Method for post-processing perfusion acquisition sequences with a medical imaging device - Patents.com

Abstract

The present invention relates to a method (100) for post-processing sampled time-dependent experimental perfusion signals (S, 10, 12) in order to generate (140) pharmacokinetic parameters (QI). Such a method is implemented by a processing unit (4) of a medical imaging analysis system, the unit being pre-trained in a process (101) that enables learning of the destructive influence of the acquisition of a perfusion sequence on the arterial signal and the redundancy of information related to the arterial input function (AIF, AIF´) shared by at least two sets of tissue signals. Such an arterial input function (AIF, AIF´) is thus directly generated (110) in step (120) by the processing unit (4) trained from the first arterial input function (AIF0) and the pre-selected tissue signal (S) as such.

Description

[Technical Field]

[0001] The present invention relates to a method for post-processing a perfusion acquisition sequence by a medical imaging device. [Background technology]

[0002] The medical imaging device may be a magnetic resonance imaging (MRI) device, or an imaging device based on the use of X-rays, such as a CT scanner. Such a post-processing method allows for the estimation of a sample of the arterial input function, also known by the acronym AIF. This sample has the remarkable feature of being highly insensitive to the effects of acquisition and therefore remaining linear with respect to the concentration of the contrast agent, which is essential for the quantitative estimation of pharmacokinetic parameters, which ultimately allow for the characterization of lesions such as tumors or ischemic tissue.

[0003] Magnetic resonance imaging is based on analyzing the response of water molecule protons when they are excited in a magnetic field. This response depends on the environment of those protons, thus making it possible to distinguish different types of tissue. Nuclear magnetic resonance imaging devices, such as the device 1 of the medical imaging analysis system SAIM shown as a non-limiting example in Figures 1 and 2, are commonly used. This provides multiple digital image sequences 12 of one or more parts of a patient's body, such as the brain, heart, and lungs, as non-limiting examples. For this purpose, the device applies a combination of high-frequency electromagnetic waves to the body part in question and measures the signals re-emitted by specific atoms, such as hydrogen for nuclear magnetic resonance imaging, as a non-limiting example. This device thus makes it possible to determine the magnetic properties and, consequently, the chemical composition and therefore the properties of biological tissue in each elementary volume, commonly called a voxel, of the imaged volume. As shown in Figure 1, the nuclear magnetic resonance imaging device 1 is controlled by a console 2. A user 6, e.g., an operator, physician, or researcher, can select commands 11 to control the device 1 based on parameters or instructions 16 entered via an input human-machine 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 represent gestural commands or instructions issued by a human 6 as control or configuration data. Based on the information 10 generated by said device 1, a plurality of digital image sequences 12 of a part of the human or animal body are obtained. Such information 10 or images 12 may also be called "experimental data".

[0004] CT scanners in modern radiology share some usage similarities with magnetic resonance imaging devices such as those shown in Figures 1 and 2. These similarities may explain a variant of the medical imaging analysis system SAIM that incorporates an imaging device 1 in the form of a CT scanner instead of a magnetic resonance imaging device. The invention will be described below primarily with the aim of addressing the effects of acquisition with a magnetic resonance imaging device. However, the specific description may also apply to other imaging devices (such as CT scanners) as appropriate.

[0005] The image sequence 12 may optionally be stored in a server 3, i.e., a computer equipped with its own storage means, and constitute a patient medical file 13. Such file 13 may contain different types of images, such as functional images showing tissue activity or anatomical images reflecting tissue properties. The image sequence 12, or more generally, experimental data, is analyzed by a processing unit 4, which is provided for this purpose. Such processing unit 4 may, for example, consist of one or more microprocessors or microcontrollers executing appropriate application program instructions stored in the storage means of the imaging analysis system. "Storage means" refers to any volatile or, preferably, non-volatile computer memory. Non-volatile memory is computer memory with a technology that allows it to retain its data in the absence of an electrical energy supply. It may contain data resulting from inputs, calculations, measurements, and / or program instructions. Currently available non-volatile memories are primarily of the electrically programmable type, such as EPROM (Erasable Programmable Read-Only Memory), or of the more electrically programmable and erasable type, such as EEPROM (Electrically Erasable Programmable Read-Only Memory), flash, or SSD (Solid State Drive). Non-volatile memory is distinguished from memory known as "volatile," which loses data when power is removed. The main types of volatile memory currently available are RAM (random access memory, also known as "read-write memory"), DRAM (dynamic random access memory, which requires periodic refreshing), SRAM (static random access memory, which requires such refreshing when there is a power drop), DPRAM, or VRAM (particularly suitable for video). In the remainder of this document, "data memory" may be volatile or non-volatile, depending on the application in question.

[0006] The processing unit 4 includes means for communicating with the outside world to acquire images. Moreover, the communication means enable the processing unit 4 to ultimately provide or output, via an output human-machine interface 5, to a user 6 of the imaging analysis system, e.g., a graphical and / or acoustic representation of the estimates or determined quantities of biomarkers or pharmacokinetic parameters QIs created by the processing unit 4 based on experimental data 10 and / or 12 obtained by magnetic resonance imaging. Throughout this document, "output human-machine interface" refers to any device, used alone or in combination, that allows outputting or providing a graphical, haptic, acoustic, or more generally human-perceptible representation of the reconstructed physiological signal, in this case a biomarker, to a user 6 of the magnetic resonance imaging analysis system. Such an output human-machine interface 5 may, non-exhaustively, consist of one or more screens, speakers, or other suitable alternatives. Thus, the user 6 of the analysis system can confirm or invalidate the diagnosis, decide on therapeutic actions he or she deems appropriate, initiate further studies, refine adjustment parameters of measurement devices, etc. Optionally, this user 6 may further configure the operation of the processing unit 4 or the output human-machine interface 5 by means of operation and / or acquisition parameters 16. Thus, for example, the user 6 may define display thresholds or select biomarkers, indicators or estimated or quantified parameters that he wishes to display. For this purpose, he makes use of the input human-machine interface 8 described above or a second input interface provided for this purpose. Advantageously, the input human-machine interface 8 and the output human-machine interface 5 may constitute one and the same physical entity. The input human-machine interface 8 and the output human-machine interface 5 of the imaging analysis system may further be integrated into the acquisition console 2. Variations exist which are explained in relation to FIG. 2.In contrast, the imaging system as described above further comprises a pre-processing unit 7 for analyzing the image sequence 12, estimating experimental signals 15 therefrom and sending these experimental signals 15 to a processing unit 4 which is thus relieved of this task.

[0007] Among the techniques or methods based on magnetic resonance imaging, perfusion magnetic resonance imaging can be distinguished. Such a technique is illustrated in FIG. 3, which shows a cardiac perfusion acquisition sequence, and consists of applying a radiofrequency pulse, called a magnetization preparation radiofrequency pulse, to bring the magnetization to its steady state (thereby also eliminating the magnetization history), followed by a series of acquisition radiofrequency pulses to spatially sample the volume of interest. The set of parameters chosen by the technician aims to maximize or minimize various effects. Among the most important are T1-weighted enhancement or T2 * A selection of pulse types allows obtaining a sequence called enhanced susceptibility. During acquisition, the technician injects a tracer in the form of a contrast agent into the venous system of the patient undergoing the examination. The magnetization of the volume of interest therefore changes during the movement of the contrast agent, which is distributed throughout the body by the circulatory system. This change depends on the amount of tracer. Figure 3 therefore shows such changes in the magnetization of the volume of interest at three separate times t0, t2, and t3, respectively, before and after the arrival of the tracer at the volume of interest at the moment t1. A perfusion sequence thus consists of periodically acquiring the magnetization state of the volume of interest by quasi-periodically repeating the application of acquisition radio-frequency pulses, which in certain cases may be preceded by a preparatory pulse. A perfusion sequence thus allows obtaining, for each elementary volume or voxel, a time sample S(t) of the change in the magnetization state due to the acquisition parameters selected by the technician.

[0008] CT scanners differ from MRI by their fundamental principle of higher or lower opacity of the anatomical structures through which the X-rays pass. The principle of perfusion sequences therefore consists in injecting a contrast agent that makes it possible to modify this opacity. The main disruptive effect that can be mentioned is the subsampling of the AIF.

[0009] As depicted in Figure 4, the estimation of the pharmacokinetic parameters is based on the theory of tracer dilution, which allows the AIF to be related to the concentration S of the contrast agent in the tissue (for five regions of interest A1 to A5 as shown in Figure 4) by convolution with a filter called "IRF", which represents the "impulse response" or "impulse response function" related to the recirculation of said contrast agent, as follows: S = AIF * IRF. Figure 5 therefore shows five time signals S associated with said regions A1 to A5, respectively.

[0010] By directly collecting samples of the contrast agent concentration in the tissue in the acquired images and selecting a reasonable sample of the AIF, the IRF and therefore the pharmacokinetic parameters can be estimated.

[0011] Perfusion sequences can reveal lesions that are not visible in their natural state, which is extremely useful for physicians seeking to establish a diagnosis and make therapeutic decisions regarding the treatment of the lesion. However, for complete decision-making, whether in clinical practice or research, it is necessary to obtain an AIF that is largely unaffected by the acquisition. Figure 5 shows the signal S acquired by a perfusion sequence depending on the concentration C of the contrast agent or tracer according to multiple acquisition parameters that describe multiple curves S. In fact, magnetic resonance images capture the magnetization state of tissue, and the presence of contrast agent indirectly affects the acquired signal. The contrast agent will affect the rate at which magnetization returns to a steady state. In other words, for the same set of acquisition parameters, magnetization will change depending on the concentration of contrast agent injected into the patient. This relationship is not linear or even bijective. However, as shown by the right-hand diagram in Figure 5, for low amounts of contrast agent, it is common to consider the relationship between the acquired signal S and the contrast agent concentration C to be linear. This right-hand diagram shows an enlarged view of the case where the contrast agent concentration C is low. Regardless of the acquisition parameters, the curves S and C converge to a straight line identified as DI, represented by a dashed line in the region ZL, referred to as the gray linearity region in FIG. 5. Therefore, when the concentration C is low, the directly acquired signal S can be used without correcting for acquisition effects, which is generally accepted for tissue signals. However, as shown by the curve in the left diagram of FIG. 5, the higher the concentration C, the greater the acquisition effects. Therefore, the accepted approximation for tissue signals may not apply to arterial signals, which are characterized by the maximum concentration of contrast agent. Therefore, it is necessary to correct for acquisition effects to estimate valid AIF samples.

[0012] To date, several post-processing methods have been proposed in the literature to correct for this acquisition effect, such as those disclosed in the paper "Physics-informed neural networks for myocardial perfusion MRI quantification" by van Herten et al. 2022. Some of them are rarely used, such as those involving the use of AIFs, called population AIFs (which use an AIF database corresponding to a given injection protocol and patient type), or those that require a blood sample simultaneously with the acquisition in order to truly know the amount of contrast agent present at a given moment. On the other hand, two methods are considered that the present invention directly competes with, as well as a third method that the present invention proposes to improve.

[0013] A first known method consists in digitally simulating the effects of the acquisition by means of physical models derived from the Bloch equations. These models require a very detailed knowledge of the perfusion sequence and therefore must be reevaluated for each case. In addition, the signals S acquired in one and the same perfusion study are subject to parameters specific to the tissue that generated them, in particular T1 and T2. * will depend on T1 and T2 * Therefore, it would need to be estimated by other acquisition sequences and evaluated by post-processing methods that inevitably have biases. Thus, the acquisition of additional parameters whose values ​​vary for each voxel of the image introduces a source of error and a new problem to solve: multimodal registration between perfusion images and T1 and T2 maps. Finally, the set of parameters required for this method is often lost because it is not necessarily stored in daily clinical practice, thereby limiting the future reuse of perfusion acquisitions.

[0014] The second method proposes using a set of acquisition parameters that significantly limits the influence of the acquisition and thus allows the arterial signal to be considered as a realistic AIF. However, such a method reduces the image resolution. As a result, it is impossible to estimate pharmacokinetic parameters very accurately using such images. In fact, this acquisition is combined with a conventional perfusion acquisition, which offers sufficient resolution for the evaluation of these parameters by using the AIF obtained by the arterial signal of an acquisition optimized for this purpose. This method has the advantage of allowing the direct use of signals resulting from different acquisitions. However, it assumes the presence of technicians trained in this type of setup and the availability of this type of operation on the MRI system, which is currently rare. In addition, this method assumes that the two types of acquisitions are obtained simultaneously. In other words, it cannot include any previous perfusion acquisitions. However, over the past 30 years, many databases have been formed using conventional acquisition methods. This method would exclude them.

[0015] Finally, a third method, shown in FIG. 6, proposes estimating AIF samples using all redundant information shared by the tissue signals. The signal of each voxel representing a tissue region of interest, such as regions A1 through A5 shown by dashed circles in FIG. 6 as an example, is the convolution product between the IRF belonging to each voxel and the AIF shared by all tissue voxels, making it possible to simultaneously estimate the IRF of each voxel and the AIF of all voxels. This method is denoted by the term blind deconvolution, since neither the IRF nor the AIF are known. Such a method 100 is implemented, for example, by the processing unit 4 of the medical analysis system shown in FIG. 1 or 2, which is supplied with experimental data 10 and / or 12. It consists of first selecting a tissue signal S and a first arterial input function AIF0 in the arterial region A0, generally represented by a dashed box in FIG. 6, in step 110, and then iteratively minimizing the residual between the measured tissue signal S and the arterial input function AIF reconstructed based on a parametric model AIFm. The goal is to minimize the residual by adjusting the parameters of the arterial input function model AIFm. A wide variety of methods for performing this minimization 120 have been described in the literature. They are generally distinguished by the constraints they impose on the arterial input function model. Some methods are said to be "unconstrained" because the arterial input function samples can take any shape. This first type of method is very sensitive to acquisition noise and is therefore rarely used. Other methods are parameterized by a given model. However, this second type of method is limited by the nature of the arterial input function model used, thereby limiting the shape the arterial input function can take.

[0016] The choice and number of tissue signals selected in step 110 varies according to the method chosen. Even if it has been shown that theoretically, including methods that do not require any constraints on the shape of the sought arterial input function, their number can be reduced to two, limiting this number of signals to two assumes a great many assumptions that are not very realistic in practice, especially regarding the influence of noise and the type of perfusion model representing these two signals. Generally, the number of tissue signals used in the literature varies from 6 to 12.

[0017] The second step 120, which minimizes the error, generally consists of minimizing the sum of squares of the residuals between the signals selected in step 110 and those reconstructed based on the adjustment of the parameters of the arterial input function model and the pharmacokinetic parameters. The iterative step 120 then consists of searching for a set of parameters that allows reducing this total error. The adjustment is customarily performed by gradient descent, with each iteration involving a large number of mathematical operations that are costly in terms of computational time. The stopping criterion determining the final number of iterations varies from study to study but may relate to the minimum change in error from one iteration to another, or even directly to the error value. In the literature, methods can be found that suggest 50 to several hundred iterations, resulting in an implementation time of several minutes. Once this second step is performed, the method allows for the generation of a second arterial input function (AIF) that would be considered "realistic" for the set of selected tissue signals. However, by its very nature, the method only reveals the correct shape of the determined second arterial input function (AIF). Therefore, it is necessary to perform step 130, which consists of adjusting the scale (sample height) of the AIF. Several methods for performing this scaling or normalization step exist in the literature. For example, it can be assumed that the determined AIF shares the area under the curve of AIF0, or at least the part that is largely unaffected by the acquisition. The second corrected arterial input function AIF' thus obtained can then be used to estimate one or more of its pharmacokinetic parameters QI in a subsequent step 140 over all or a portion of the voxels of the perfusion image, thus generating information relevant for making therapeutic decisions. Such pharmacokinetic parameters QI can be optionally formed into one or more graphical representations and output via a suitable human-machine interface, such as the interface 5 associated with the medical imaging analysis system according to FIG. 1 or 2, so that a member of medical personnel can, for example, make a diagnosis. Summary of the Invention

[0018] The present invention makes it possible to overcome all or part of the drawbacks presented by known or previously mentioned solutions.

[0019] Among the many advantages provided by the present invention: - a free, i.e. unconstrained, form for the arterial input function that is less affected by acquisition noise, - reducing the number of tissue signals required to perform the method; Mention may be made of the implementation of a post-processing method that allows a non-iterative execution of the estimation of the arterial input function, which makes it possible to obtain execution times of the order of seconds on a standard computer for general use, as opposed to several minutes according to the prior art.

[0020] To this end, the invention relates to a method for post-processing sampled temporal experimental signals resulting from a perfusion acquisition sequence by a medical imaging device and resulting from the movement of a tracer within an elementary volume of an organ, said method being implemented by a processing unit of a medical imaging analysis system, said method comprising: - selecting a first arterial input function related to an arterial region of an organ and a set of tissue signals respectively related to separate tissue regions of said organ; generating a second arterial input function based on the first arterial input function and the selected set of tissue signals; - generating pharmacokinetic parameters based on the second arterial input function and the experimental signal.

[0021] In order to increase the performance of such a method with respect to known techniques, the step of generating a second arterial input function according to the present invention consists of performing basic operations by said processing unit pre-trained according to a process of learning the disruptive effect of the acquisition of perfusion sequences on arterial signals and the redundancy of information related to such second arterial input function shared by at least two sets of tissue signals.

[0022] According to certain embodiments, such a method may include a step of correcting the generated second arterial input function by scaling it before performing a step of generating pharmacokinetic parameters based on the corrected second arterial input function and the experimental signal.

[0023] Advantageously, if the medical imaging system comprises an output human-machine interface, the method according to the invention may comprise the step of creating a graphical representation of the second arterial input function and outputting said graphical representation by means of said output human-machine interface, in order to provide a member of medical staff with the opportunity to verify the validity of the generated second arterial input function.

[0024] According to a preferred embodiment, the learning process may consist of deep learning based on minimizing the mean value of the squared error between the actual samples of the arterial input function that made it possible to generate the tissue signal and the values ​​of the estimates of these same samples made by said learning process; such a learning process may be carried out by an Adam optimizer.

[0025] According to a second subject, the invention relates to a computer program comprising one or more program instructions that can be executed by a processing unit of a computer, said program instructions being able to be located in a non-volatile memory of said computer, and execution of the program instructions by said processing unit resulting in the implementation of the post-processing method according to the invention.

[0026] Similarly, according to a third subject, the invention relates to a computer-readable storage medium containing instructions for such a computer program.

[0027] Finally, the invention also relates to a medical imaging analysis system comprising a processing unit in communication with the outside world and arranged to receive a set of samples of time-experimental signals resulting from a perfusion acquisition sequence by a medical imaging device and resulting from the movement of a tracer within an elementary volume of an organ, such processing unit having been pre-trained according to the aforementioned learning process and comprising storage means containing the program instructions of the computer program according to the invention.

[0028] Other features and advantages will become more clearly apparent upon reading the following description and examining the accompanying drawings. [Brief explanation of the drawings]

[0029] [Figure 1] FIG. 1, already described, shows a medical imaging system comprising a medical imaging platform. [Figure 2] FIG. 2, already described, shows a variant of such a medical imaging system comprising a medical imaging platform. [Figure 3] FIG. 3, already described, illustrates a cardiac perfusion sequence by medical imaging. [Figure 4] FIG. 4, previously described, shows an example of a known process for estimating the impulse response for a voxel of interest and thus generating pharmacokinetic parameters. [Figure 5] FIG. 5, already explained, shows the destructive effect of the acquisition on the linearity of the signal as a function of the concentration of the contrast agent. [Figure 6] FIG. 6, already described, shows a known method for obtaining an estimate of the arterial input function by blind deconvolution. [Figure 7] FIG. 7 illustrates a method for generating realistic arterial input functions by blind deconvolution, and ultimately the pharmacokinetic parameters of interest, in accordance with the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0030] As shown in FIG. 7, the present invention relates to a method 100 for generating a pharmacokinetic parameter of interest (QI) based on one or more sampled time-experimental signals (10, 12) resulting from acquisitions by a perfusion medical imaging device (1). Such a method (100) can be implemented by a processing unit (4) of such a medical imaging analysis system. Such a method (100) relies on a process (101) of learning the disruptive effects of perfusion sequence acquisitions on arterial signals combined with learning the redundancy of information related to the arterial input function shared by at least two sets of tissue signals. Therefore, the processing unit 4 is provided to include a convolutional neural network or any other equivalent solution preconfigured by implementing the learning process 101.

[0031] In fact, all of the tissue signals result from the convolution of one and the same arterial input function AIF with their respective impulse responses IRF. Therefore, it is possible to extract shared information from them, which is the desired arterial input function AIF. Learning the destructive effects of acquisition is performed by providing an arterial signal AIF0, which is itself a degraded version of the desired AIF. Therefore, it becomes a question of "transforming" such a selected arterial signal AIF0 so that it corresponds to the information shared by the tissue signals while avoiding proposing less likely solutions. In other words, this aspect promotes the stability of the algorithm. This learning process 101 can be performed using a combination of physically acquired signals from volunteers, patients, or physical phantoms and / or simulated signals, which allow for the generation of more diverse examples and thus promote the inference capabilities of the method for generating such a realistic arterial input function AIF. Using the database ADB thus generated, learning 101 can be performed, for example, using a deep learning algorithm. The estimation of the learning algorithm parameters may be derived by minimizing the estimation error between the desired AIF and the one used to generate the tissue signal of the database ADB. Said estimation may also be derived directly by minimizing the error between said tissue signal originating from the database ADB and the tissue signal estimated by the algorithm, if a step of reconstructing said signal using the AIF' generated by the algorithm is proposed. Thus, the database ADB may optionally include: - the AIFs to be obtained, their variants AIF0 affected by the acquisition, and the signals they generate, - AIF0 affected by the acquisition and the signal generated by the AIF sought.

[0032] Once the learning step 101 allows a sufficient performance to be achieved, the post-processing method 100 may use a convolutional neural network (or any other equivalent solution), hereinafter referred to as a "network") thus trained in the processing unit 4 implementing said method 100. Similar to the post-processing method 100 described in relation to the third prior art method, the method 100 includes a step 110 of selecting an initial proposed arterial input function AIF0 and a set of tissue signals S that do not correspond to blood vessels or voids. Such a step 110 may be performed according to various techniques. Thus, such a selection may consist of selecting a mean arterial input function derived from a known population. It may also result from a manual selection by an engineer on images resulting from the acquisition of sampled time-experimental signals of several regions of the sample from which this arterial input function is derived, as well as the set of tissue signals, or from the implementation of any post-processing method that allows the arterial input function and the set of tissue signals to be estimated automatically from images. The method 100 includes a step 120 of generating an estimate of a second realistic arterial input function AIF for a set of input signals that has not been affected by the acquisition of the first proposed arterial input function AIF0. Unlike the previously known method 100, step 120 of the method 100 according to the present invention consists in generating the second realistic arterial input function AIF directly, i.e., without performing the lossy iterative process 120, based on the basic operations performed by the network trained according to process 101. Therefore, performing step 120 is significantly faster and simpler than prior art iterative solutions. The method 100 according to the present invention may further include a step 130 of scaling the generated second arterial input function AIF so that it is finally used, after correction, in step 140 to generate one or more pharmacokinetic parameters QI of interest for all or a portion of the voxels of the perfusion image, which pharmacokinetic parameters QI may be subject to output in graphical form by an output human-machine interface, such as the interface 5 of the medical imaging analysis system SAIM according to FIG. 1 or 2.However, the present invention provides that the step 130 of normalizing the generated AIF can be further integrated directly into the step 120 of directly generating a realistic, normalized, i.e., scaled, AIF' that can be used to perform step 140 of generating one or more pharmacokinetic parameters QI of interest for all or a portion of the voxels of the perfusion image. A reasonable scaling factor can be calculated in such a way as to preserve mass conservation between the generated arterial input function and the experimental signal, for example by matching low distortion regions in the two signals.

[0033] An embodiment of the network training process 101 according to the invention that makes it possible to verify the validity of the invention will now be considered.

[0034] As previously mentioned, the ADB database can be composed of over 3 million AIF samples simulated by randomly varying the parameters of a model called Parker's model. Thus, each AIF sample can be used to generate five tissue signals with different pharmacokinetic parameters. The model used to generate these signals can be the Toft-Ketty model, extended by adding a delay between the AIF sample and the tissue.

[0035] To generate a second realistic arterial input function AIF in step 120 of the post-processing method 100 according to the present invention, a deep learning algorithm is selected. Such a deep learning algorithm is advantageously divided into two branches. The first branch processes samples of the arterial input function AIF0. It consists of a one-dimensional convolutional network with three layers. The second branch processes a set of tissue signal samples. It consists of a two-dimensional convolutional network with three layers. The two branches each end with a simple neuron network with only one layer, which is then joined by a simple neuron network with three layers. The last layer proposes at its output samples of the arterial input function AIF of the same size as the first proposed arterial input function AIF0.

[0036] Advantageously, the learning process 101 can be performed by an optimizer called the Adam optimizer, based on minimizing the mean squared error between the actual samples of the arterial input function that allowed the tissue signals to be generated and the values ​​of these same samples estimated by the learning algorithm. The number of iterations spent on this training was 30 on 80% of the generated database, and validation was then performed on the remaining 20%. The result obtained from validation was, in terms of mean squared error, 0.69% of the peak value of the true AIF samples, both for the training database and for those used for validation. This method is generally described in the literature as achieving the best results. However, this deep learning can be replaced by a table-building method that allows for a correspondence between inputs represented by a first set of arterial input functions and tissue signals and outputs representing a second arterial input function that is not affected by the acquisition. In this case, learning consists of establishing such a table.

[0037] To ensure the clinical feasibility of the post-processing method 100 according to the present invention, this same deep learning algorithm was tested on a database of 43 MRI cardiac perfusion images. These images had the notable feature of being acquired using the second prior art method described above. In doing so, the available arterial input function samples were largely unaffected by the acquisitions associated with each conventional acquisition. By manually selecting an area of ​​the myocardium from which five samples of tissue signal were extracted and an area of ​​the left ventricle from which a proposed arterial input function sample AIF0 affected by multiple acquisitions was selected, allowing the learning algorithm to be evaluated on real data by comparing the values ​​of the arterial input function sample estimates made by the learning algorithm with those largely unaffected by acquisitions. The results were evaluated in terms of mean square error and coefficient of determination (R2) and compared to the values ​​obtained by the proposed arterial input function sample AIF0 affected by multiple acquisitions. The median values ​​for mean square error were less than 0.08 compared to 0.15 for the arterial input function samples affected by multiple acquisitions, and greater than 0.85 compared to less than 0.3 for R2. Thus, for practical acquisitions, the present invention allows obtaining realistic and correct estimates of the samples of the sought arterial input function AIF.

[0038] For the purpose of verification of the validity of the second arterial input function generated in step 120 by a medical professional using the post-processing method 100 according to the invention, the post-processing method 100 according to the invention may further comprise a step 150 of creating a graphical representation of said second arterial input function AIF before or after its correction. Such step 150 may furthermore consist in outputting said graphical representation by an output human-machine interface, if the medical imaging analysis system implementing said method 100 comprises an output interface such as the interface 5 of the system shown in Figures 1 and 2.

[0039] The present invention has been described with reference to the non-limiting example of signals resulting from a cardiac acquisition sequence with a perfusion imaging device. The invention is not limited to this single considered organ and can be used to generate pharmacokinetic parameters of any other organ of interest, such as the brain.

Claims

1. A method (100) for post-processing sampled temporal experimental signals (S, 10, 12) resulting from a perfusion acquisition sequence by a medical imaging device (1) and resulting from the movement of a tracer within an elementary volume of an organ, said method being implemented by a processing unit (4) of a medical imaging analysis system (SAIM), said method (100) comprising: selecting (110) a first arterial input function (AIF0) related to an arterial area (A0) of said organ, and a set (S) of tissue signals respectively related to separate tissue areas (A1, A2, A3, A4, A5) of said organ; generating (120) a second arterial input function (AIF, AIF') based on the first arterial input function (AIF0) and the selected set of tissue signals (S); generating (140) a pharmacokinetic parameter (QI) based on the second arterial input function (AIF') and the experimental signal (S, 10, 12); 10. The method of claim 1, wherein the step of generating a second arterial input function (AIF) (120) consists of performing basic operations by said processing unit (4), said processing unit having been previously trained according to a process (101) for learning the disruptive effects of perfusion sequence acquisitions on arterial signals and the redundancy of information related to such second arterial input function shared by at least two sets of tissue signals.

2. 2. The method (100) of claim 1, comprising a step (130) of correcting the generated second arterial input function (AIF) by scaling it before performing a step (140) of generating a pharmacokinetic parameter (QI) based on the thus corrected second arterial input function (AIF') and the experimental signal (S, 10, 12).

3. 3. The method (100) according to claim 1 or 2, wherein the medical imaging system (SAIM) comprises an output human-machine interface (5), and the method (100) comprises the step of creating (150) a graphical representation of the second arterial input function (A1F, A1F′) and outputting (150) the graphical representation by the output human-machine interface (5).

4. 3. The method according to claim 1 or 2, wherein the learning process (101) consists of deep learning based on minimizing the mean value of the squared error between the actual samples of the arterial input function that allowed to generate the tissue signal and the values ​​of the estimates of these same samples made by the learning process.

5. 5. The method of claim 4, wherein the learning process (101) is performed by an Adam optimizer.

6. A computer-readable storage medium comprising program instructions that can be executed by a processing unit of a computer, the program instructions being capable of being placed in a non-volatile memory of the computer, and execution of the program instructions by the processing unit performing the method of claim 1 or 2.

7. 1. A medical imaging analysis system (SAIM) comprising: a processing unit (4) in communication with the outside world, arranged to receive sets of samples of time-experimental signals (S, 10, 12) resulting from perfusion acquisition sequences by a medical imaging device (1) and resulting from the movement of a tracer within an elementary volume of an organ, said processing unit (4) having been pre-trained according to a process (101) for learning the disruptive effects of the acquisition of perfusion sequences on arterial signals and the redundancy of information related to arterial input functions shared by at least two sets of tissue signals; and storage means comprising program instructions executable by said processing unit (4), wherein execution of said program instructions by said processing unit (4) results in the implementation of the method according to claim 1 or 2.

Images