indices.

[0058] The present invention further provides the ability to modify the k principal components thus selected (represented by vector Φ' in FIG. 7) in step 130, before the execution of step 140. The purpose is to increase the contribution of the most "information-providing" (lowest rank) principal components to those with lower and higher ranks, each with a lower ability to represent spatial information. For this purpose, such step 130 further includes a sub-step 133 that applies an operation for filtering the selected k principal components, and the filtering intensity is specific to each of the k principal components of the selected ranks 1 to k (q or less), and its information-providing index σ1 r from σk r is proportional. Such a filtering operation 133 can consist of using a Gaussian filter, an average filter, a median filter, an anisotropic diffusion filter, a wavelet filter, or any other spatial or other filter available in this situation, either alone or in combination. As a preferred but non-limiting example, such a sub-step 133 can consist of adjusting the weight wj r specific to the principal component of rank j executed based on the value σj r . In this example, the weight can be multiplied linearly by the strength A of the Gaussian filter. Each selected component of rank j, i.e., j is less than or equal to k, is filtered by a factor of wj r ×A. In a variant or in addition to the linear multiplication, the weight can be represented in advance in the form of an input to a logarithmic, exponential, or polynomial function f. In this way, each selected score image of PCj, i.e., of rank j where j is less than or equal to k, can be filtered by a factor determined by f(wj r )×A. Finally, the "filtered" score image is used to regenerate the matrix U of scores. As shown in FIG. 15, such a weight wj r can be calculated as shown in

Number

[0059] The present invention should not be considered limited only to these examples of the calculation of the weight wj r in sub-step 133.

[0060] FIGS. 10 to 12 can show the benefits resulting from the execution of a method for attenuating the noise of experimental data resulting from multiple acquisitions by a magnetic resonance imaging apparatus in the field of HARDI diffusion imaging. Therefore, the acquisition of such experimental data can relate to a plurality of slices for a plurality of phases, in this case, the human brain. Therefore, FIG. 10 shows the raw data of phases 2, 12, 22, 32, 42, 52, and 62, respectively referred to as P2, P12, P22, P32, P42, P52, and P62 in FIG. 10, for a given slice. The first image series referred to as L1 shows the raw and thus noisy experimental data for a plurality of voxels and for each of the various phases P2 to P62 described above. The second and third image series L2 and L3 show the experimental data denoised by the implementation of the present invention. According to this, the selected k principal components are subjected to filtering (step 133 of the method according to FIG. 7) before (L3) or not before (L2) the projection of the experimental data onto the k principal components. The increased performance improvement provided by the present invention from series L1 to series L3 can be visually noticed.

[0061] In addition, the performance of method 100 according to the present invention can be shown in FIG. 11. It depicts the noise removed in step 140 of the method according to the present invention for various phases referenced by P2, P12, P22, P32, P42, P52, and P62 in FIG. 10, which relates to the noisy experimental data of series L2 and L3 of FIG. 10. It is clear that such noise does not show any spatial information of interest.

[0062] FIG. 12 depicts the use of HARDI diffusion data described with reference to FIG. 10 as part of tractography. FIG. 12 shows three color images I1, I2, I3 represented in grayscale for a given slice, and partial enlarged images PI1, PI2, and PI3 of the respective images I1, I2, I3. These images I1, I2, I3 are generated in step 150 based on the raw experimental data (series L1 of FIG. 10), the data denoised by the implementation of the present invention and shown by series L2 of FIG. 10, and series L3 of FIG. 10, respectively. With the enlarged images PI1, PI2, and PI3, the enhanced emphasis of the fibers can be seen with the naked eye. In the enlarged image PI1, the area of the organ (surrounded by the white circle) may be considered dead. This is different from the fact in the enlarged images PI2 and PI3. Moreover, with such use, such fibers can be counted. Based on the raw data of series L1 of FIG. 10, approximately 27,000 fibers can be counted. This is approximately 34,000 when the denoised data forming series L2 of FIG. 10 is used, and more than 37,000 when the denoised data forming series L3 of FIG. 10 is used. Therefore, the denoising of the experimental data according to the present invention provides a particularly large performance improvement.

[0063] Figures 13 and 14, like Figures 10 and 11, show the performance improvement provided by the present invention based on perfusion data. Thus, the acquisition of experimental data relates to multiple slices for multiple phases, in this case, related to the human brain. Thus, Figure 13 shows the raw data for phases 1, 5, 10, 15, 20, 25, 30, 35, and 40, referred to as P1, P5, P10, P15, P20, P25, P30, P35, and P40 respectively, for a given slice. The first image series referred to as L1 shows the raw, and thus noisy, experimental data for multiple voxels and for the various phases P1 to P40 described above respectively. The second and third image series L2 and L3 show the experimental data denoised by the implementation of the present invention. According to that, the selected k principal components either undergo (L3) or do not undergo (L2) filtering (step 133 of the method according to Figure 7) prior to the projection of the experimental data onto the k principal components. An increased performance improvement provided by the present invention can be visually noticed from series L1 to series L3. This performance improvement is more pronounced when the spectra of the experimental data for the voxel V of interest, which are denoised by the implementation of the present invention, are examined for the raw (A), the case where the selected k principal components did not undergo filtering (B), and the case where the filtering was performed (C) in Figure 13 respectively. It is worth noting that the curves depicted by the respective spectra become smoother and smoother.

[0064] In addition, the performance of method 100 according to the present invention can be shown in Figure 14. It depicts the removed noise related to series L2 and L3 of Figure 13. It is clear that such noise does not show any spatial information of interest.

Claims

1. A method (100) for attenuating noise of experimental data (Zi) resulting from multiple acquisitions by a magnetic resonance imaging apparatus (1) related to a basic volume of interest, called a "voxel" of interest hereinafter, wherein the voxel of interest is among a plurality of voxels, the method (100) is executed by a processing unit (4) of a medical imaging analysis system (S), and the method (100) comprises: a step (110) of collecting the experimental data (Z1,..., Zi,..., Zn) related to the plurality of voxels (V1,..., Vn); a step (120) of generating, for the plurality of voxels (V1,..., Vn), an ordered set of q principal components ordered according to their respective eigenvalues, and respective scores thereof, the scores corresponding to projections of the experimental data onto the q principal components; a step (130) of selecting the first k principal components from among the generated q principal components; a step (140) of projecting the experimental data (Zi) onto the selected k principal components to generate noise-attenuated experimental data (Zi') related to the basic volume of interest; and the step (130) of selecting k principal components of the method (100) comprises: An information providing index (σ1 ｒ , …, σq ｒ ) that determines the amount of spatial information included in the score image related to the principal component, in the form of the score image of the principal component for the plurality of voxels (V1, …, Vn) and the reduction rate of the value of the defined characteristic of the score before and after application of the smoothing filter to the score, and a sub-step (131) belonging to a set of characteristics including, alone or in combination, variance, standard deviation, median, and mean, as the defined characteristics of the score for the plurality of voxels (V1, …, Vn), and said generated information providing index (σ1 ｒ , …, σq ｒ ) to determine a rank k in a sub-step (132) of selecting the first k principal components, characterized in that it comprises a method (100).

2. A method (100) for attenuating noise of experimental data (Zi) resulting from multiple acquisitions by a magnetic resonance imaging apparatus (1) related to a basic volume of interest, called a "voxel" of interest hereinafter, wherein the voxel of interest is among a plurality of voxels, the method (100) is executed by a processing unit (4) of a medical imaging analysis system (S), and the method (100) comprises: a step (110) of collecting the experimental data (Z1,..., Zi,..., Zn) related to the plurality of voxels (V1,..., Vn); a step (120) of generating, for the plurality of voxels (V1,..., Vn), an ordered set of q principal components ordered according to their respective singular values, and respective scores thereof, the scores corresponding to projections of the experimental data onto the q principal components; a step (130) of selecting the first k principal components from among the generated q principal components; Projecting the experimental data (Zi) onto the selected k principal components to generate noise-reduced experimental data (Zi') related to the basic volume of interest (step 140); The step (130) of selecting k principal components of the method (100) comprises: Generating information-providing indicators (σ1r,..., σqr) that define the amount of spatial information contained in the score images related to the principal components, in the form of the score images of the principal components for the plurality of voxels (V1,..., Vn) and the rate of decrease of the defined characteristics of the scores before and after application of a smoothing filter to the scores, the defined characteristics of the scores for the plurality of voxels (V1,..., Vn) belonging to a set of characteristics including variance, standard deviation, median, and mean, either alone or in combination (sub-step 131); Determining rank k to select the first k principal components based on the generated information-providing indicators (σ1r,..., σqr) (sub-step 132); the method (100) being characterized by comprising this.

3. The sub-step (132) for determining the rank k comprises calculating a mathematical difference between values of the information providing indicators (σ1 ｒ ,..., σq ｒ ) when the information providing indicators (σ1 ｒ ,..., σq ｒ ) are ordered according to the q generated principal components, and the rank k is that of the principal component for which the mathematical difference becomes equal to or less than a determined threshold value. The method (100) according to claim 1 or 2.

4. The sub-step (132) for determining the rank k is such that when the information-providing indicators (σ1 ｒ , …, σq ｒ ) are sorted according to the q generated principal components from the information-providing indicator (σq ｒ ) of the principal component of rank q to the information-providing indicator (σ1 ｒ ) of the principal component of rank 1, it consists of detecting a flat portion (σp ｒ , …, σq ｒ ) depicted by the values of the information-providing indicators (σ1 ｒ ), and the rank k is that of the first principal component whose value of the information-providing indicator (σk ｒ ) deviates from the flat portion (σp ｒ ), the method (100) according to claim 1 or 2.

5. including a sub-step (133) of applying a filtering operation to the selected k principal components, the filtering strength of the filtering operation being specific to each of the selected k principal components and proportional to its information providing index (σ1 ｒ , …, σ11 ｒ ), the method (100) according to claim 1 or 2.

6. The filtering operation (133) consists of using a Gaussian filter, an average filter, a median filter, an anisotropic diffusion filter, either alone or in combination; the method (100) according to Claim 5.

7. Comprising one or more program instructions interpretable by the processing unit (4) of the medical imaging analysis system (S), the program instructions being capable of being placed in the non-volatile storage device of the medical imaging analysis system (S), and the execution of the instructions by the processing unit (4) causing the method (100) according to Claim 1 or 2 to be executed; a computer program.

8. A computer-readable storage medium containing the instructions of the computer program according to Claim 7.

9. Including a processing unit (4) provided to attenuate the noise of experimental data (Zi) resulting from multiple acquisitions by a magnetic resonance imaging device (1) related to a basic volume of interest, hereinafter referred to as a "voxel" of interest, the voxel of interest being among a plurality of voxels, the processing unit (4) Collecting the experimental data (Z1,..., Zi,..., Zn) related to the plurality of voxels (V1,..., Vn); For the plurality of voxels (V1, …, Vn), generate an ordered set of q principal components ordered according to their respective eigenvalues, and their respective scores, the scores corresponding to the projection of the experimental data onto the q principal components, select the first k principal components from among the q generated principal components, project the experimental data (Z1, …, Zi, …, Zn) onto the selected k principal components to generate experimental data (Z1′, …, Zi′, …, Zn′) with attenuated noise related to the plurality of voxels (V1, …, Vn), and is configured to the processing unit (4) is An information providing index (σ1 ｒ , …, σq ｒ ) that determines the amount of spatial information included in the score image related to the principal component in the form of the score image of the principal component for the plurality of voxels (V1, …, Vn) and the reduction rate of the value of the defined characteristic of the score before and after application of the smoothing filter to the score is generated, and the defined characteristics of the score for the plurality of voxels (V1, …, Vn) belong to a set of characteristics including variance, standard deviation, median, and mean, either alone or in combination. configured to select the first k principal components based on the generated information providing index, characterized in that the medical imaging analysis system (S).

10. A processing unit (4) provided to attenuate noise of experimental data (Zi) resulting from multiple acquisitions by a magnetic resonance imaging apparatus (1) related to a basic volume of interest, hereinafter called a "voxel" of interest, the voxel of interest being among a plurality of voxels, the processing unit (4) is collect the experimental data (Z1, …, Zi, …, Zn) related to the plurality of voxels (V1, …, Vn), for the plurality of voxels (V1, …, Vn), generate an ordered set of q principal components ordered according to their respective singular values, and their respective scores, the scores corresponding to the projection of the experimental data onto the q principal components, select the first k principal components from among the q generated principal components, project the experimental data (Z1, …, Zi, …, Zn) onto the selected k principal components to generate experimental data (Z1′, …, Zi′, …, Zn′) with attenuated noise related to the plurality of voxels (V1, …, Vn), and is configured to the processing unit (4) is The score images of the principal components for the plurality of voxels (V1, …, Vn), and information providing indicators (σ1r, …, σqr) that determine the amount of spatial information included in the score images related to the principal components, in the form of the reduction rate of the determined characteristics of the scores before and after application of the smoothing filter to the scores, wherein the determined characteristics of the scores for the plurality of voxels (V1, …, Vn) belong to a set of characteristics including variance, standard deviation, median, and mean, alone or in combination, A medical imaging analysis system (S), characterized in that it is configured to select the first k principal components based on the generated information providing indicators. The medical imaging analysis system (S) according to claim 9, including a program memory including the program instructions of the computer program according to claim 7, which cites claim 1.

Images