Method for post-processing a sequence of acquisition of perfusion by a medical imaging device

Abstract

The invention relates to a method (100) for post-processing a sampled time-dependent experimental perfusion signal (S, 10, 12) to generate (140) a pharmacokinetic parameter (QI). Such a method is implemented by a processing unit (4) of a medical-imaging analysis system, said unit having been trained beforehand in a process (101) allowing the disrupting effect of acquisition of a perfusion sequence on arterial signals and redundancy of information related to an arterial input function (AIF, AIF') shared by a set of at least two tissual signals to be learnt. Such an arterial input function (AIF, AIF') is produced directly in a step (120) by said processing unit (4) thus trained from a first arterial input function (AIF0) and from tissual signals (S) selected beforehand (110).

Description

[0001] Method for post-processing a perfusion acquisition sequence by a medical imaging device

[0002] The invention relates to a method for post-processing a perfusion acquisition sequence by a medical imaging device. The latter may be a magnetic resonance imaging device, better known by the acronym MRI or the English acronym MRI for "Magnetic Resonance Imaging", or an imaging device based on the exploitation of X-rays such as a scanner. Such a post-processing method makes it possible to estimate a sampling of the arterial input function, also known by the English acronym AIF for "Arterial Input Function". This sampling has the particularity of being much less sensitive to acquisition effects and therefore of maintaining linearity with respect to the concentration of the contrast product, essential for a quantitative estimation of the pharmacokinetic parameters, which ultimately makes it possible to characterize lesions such as tumors or ischemic tissues.

[0003] Magnetic resonance imaging is based on an analysis of the response of the proton of a water molecule when it is excited in a magnetic field. This response depends on the environment of such a proton and thus makes it possible to differentiate different types of tissue. A Nuclear Magnetic Resonance imaging device, such as the device 1 of a SAIM medical imaging analysis system illustrated by way of non-limiting example by figures 1 and 2, is generally used. This delivers a plurality of digital image sequences 12 of one or more parts of a patient's body, by way of non-limiting examples, the brain, the heart, the lungs. Said device applies for this purpose a combination of high-frequency electromagnetic waves to the part of the body considered and measures the signal re-emitted by certain atoms, such as by way of non-limiting example, hydrogen for Nuclear Magnetic Resonance imaging.The device thus makes it possible to determine the magnetic properties and, consequently, the chemical composition of biological tissues and therefore their nature, in each elementary volume, commonly called a voxel, of the imaged volume. As shown in Figure 1, a Nuclear Magnetic Resonance imaging device 1 is controlled using a console 2. A user 6, for example an operator, practitioner or researcher, can thus choose commands 11 to control the device 1, from parameters or instructions 16 entered via a human-machine input interface 8 of the analysis system. Such a human-machine interface 8 may consist, for example, of a computer keyboard, a pointing device, a touch screen, a microphone or, more generally, any interface arranged to translate a gesture or an instruction issued by a human 6 into control or parameter data.From information 10 produced by said apparatus 1, a plurality of sequences of digital images 12 of a part of a human or animal body are obtained. We will also call such information 10 or images 12 “experimental data”.

[0004] A modern radiology scanner shares several similarities in use with a magnetic resonance imaging device as illustrated in Figures 1 and 2. The latter could describe a variant of the SAIM medical imaging analysis system integrating an imaging device 1 in the form of a scanner instead of a magnetic resonance imaging device. The invention will be described below mainly with the aim of solving an acquisition effect by a magnetic resonance imaging device. However, a specific description may relate to other imaging devices (such as the scanner) when necessary.

[0005] The image sequences 12 may optionally be stored within a server 3, i.e. a computer equipped with its own storage means, and constitute a medical file 13 of a patient. Such a file 13 may comprise images of different types, such as functional images highlighting the activity of the tissues or anatomical images reflecting the properties of the tissues. The image sequences 12 or, more generally, the experimental data, are analyzed by a processing unit 4 arranged for this purpose. Such a processing unit 4 may, for example, consist of one or more microprocessors or microcontrollers implementing suitable application program instructions loaded into storage means of said imaging analysis system. The term "storage means" means any volatile or, advantageously, non-volatile computer memory.Non-volatile memory is a computer memory whose technology allows its data to be retained in the absence of an electrical power supply. It can contain data resulting from inputs, calculations, measurements and / or program instructions. The main non-volatile memories currently available are electrically writable, such as EPROM (Erasable Programmable Read-Only Memory), or electrically writable and erasable, such as EEPROM (Electrically-Erasable Programmable Read-Only Memory), flash, SSD (Solid-State Drive), etc. Non-volatile memories are distinguished from so-called "volatile" memories, the data of which is lost in the absence of an electrical power supply.The main volatile memories currently available are of the RAM type ("Random Access Memory" in English terminology or also called "live memories"), DRAM (dynamic random access memory, requiring regular updating), SRAM (static random access memory requiring such updating during a power shortage), DPRAM or VRAM (particularly suitable for video), etc. A "data memory", in the rest of the document, can be volatile or non-volatile depending on the intended application.

[0006] Said processing unit 4 comprises means for communicating with the outside world to collect the images. Said means for communicating further allow the processing unit 4 to deliver or output, ultimately, a rendering, for example graphic and / or sound, of an estimation or quantification of a biomarker or of a pharmacokinetic parameter Ql developed by said processing unit 4 from the experimental data 10 and / or 12 obtained by Magnetic Resonance Imaging, to a user 6 of the imaging analysis system via an output human-machine interface 5.Throughout the document, the term “output human-machine interface” means any device, used alone or in combination, making it possible to output or deliver a graphic, haptic, audio or, more generally, human-perceptible representation of a reconstructed physiological signal, in this case a biomarker, to a user 6 of a Magnetic Resonance imaging analysis system. Such an output human-machine interface 5 may consist, in a non-exhaustive manner, of one or more screens, loudspeakers or other suitable alternative means. Said user 6 of the analysis system can thus confirm or deny a diagnosis, decide on a therapeutic action that he deems appropriate, deepen research work, refine adjustment parameters of measuring equipment, etc.Optionally, this user 6 can also configure the operation of the processing unit 4 or the output human-machine interface 5, by means of operating and / or acquisition parameters 16. For example, he can thus define display thresholds or choose the biomarkers, indicators or estimated or quantified parameters for which he wishes to have a representation. The user uses for this the input human-machine interface 8 previously mentioned or a second input interface provided for this. Advantageously, the input 8 and output 5 human-machine interfaces can constitute a single physical entity. Said input 8 and output 5 human-machine interfaces of the imaging analysis system can also be integrated into the acquisition console 2.There is a variant, described in connection with Figure 2, for which an imaging system, as described previously, further comprises a preprocessing unit 7 for analyzing the image sequences 12, deducing experimental signals 15 therefrom and delivering the latter to the processing unit 4 which is thus relieved of this task. Among the techniques or modalities based on magnetic resonance imaging, perfusion imaging is distinguished. Such a technique, shown diagrammatically in Figure 3 which illustrates a cardiac perfusion acquisition sequence, consists of applying a so-called magnetization preparation radiofrequency pulse to allow the latter to evolve towards its equilibrium state (it can also make it possible to delete 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 operator aims to maximize or minimize different effects, among the most notable, we find the choice of the type of pulses allowing either to obtain a sequence called T1-weighted enhancement or T2*-weighted susceptibility. The operator proceeds, during the acquisition, to the injection of a tracer in the form of a contrast product into the venous system of the patient undergoing the examination. The magnetization of the volume of interest then varies during the passage of the latter - distributed to the whole body by the blood system - and this variation is a function of the quantity of the tracer. Figure 3 thus illustrates at three distinct times t0, t2, t3, such an evolution of the magnetization of the volume of interest, respectively upstream and downstream of the arrival at time t1 of the tracer in the volume of interest.The perfusion sequence thus consists of regularly acquiring the magnetization state of the volume of interest by repeating quasi-periodically the application of a radiofrequency acquisition pulse which may be preceded in certain cases by a preparation pulse. A perfusion sequence thus makes it possible to obtain for each elementary volume or voxel, a temporal sampling S(t) of the variation of the magnetization state resulting from the acquisition parameters selected by the manipulator.

[0007] The case of the scanner differs from the MRI by its basic principle: the more or less significant opacity of the anatomical structures crossed by X-rays. The principle of the perfusion sequence therefore consists of injecting a contrast product to modify this opacity. The main disruptive effect that could be cited is an undersampling of the AIF.

[0008] As described in Figure 4, the estimation of the pharmacokinetic parameters is based on the tracer dilution theory which makes it possible to link the AIF to the concentration of the contrast product S (relative to five regions of interest A1 to A5 as shown in Figure 4) in the tissue by a convolution product with a filter called "impulse response" or "IRF" an English acronym for "Impulse Response Function" in relation to the recirculation of said contrast product, such that S = AIF * IRF. Figure 5 thus illustrates five time signals S respectively associated with said regions A1 to A5.

[0009] By retrieving a sample of the contrast agent concentration in the tissue directly from the acquired images and choosing a relevant AIF sample, it is possible to estimate the IRF and therefore the pharmacokinetic parameters.

[0010] Perfusion sequences allow the updating of invisible lesions in the native state, which is very useful for a practitioner seeking to establish a diagnosis and make a therapeutic decision in the treatment of pathologies. However, in order to refine decision-making, whether in the clinic or in the field of research, it is necessary to obtain an AIF little subject to acquisition effects. Figure 5 describes a signal S acquired by a perfusion sequence as a function of the concentration C of the contrast agent or tracer according to a plurality of acquisition parameters, which explains the plurality of S curves. Indeed, magnetic resonance imaging being a capture of a state of magnetization of tissues, the presence of a contrast agent acts indirectly on the acquired signal. The contrast agent will influence the speed of return to the equilibrium state of magnetization.In other words, for the same set of acquisition parameters, the magnetization will vary according to the concentration of the product injected into the patient. This relationship is neither linear nor even bijective. However, as indicated by the right-hand view described in Figure 5, in the case of a low dose of contrast agent, it is customary to consider the relationship between acquired signal S and concentration C of the product to be linear. Such a right-hand view illustrates an enlargement for low concentrations C of contrast agent. Whatever the acquisition parameters, the curves S and C merge with the identity line DI represented by a broken line within a zone ZL called linearity, which is gray in color in Figure 5. For low concentrations C, it is therefore possible to use the signals S acquired directly without correcting for acquisition effects, which is generally accepted for tissue signals.However, as indicated by the curves in the left view of Figure 5, the higher the concentration C, the greater the acquisition effect. The approximation accepted for tissue signals cannot therefore be applied to arterial signals characterized by a maximum concentration of the contrast agent. It is therefore necessary to correct the acquisition effect to estimate relevant AIF sampling.

[0011] Currently, several post-processing methods have been proposed in the literature to correct this acquisition effect, such as that disclosed in the article "Physics-informed neural networks for myocardial perfusion MRI quantification", van Herten et al. 2022. Some of them are little used, such as those involving the use of a so-called population AIF (using an AIF database corresponding to a given injection protocol and a type of patient) or those requiring a blood sample concomitant with the acquisition to actually know the quantity of contrast agent present at a given time. On the other hand, let us examine two methods that the invention directly competes with as well as a third that the invention aims to improve.

[0012] A first known method consists of numerically simulating the acquisition effects using physical models derived from Bloch equations. These models require very detailed knowledge of the perfusion sequence and must therefore be re-evaluated for each case. In addition, the acquired signals S for the same perfusion examination will depend on parameters intrinsic to the tissues that generated them, in particular T1 and T2* which must therefore be estimated using other acquisition sequences and evaluated by post-processing methods that necessarily have a bias. The acquisition of additional parameters, whose value varies for each voxel of the image, thus introduces a source of error and a new problem to be solved: multimodal registration between perfusion images and T1 and T2 maps.Finally, all the parameters required for this method are not necessarily stored in clinical routine and are therefore often lost, limiting future reuse of perfusion acquisitions.

[0013] A second method proposes to use a set of acquisition parameters that greatly limit acquisition effects and therefore consider arterial signals as viable AIFs. However, such an approach is at the expense of image resolution. This results in the impossibility of using such images to estimate pharmacokinetic parameters in a very precise manner. Indeed, this acquisition is combined with a conventional perfusion acquisition offering sufficient resolution for the evaluation of these parameters by using the AlF obtained by the arterial signals of the acquisition optimized for this purpose. This method has the advantage of allowing the direct use of signals from the different acquisitions. However, it requires a technician trained in this type of parameterization and the availability of this type of manipulation on the MRI machine, which is currently still rare.Furthermore, this method assumes that both types of acquisitions are obtained simultaneously. In other words, any old perfusion acquisition cannot be affected by this method. However, for more than thirty years, numerous databases using traditional acquisitions have been created. This method leads to their exclusion.

[0014] Finally, a third method, illustrated in Figure 6, proposes to use all the redundant information shared by tissue signals to estimate an AIF sampling. Since the signal of each voxel describing tissue regions of interest, such as regions A1 to A5 illustrated by broken circles as an example in Figure 6, is the convolution product between the I RF specific to each voxel and the AlF shared by all tissue voxels, it is possible to simultaneously estimate the I RF of each voxel and the AlF of all voxels. This approach is referred to as blind deconvolution, since neither the I RF nor the AlF are known.Such a method 100, for example implemented by a processing unit 4 of a medical analysis system illustrated by FIG. 1 or 2 supplied with experimental data 10 and / or 12, consists first of all of a step 110 of selecting tissue signals S and a first arterial input function AIF0 generally in an arterial region A0 symbolized in FIG. 6 by a broken line square, then in a step 120 of iterative minimization of the residual errors between measured tissue signals S and those reconstructed from a parametric model AlFm of arterial input function AIF. The objective is to adjust the parameters of said arterial input function model AlFm to minimize said residual errors. A very wide variety of methods for carrying out this minimization 120 has been described in the literature. They are generally distinguished by the constraints that they impose on said arterial input function model.Some methods are called "unconstrained" because the sampling of the arterial input function can have any shape. This first type of method is rarely used because it is very sensitive to acquisition noise. 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, thus restricting the shapes that an arterial input function can take.

[0015] The choice and number of tissue signals selected in step 110 vary depending on the method chosen. While it has been shown that, theoretically, their number could be reduced to two, including for methods requiring no constraints on the shape of the desired arterial input function, limiting this number of signals to two assumes a very large number of hypotheses that are unrealistic in practice, particularly concerning the effect of noise and the type of perfusion model describing these two signals. As a general rule, said number of tissue signals used in the literature varies from six to twelve.

[0016] Said second step 120 of minimizing errors commonly consists of minimizing the quadratic sum of the residual errors between the signals chosen in step 110 and those reconstructed from the adjustment of the parameters of the arterial input function model and the pharmacokinetic parameters. This is an iterative step 120 which therefore consists of searching for this set of parameters which makes it possible to reduce this overall error. The adjustment is usually done by a gradient descent process and each iteration involves a large number of mathematical operations which are costly in terms of computation time. The stopping criterion which therefore determines the final number of iterations is variable from one study to another but may relate to a minimal change in the error from one iteration to another or even directly to the value of the error.In the literature, we can find methods proposing from fifty to several hundred iterations generating an implementation time of several minutes. Once this second step is completed, the method makes it possible to generate a second arterial input function AIF that we will call "viable" for all the chosen tissue signals. However, by the very nature of the method, the latter only represents the correct form of the second arterial input function AIF sought. It is therefore necessary to implement a step 130 consisting of adjusting the scale of the AlF (sampling height). There are multiple ways in the literature to carry out this scaling or normalization step. For example, we can assume that the desired AlF shares the area under the AIFO curve or, at least, a part little subject to acquisition effects.The second corrected arterial input function AIF' thus obtained can now be used on all or part of the voxels of the perfusion image to estimate in a subsequent step 140 one or more pharmacokinetic parameters Ql and therefore develop relevant information for therapeutic decision-making. Such pharmacokinetic parameters Ql can be the subject of possible constitution of one or more graphical representations and be the subject of an output via a suitable human-machine interface, such as the interface 5 in connection with the medical imaging analysis system according to figure 1 or 2 so that a health personnel can make a diagnosis for example.

[0017] The invention makes it possible to address all or part of the drawbacks raised by the known or previously mentioned solutions.

[0018] Among the many advantages brought by the invention, we can mention the implementation of a post-treatment method which allows:

[0019] - a free form, i.e. unconstrained, for the arterial input function with little sensitivity to acquisition noise;

[0020] - a reduction in the number of tissue signals required to implement the process;

[0021] - a non-iterative execution of the estimation of the arterial input function allowing execution times of the order of a second to be obtained on a standard consumer computing machine versus several minutes according to the state of the art.

[0022] To this end, the invention relates to a method for post-processing a sampled temporal experimental signal resulting from a perfusion acquisition sequence by a medical imaging device and resulting from the passage of a tracer within an elementary volume of an organ, said method being implemented by a processing unit of a medical imaging analysis system. Such a method comprises:

[0023] - a step of selecting a first arterial input function linked to an arterial region of the organ and a set of tissue signals linked respectively to distinct tissue regions of said organ;

[0024] - a step of producing a second arterial input function from said first arterial input function and said set of selected tissue signals; - a step of developing a pharmacokinetic parameter from said second arterial input function and said experimental signal.

[0025] To increase the performance of implementing such a method with respect to known techniques, the step of producing a second arterial input function according to the invention consists of the implementation of elementary operations by said processing unit previously trained according to a process of learning the disruptive effect of the acquisition of a perfusion sequence on arterial signals and the redundancy of the information in connection with such a second arterial input function shared by a set of at least two tissue signals.

[0026] According to a particular embodiment, such a method may comprise a step of correction by scaling of said second arterial input function, produced prior to the implementation of the step of developing a pharmacokinetic parameter from said second arterial input function thus corrected and said experimental signal.

[0027] Advantageously and to offer the possibility to a health personnel to verify the relevance of said second arterial input function produced when said medical imaging system comprises an output human-machine interface, a method according to the invention can comprise a step of developing a graphical representation of said second arterial input function and outputting said graphical representation by said output human-machine interface.

[0028] According to a preferred embodiment, the learning process may consist of deep learning based on minimizing the mean value of the squared errors between real samples of arterial input functions that have made it possible to generate tissue signals and an estimation of these same samples carried out by said learning process, such a learning process being able to be carried out via the Adam optimizer. According to a second subject, the invention relates to a computer program product comprising one or more program instructions executable by the processing unit of a computer, said program instructions being loadable into a non-volatile memory of said computer and the execution of which by said processing unit causes the implementation of a post-processing method according to the invention.

[0029] Likewise, according to a third object, the invention relates to a storage medium readable by a computer comprising the instructions of such a computer program product.

[0030] Finally, the invention further relates to a medical imaging analysis system comprising a processing unit arranged to communicate with the outside world and receive a set of samples of a temporal experimental signal resulting from a perfusion acquisition sequence by a medical imaging device and resulting from the passage of a tracer within an elementary volume of an organ. Such a processing unit has previously been trained according to a previously mentioned learning process and comprises storage means comprising the program instructions of a computer program product according to the invention.

[0031] Other features and advantages will become more apparent upon reading the following description and examining the accompanying figures, including:

[0032] - figure 1, already described, illustrates a medical imaging system comprising a medical imaging platform;

[0033] - figure 2, already described, illustrates a variant of such a medical imaging system comprising a medical imaging platform;

[0034] - Figure 3, already described, describes a cardiac perfusion sequence by medical imaging;

[0035] - Figure 4, already described, illustrates an example of a known process for estimating impulse responses in relation to voxels of interest and therefore producing pharmacokinetic parameters;

[0036] - Figure 5, already described, illustrates the disruptive effect of an acquisition on the linearity of the signal according to the concentration of the contrast product;

[0037] - Figure 6, already described, illustrates a known method for obtaining an estimate of an arterial input function by blind deconvolution;

[0038] - Figure 7 illustrates a method for producing a viable arterial input function by blind deconvolution and ultimately of pharmacokinetic parameters of interest in accordance with the invention.

[0039] As shown in Figure 7, the invention relates to a method 100 for producing a pharmacokinetic parameter of interest (Ql) from one or more sampled temporal experimental signals (10, 12) resulting from an acquisition by a medical perfusion imaging device (1). Such a method (100) can be implemented by the processing unit (4) of such a medical image analysis system. Such a method (100) is based on a learning process (101) of the disruptive effect of the acquisition of the perfusion sequence on the arterial signals combined with learning of the redundancy of the information in connection with said arterial input function shared by a set of at least two tissue signals. The processing unit 4 is thus arranged to comprise a convolution neural network or any other equivalent solution, having been previously configured by the implementation of said learning process 101.

[0040] Indeed, said tissue signals all come from the convolution product of the same AIF arterial input function and their respective IRF impulse responses. It is therefore possible to learn to extract the shared information which is none other than the desired AIF arterial input function. Learning the disruptive effect of the acquisition is done by providing an AIFO arterial signal which is none other than a degraded version of the desired AlF. Thus, it will be a question of learning to "transform" such a selected AIFO arterial signal in such a way that the latter can correspond to the information shared by the tissue signals, avoiding proposing unlikely solutions. In other words, this point favors the stability of the algorithm.This learning process 101 can be carried out using a combination of signals acquired physically on volunteers, patients or even using physical phantoms and / or simulated signals making it possible to generate a wider variety of cases and therefore promoting the ability to extrapolate for the production process of a viable AIF arterial input function sought as such. By using an ADB database thus generated, the learning 101 can for example be done using a deep learning algorithm. The estimation of the parameters of the learning algorithm can be guided by a minimization of the estimation error of the sought AlF and that used to generate the tissue signals of the ADB database.Said estimation can also be guided directly by minimizing the error between said tissue signals from the ADB database and the tissue signals estimated by the algorithm if a step of reconstructing said signals using the AlF' produced by the algorithm is proposed. Thus, the ADB database can be constituted as desired:.

[0041] - sought AIFs, their AIFO transforms subject to acquisition effects and the signals they generate;

[0042] - AIFOs subject to acquisition effects and signals generated by the sought-after AIFs.

[0043] Once the learning step 101 makes it possible to achieve satisfactory performance, the post-processing method 100 can use a convolution neural network (or any other equivalent solution), hereinafter referred to as a “network”, thus trained within the processing unit 4 implementing said method 100. The latter comprises, like the post-processing method 100 described in connection with the third method according to the state of the art, a step 110 of selecting an initial proposal for an arterial input function AIFO and a set of tissue signals S not corresponding to blood vessels or to a vacuum. Such a step 110 can be carried out using different techniques. Such a selection can thus consist of choosing an average arterial input function from a known population.It can also result from a manual selection by the manipulator on the image resulting from the acquisition of a temporal experimental signal sampled from several sampling zones from which this arterial input function will be derived as well as all of the tissue signals, or from the implementation of any post-processing method making it possible to automatically deduce from the image this arterial input function and all of the tissue signals. Said method 100 comprises a step 120 of producing an estimate of a second viable arterial input function AIF for all of the input signals not subject to the acquisition effects of the initial proposal of arterial input function AIF0.Unlike said prior known method 100, step 120 of a method 100 according to the invention consists of producing directly, that is to say without implementing a costly iterative process 120, a second viable arterial input function AIF from elementary operations implemented by a network trained according to the process 101. Step 120 is therefore drastically faster and simpler to implement than the iterative solution according to the state of the art.A method 100 according to the invention may also comprise a step 130 of scaling said second arterial input function AIF produced so that it is used after final correction in a step 140 of producing one or more pharmacokinetic parameters of interest Ql for all or part of the voxels of a perfusion image, the latter being able to be the subject of an output, for example in a graphical form via a human-machine output interface, such as the interface 5 of a medical imaging analysis system SAIM according to FIG. 1 or 2. The invention however provides that the step 130 of normalizing the produced AlF may also be integrated directly into the step 120 directly generating a viable and normalized AIF', i.e. scaled and available to implement the step 140 of producing one or more pharmacokinetic parameters of interest Ql for all or part of the voxels of an infusion image.A relevant scaling factor can be calculated so as to respect a conservation of mass between the produced arterial input function and the said experimental signal, for example by matching a low distortion zone in the two signals.

[0044] Let us examine an example of implementation of the learning process 101 of a network according to the invention which made it possible to validate the relevance of the invention.

[0045] As mentioned earlier, an ADB database can consist of more than three million AlF samples simulated using a random variation of the parameters of the so-called Parker model. Each AlF sample can then be used to generate five tissue signals with different pharmacokinetic parameters. The model used to generate these signals can be the so-called Toft-Ketty model augmented by the addition of a delay between the AlF and tissue sampling.

[0046] A deep learning algorithm is selected to produce a second viable arterial input function AIF in a step 120 of a post-processing method 100 according to the invention. Such a deep learning algorithm is advantageously divided into two branches. The first branch processes the sampling of the arterial input function AIF0. It consists of three layers of convolution networks in one dimension. The second branch processes all the tissue signal samplings. It consists of three layers of convolution networks in two dimensions. The two branches are each terminated by a single-layer simple neural network and then grouped by a three-layer simple neural network. The last layer proposes as output an arterial input function AIF sampling of the same size as that of the initial proposal of arterial input function AIF0.The learning process 101 can be advantageously carried out by a so-called Adam optimizer based on a minimization of the mean value of the squared errors between the real samples of arterial input functions which made it possible to generate the tissue signals and the estimation of these same samples carried out by the learning algorithm. The number of iterations used for this training was thirty covering eighty percent of the generated database, the validation being therefore carried out on the remaining twenty percent. The validation results obtained were, in terms of mean squared error, 0.69% of the peak value of the real AlF samples for the training database as well as for that used for validation. This approach is generally described in the literature as the most efficient.However, this deep learning can be replaced by a table creation method allowing the mapping between an input, represented by an initial arterial input function and a set of tissue signals, and an output representing a second arterial input function not subject to acquisition effects. In this case, the learning consists of setting up such a table.

[0047] In order to ensure clinical viability of a post-processing method 100 according to the invention, this same deep learning algorithm was tested on a cardiac perfusion imaging database by MRI consisting of forty-three elements. These images have the particularity of having been acquired using the second state-of-the-art method described previously. In doing so, the available arterial input function samples were little subject to the acquisition effects associated with each conventional acquisition.By manually selecting the myocardium region to extract five tissue signal samples and the left ventricle region to extract one arterial input function sampling proposal AIF0 highly subject to acquisition effects, the learning algorithm could be evaluated on real data by comparing the arterial input function sampling estimates made by the latter to the samples that were not very subject to acquisition effects. The results were measured in terms of root mean square error and coefficient of determination, denoted R2, and were compared to the values ​​obtained by the arterial input function sampling proposal AIFO highly subject to acquisition effects. Their median values ​​were respectively less than 0.08 in terms of root mean square error, compared to 0.15 of the samplings of the arterial input function subject to acquisition effects, and more than 0.85 in terms of R2, compared to less than 0.3. Thus, in the case of real acquisition, the invention made it possible to obtain a viable and correct estimation of the sampling of the desired AIF arterial input function.

[0048] For the purposes of validating the relevance of the second arterial input function produced in step 120 by the healthcare personnel using a post-processing method 100 according to the invention, the latter may further comprise a step 150 of developing a graphical representation of said second arterial input function AIF before or after correction thereof. Such a step 150 may further consist of causing an output of said graphical representation by a human-machine output interface when the medical imaging analysis system implementing said method 100 comprises such an output interface, like the interface 5 of the system illustrated by FIGS. 1 and 2.

[0049] The invention has been described in connection with a non-limiting example of signals from a cardiac acquisition sequence using a perfusion imaging device. The invention cannot be limited to this single organ examined and can be used to produce pharmacokinetic parameters for any other organ of interest, such as the brain for example.

Claims

CLAIMS Method (100) for post-processing a sampled temporal experimental signal (S, 10, 12) resulting from a perfusion acquisition sequence by a medical imaging device (1) and resulting from the passage 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: - a step (110) of selecting a first arterial input function (AIF0) linked to an arterial region (A0) of the organ and a set of tissue signals (S) linked respectively to distinct tissue regions (A1, A2, A3, A4, A5) of said organ; - a step (120) of producing a second arterial input function (AIF, AIF') from said first arterial input function (AIF0) and said set of selected tissue signals (S); - a step (140) of developing a pharmacokinetic parameter (Ql) from said second arterial input function (AIF') and said experimental signal (S, 10, 12); said method being characterized in that the step (120) of producing a second arterial input function (AIF) consists of the implementation of elementary operations by said processing unit (4), the latter having been previously trained according to a learning process (101) of the disruptive effect of the acquisition of a perfusion sequence on arterial signals and of the redundancy of the information linked to such a second arterial input function shared by a set of at least two tissue signals.

2. Method (100) according to the preceding claim, comprising a step (130) of correction by scaling of said second arterial input function (AIF) produced, prior to the implementation of the step (140) of development of a pharmacokinetic parameter (Ql) from said second arterial input function thus corrected (AIF') and said experimental signal (S, 10, 12).

3. Method (100) according to claim 1 or 2, for which said medical imaging system (SAIM) comprises an output human-machine interface (5), said method (100) comprising a step of developing (150) a graphical representation of said second arterial input function (AIF, AIF') and outputting said graphical representation by said output human-machine interface (5).

4. Method according to any one of the preceding claims, for which the learning process (101) consists of deep learning based on a minimization of the mean value of the quadratic errors between real samples of arterial input functions which made it possible to generate tissue signals and an estimation of these same samples carried out by said learning process.

5. Method according to the preceding claim, for which the learning process (101) is carried out via the Adam optimizer.

6. Computer program product comprising one or more program instructions executable by the processing unit of a computer, said program instructions being loadable into a non-volatile memory of said computer and the execution of which by said processing unit processing causes the implementation of a method according to any one of the preceding claims. Computer-readable storage medium comprising the instructions of a computer program product according to the preceding claim.Medical imaging analysis system (SAIM) comprising a processing unit (4) arranged to communicate with the outside world and receive a set of samples of a temporal experimental signal (S, 10, 12), resulting from a perfusion acquisition sequence by a medical imaging device (1) and resulting from the passage of a tracer within an elementary volume of an organ, said processing unit (4) having previously been trained according to a learning process (101) of the disruptive effect of the acquisition of a perfusion sequence on arterial signals and of the redundancy of the information in connection with an arterial input function shared by a set of at least two tissue signals and comprising storage means comprising the program instructions of a computer program product according to the preceding claim.

Images