Method for determining the presence of at least one data point representative of malignant cell proliferation within an image, device and corresponding program.

Abstract

The disclosure relates to a method for detecting the presence of at least one data point representative of malignant cell proliferation within a histopathological image, said method being implemented within an electronic processing device.Such a process comprises: - a segmentation (S01) of the representative image of the tissue section, yielding a set of tiles of predetermined size, each tile having a position within the representative image of the tissue section; - a determination (S02), for each tile in the set of tiles, of a data point representing the number of malignant cells and a data point representing the total number of cells, these data points being inserted respectively into a first matrix Mp and a second matrix Mt, each element of which is associated with a tile position within the representative image of the tissue section; - a detection (S03) of the presence of malignant cell proliferation within said histopathological image, said detection implementing calculations of data groupings from the first matrix Mp and the second matrix Mt, according to an algorithm for distributing calculations in parallel. Fig. 1.

Description

Title of the invention: Method for determining the presence of at least one data point representative of malignant cell proliferation within an image, device and corresponding program.

[0001] Domain

[0002] The disclosure relates to the field of quantifying malignant cell proliferation. More particularly, the disclosure relates to a method for determining the degree of malignant cell proliferation in representative images of biolabeled tissues taken from one or more patients. Previous art

[0003] The precise quantification of nuclear biomarkers plays an essential role in the therapeutic decision-making process in oncology. This approach relies on the detection and counting of specifically labeled cells using an antigen-antibody reaction applied to tissue samples taken from patients. Once this technique is applied, tumor cells become visually distinguishable, thus allowing medical pathologists to perform an accurate count. For example, with the biomarker "KI67," tumor cells are stained in two variations: "brown" for positive results (cancer cells) and "blue" for negative results. By combining this information, the pathologist can calculate a proliferation index, a crucial measure for assessing cellular activity.It is worth noting that Ki67 is a protein associated with cell proliferation, and its immunostaining is commonly used in pathology to assess the rate of cell proliferation.

[0004] This proliferation index, for example based on the KI67 biomarker, reflects the percentage of positively labeled tumor cells relative to the total number of labeled cells in specific areas of the sample, commonly called "hotspots." These "hotspots" correspond to areas of the image in which a significant number of cells are present, generally defined by a threshold (for example, greater than 500 cells). This approach ensures that the analyzed regions are representative of cellular activity in the tumor, avoiding potential biases related to the selection of less characteristic areas.

[0005] The proliferation index obtained from this quantification allows oncologists to better understand the aggressive nature of the tumor and to make more informed treatment decisions. Highly proliferative tumors are often associated with faster tumor growth, which can influence The decision to resort to more aggressive treatments such as chemotherapy. On the other hand, tumors with a low proliferation index could potentially respond to more targeted treatments.

[0006] In order to calculate the proliferation index, particularly of the KI67 biomarker, more accurately and efficiently, modern approaches based on artificial intelligence and machine learning have been developed. These approaches are essentially based on the global detection of areas of interest in labeled histopathological images (e.g., KI67). More specifically, deep learning methods are used to identify potential areas of proliferation in labeled histopathological images. Thus, the process of training a convolutional neural network (CNN) to detect hotspots in tissue images generally includes the following steps:

[0007] - Data acquisition and preprocessing: Labeled histopathological images are acquired and annotated manually by practitioners;

[0008] - Training of CNNs: The CNN models are trained on the annotated images. The training phase may involve the use of mutual information from images of different types (e.g., H&E and Ki67) of the whole slide to identify hotspots of marker expression (e.g., KI67);

[0009] - Identification of regions of interest (ROI): The CNN(s) formed put in evidence of regions with a high density of positively labeled cells, which are potential hotspots, which allows for the identification of areas of interest for further analysis and quantification.

[0010] - Calculation of the proliferation index: After the identification of hotspots, other CNNs are used to segment tumor cells and calculate the proliferation index in these hotspots.

[0011] Within these processes, deep neural networks are assumed to: highlight regions with a high density of relevant cells; and eliminate any coloring artifacts that could mislead the segmentation phase. The problem with this approach lies primarily in the large amount of computation required to identify global areas and generally in the error rates that can be due to weak training or an insufficient training sample.

[0012] The present technique aims to improve the situation. Summary of the invention

[0013] Thus, the disclosure relates to a method for detecting the presence of at least one data point representative of malignant cell proliferation within an image histopathological, the said process being implemented within an electronic processing device.

[0014] This process comprises at least one iteration of the following steps:

[0015] - cutting of the representative image of the tissue section, delivering a set of predetermined-size tiles, each tile having a position within the representative image of the tissue section;

[0016] - determination, for each tile in the set of tiles, of a data representative of a number of malignant cells and of a data representative of a total number of cells, these data being inserted respectively into a first matrix Mp and into a second matrix Mt, each element of which is associated with a tile position within the image representative of the tissue section;

[0017] - detection of the presence of malignant cell proliferation within said histopathological image, said detection implementing calculations of data groupings of the first matrix Mp and the second matrix Mt, according to an algorithm of parallel calculation distribution.

[0018] Thus, unlike existing processes, which use artificial intelligence to directly determine proliferation zones, the process of the invention works at the cellular level, by counting cells on small image tiles (which is therefore parallelizable for the first time) and then performing calculations of grouping matrix cells based on the number of cells counted, which is faster and more efficient than current methods.

[0019] According to a particular feature, the determination step includes, for a current tile of the set of tiles, a step of determining the data representative of a number of malignant cells of the current tile using a first neural network, previously trained to identify malignant cells.

[0020] According to a particular feature, the determination step includes, for a current tile of the set of tiles, a step of determining the data representative of a total number of cells of the current tile using a second neural network, previously trained in cell identification.

[0021] Thus, more specialized cell search networks are available, which can be trained separately to increase the detection and counting rate of cells within the tissue.

[0022] According to a particular feature, the first and second neural networks form a single neural network. Thus, in this situation, the speed of obtaining cell counts is increased, making it possible to implement the described process in real or near real time.

[0023] According to a particular feature, the step of detecting the presence of malignant cell proliferation within said histopathological image comprises at least one detection iteration comprising the following steps:

[0024] - cutting a current matrix of positively labeled cells into minus three resulting matrices of the positively marked cells, each resulting matrix including an area of ​​overlap with another of the at least two remaining resulting matrices;

[0025] - identical partitioning of a current matrix of the number of cells, delivering to the minus three resulting matrices of the number of cells;

[0026] - obtaining, for each pair of resulting matrices, the marked cells positively and resulting matrices of the number of cells, of a temporary proliferation index, which is recorded within a data structure;

[0027] each pair of resulting matrices forming the current matrices of the next iteration of the detection step until a predetermined stopping parameter is reached.

[0028] This methodology of grouping cell boxes together is effective for distributing calculations over several calculation units in parallel, in an increasing manner, in order to delimit the proliferation zones precisely.

[0029] According to a particular feature, the detection step further includes, after the detection iterations, a ranking step, in descending order, of the temporary proliferation indices so that the highest proliferation indices are associated with the elements of the resulting matrices that identify said less a representative data of malignant cell proliferation within the histopathological image.

[0030] It is thus possible to process the most promising groupings first if it happens that parallel computing resources are limited.

[0031] According to a particular feature, the step of detecting the presence of malignant cell proliferation within said histopathological image includes at least one step of determining, from at least one element, the first matrix Mp, presenting the highest value, of at least one group of contiguous elements whose values ​​allow the calculation of a temporary proliferation index of a value greater than those of other groups of contiguous elements, said group constituting a core for the iterative search of a proliferation zone within said histopathological image.

[0032] Thus, this approach makes it possible to target several potential areas in parallel, depending on the threshold chosen for the selection of the highest values.

[0033] According to a particular feature, the detection step further comprises, subsequent to the determination step of said at least one core of elements, at least one iteration of an expansion step of said at least one kernel so as to maximize the temporary proliferation index obtained by selecting contiguous elements of said first matrix Mp and corresponding contiguous elements of the second matrix Mt.

[0034] According to a particular feature, the kernel expansion iterations are implemented using parallelizable computing means.

[0035] According to another aspect, the disclosure also relates to an electronic device for detecting the presence of at least one data point representative of malignant cell proliferation within a histopathological image. Such a device comprises modules for:

[0036] - cutting of the representative image of the tissue section, delivering a set of predetermined-size tiles, each tile having a position within the representative image of the tissue section;

[0037] - determination, for each tile in the set of tiles, of a data representative of a number of malignant cells and of a data representative of a total number of cells, these data being inserted respectively into a first matrix Mp and into a second matrix Mt, each element of which is associated with a tile position within the image representative of the tissue section;

[0038] - detection of the presence of malignant cell proliferation within said histopathological image, said detection implementing calculations of data groupings of the first matrix Mp and the second matrix Mt, according to an algorithm of parallel calculation distribution.

[0039] According to another aspect, the invention also relates to a computer program capable of implementing the described process and to a data carrier for recording this computer program.

[0040] The determination device has the architecture of a computer. It is equipped with one or more processors capable of executing all types of computer programs, from operating systems to application software, written in compiled or interpreted languages. The various components of the device are interconnected by a communication bus. The device may optionally be equipped with a communication system to communicate via protocols such as Bluetooth, Ethernet, or Wi-Fi with other systems and connect to mobile or fixed telecommunications networks. The device also includes memory components that store the data and programs necessary for its operation.The device is also modified to be able to perform determination operations on a large number of images and manage several thousand simultaneous operations per second, notably through parallel implementation of calculations for grouping areas of cell proliferation.

[0041] Data carriers can be any entity or device capable of storing programs. For example, the carriers can include a storage means, such as a ROM, for example a CD-ROM or a microelectronic circuit ROM, or a magnetic recording means such as a hard drive, or more commonly, flash memory. Alternatively, the carriers can be transmissible media such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means. The programs according to the invention can, in particular, be downloaded from a network such as the Internet. Alternatively, the information carrier can be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the process in question. Description of the drawings

[0042] Other features and advantages of the invention will become more apparent upon reading the following description of a particular embodiment, given by way of simple illustrative and non-limiting example, and the accompanying drawings, among which:

[0043] - [Fig.1] represents the main steps of the detection process;

[0044] - [Fig.2] illustrates the cutting and summation phases of a first example of realization ;

[0045] - [Fig.3] illustrates the phases of core determination and expansion of a second example of implementation;

[0046] - [Fig.4] illustrates a device for implementing the detection process.

[0047] Description of an embodiment

[0048] As explained above, one objective of the disclosure is to enable the acquisition of data representative of a proliferation index that is faster and more efficient than current techniques. Figure 1 illustrates the steps of the detection process that is the subject of the disclosure.

[0049] Such a process includes:

[0050] - the slice (SOI) of the representative image of the IRCT tissue section, delivering a set of predetermined-size ETU tiles, each tile having a position within the representative image of the tissue section;

[0051] - the determination (SO2) (e.g., the counting), for each tile in the set of tiles, a representative data of a number of malignant cells and a representative data of a total number of cells, these data being inserted respectively into a first matrix Mp and into a second matrix Mt, each element of which is associated with a tile position within the representative image of the tissue section;

[0052] - the detection (S03) of the presence of malignant cell proliferation within of said histopathological image, said detection implementing calculations of data groupings of the first matrix Mp and the second matrix Mt, according to an algorithm of parallel calculation distribution.

[0053] Thus, a process is implemented in several successive phases. The second of these phases consists of counting the positively marked cells and counting the negatively marked cells. This counting is performed on a set of tiles obtained from a segmentation of the original image. The tiles are of a fixed size, and their extraction is performed without overlap (this means that there are no overlapping or sliding areas on the tiles). Each tile is provided to a detector, this detector being configured to deliver the total number of positively marked cells and the total number of negatively marked cells within a tile.Thus, at the end of this first phase, a first matrix of the number of positively labeled cells (called the first matrix Mp) and a matrix including the number of negatively labeled cells (or the total number of cells depending on the implementation methods, called the second matrix Mt) are associated with the input histopathological image.

[0054] Each element (cell) of these two matrices Mp and Mt corresponds to a tile of the input histopathological image.

[0055] For example, purely to illustrate the technique used, an input histopathological image could be 4096*8192 pixels in size. Each tile is of a fixed size, for example 128*128 pixels. The first matrix Mp and the second matrix Mt then each have a size of 32*64 cells (elements) respectively, and the values ​​of these elements correspond to cell counts.

[0056] The final phase consists of identifying one or more proliferation zones (hotspots) from the first matrix Mp and the second matrix Mt by grouping the cells (elements, cells) of these matrices. It is important to note that these groupings are performed based on the matrix data, and not on the images or tiles, as is the case with prior art techniques. This aspect of the proposed methodology is important because it allows for efficient and rapid processing of the input histopathological image, avoiding excessive resource consumption.

[0057] At least two different methods are proposed for performing this detection of proliferation zones (hotspots) from the first matrix Mp and the second matrix Mt. In the following, in connection with an example embodiment, a first detection method is presented, then in connection with a second example embodiment, a second detection method is presented.

[0058] Thus, unlike the techniques presented in relation to the prior art, which are based first on the detection of regions of interest (on the basis of histopathological images) and then on the calculation of the proliferation index within these regions of interest, the invention makes it possible to obtain this information in a much simpler and faster way by transforming the initial problem of recognition (identification) of the regions of interest (ROI) into a cell counting problem, which is much faster to process thereafter, this counting making it possible both to identify the hotspot(s) and to determine the proliferation index.

[0059] Counting negatively labeled and positively labeled cells

[0060] A cell counting technique that can be implemented for a standard histopathological image is presented below. This technique is based on the implementation of one or more neural networks, as detailed below. Prior to implementing cell counting within the standard histopathological image, it should be noted that all images (i.e., the standard histopathological image, but also the histopathological images used for training the neural network(s)) were preprocessed to improve image quality, segmentation accuracy, and facilitate subsequent analysis, to prepare the images in such a way as to minimize variations, highlight relevant features (particularly staining), and facilitate accurate and reproducible analysis.

[0061] As stated, in at least one embodiment, cell counting is performed using at least one neural network. This neural network is trained, according to the present embodiment, by providing it with a set of images of predetermined size that have been manually labeled (annotated). Unlike existing techniques, however, the manual labeling was performed at the level of the cell itself (and not at the level of regions of interest). More specifically, the manual labeling was performed by practitioners who identified, within the images presented to them, positively labeled cells, negatively labeled cells, and, in one variant, lymphocytes. These training images preferably have a resolution of 0.25 µm (corresponding to a zoom of approximately 40x) to allow for precise detection of individual cells and their nuclear characteristics.To train the cell detection model, annotations were prepared for each image, indicating the image regions containing Ki67-positive and Ki67-negative cells, as well as the image regions containing lymphocytes. These annotations are provided at the pixel level, meaning they specify which parts of the image belong to each category. These annotations serve as the model's "ground truth" during training, allowing it to learn how to segment. The algorithm correctly identifies Ki67 positive / negative cells as well as lymphocytes. Including lymphocytes allows the algorithm to distinguish them from tumor cells, thus reducing any segmentation bias.

[0062] According to the present embodiment, the training is supervised and performed offline. The images used for training are normalized, and information on the presence of positively labeled cells, all cells, and lymphocytes is provided to the neural network to enable learning. At the end of the training, one or more neural networks are thus available that are capable of performing cell identification as described above.

[0063] According to the present embodiment, in at least one example, after this training, once the models have stabilized, the neural network(s) can be modified to add an output layer dedicated to counting cells of the different categories (positively labeled, negatively labeled, and lymphocytes). In another example, a counting module is additionally implemented, taking as input the classification data from the neural network(s).

[0064] In any event, using the trained neural network(s) as described above, the current histopathological image is used to obtain both the first matrix Mp and the second matrix Mt. To do this, the following method is implemented, comprising:

[0065] - a tissue segmentation step of the current histopathological image, allowing the fabric to be isolated from the rest of the image (especially the background and artifacts); depending on the situation, the segmentation may have been carried out beforehand or may not be necessary;

[0066] - a tiling step in which the segmented histopathological image area containing the tissue is cut into tiles, for example of size 256x256 pixels, the coordinates of each tile in the histopathological (segmented) image being preserved, so as to be able to associate the quantities of cells revealed subsequently with a location in the original image;

[0067] - for each tile, a prediction step, via the network(s) of neurons described above, estimation of the total number of cells (or the number of negative cells) and of the number of positively marked cells, these numbers being recorded, in the cell (the element) of the corresponding matrix, at the location relative to that tile.

[0068] Depending on the operational implementation conditions, the prediction step can be implemented serially or in parallel, in particular when several parallelizable computing resources can be implemented on the detection device.

[0069] In all cases, at the end of this step, we have, for the segmented histopathological image, the first matrix Mp and the second matrix Mt.

[0070] Depending on the embodiment, two different neural networks can be used: for example, one network responsible for detecting positively labeled cells for each tile, and one network responsible for detecting the total number of cells (or one network responsible for detecting the number of negatively labeled cells). This type of configuration allows for greater specialization of the neural networks, and therefore, under certain conditions, for obtaining more satisfactory results for the detection of specific cells, depending, for example, on the tissues from which the histopathological images are extracted.

[0071] Depending on the embodiment, several neural networks can be implemented to identify cells (in particular to detect positively labeled cells). In such a situation, each network is trained separately. The advantage of such a configuration, although more computationally intensive (for both training and inference), is that it allows the individual results (i.e., the number of cells detected) of each neural network to be averaged so that the averaged results are even more representative of the actual number of cells in each tile.

[0072] Identification of hotspots and calculation of proliferation indices

[0073] The second step consists of determining the locations of the hotspots, if any, in the original image.

[0074] In a first embodiment, described in relation to [Fig. 2], this step of determining areas of malignant cell proliferation is implemented by performing a plurality of cuts and summative calculations. More specifically, in this embodiment, the identification of proliferation areas is implemented by performing increasingly fine cuts within the first matrix Mp and the second matrix Mt, each of these two matrices having identical starting sizes. In [Fig. 2], a fictitious example of the matrix Mp is illustrated. The cutting of each matrix is ​​performed alternately horizontally and vertically. At each cut, at least three resulting level 0 submatrices are obtained (Mp01, Mp02, Mp03), each of the resulting matrices having an overlap zone with the preceding resulting matrix.Then, for each of the resulting level 0 matrices, a summation of the number of cells (a summation of the numbers inscribed in the cells of the resulting matrix) is performed (one summation for the resulting level 0 matrix of positively labeled cells and one summation for the resulting level 0 matrix of negatively labeled cells). A temporary proliferation index is then obtained for each "binomial" of resulting level 0 matrices. The... "Binomials" of resulting level 0 matrices are then ranked in descending order of their temporary proliferation indices.Then, for each "pair" of resulting level 0 matrices, either in descending order or in parallel (depending on the operational implementation conditions, particularly when parallel computing resources are available, such as a graphics card), a partitioning (for example, horizontal if the first partitioning was vertical, and vice versa) is carried out according to the same logic (MpOll, Mp012, Mp013): at least three resulting level 1 sub-matrices are obtained (for the current resulting level 0 matrix of positively marked cells and for the current resulting level 0 matrix of negatively marked cells), each of the resulting level 1 matrices having an overlap area with the resulting level 1 matrix that precedes it in the partitioning, and this for each resulting level 0 matrix.A temporary proliferation index is calculated for each of the resulting level 1 matrix pairs, and these pairs are ranked in descending order of their temporary proliferation indices. The procedure is then continued, as in previous iterations, by alternating horizontal and vertical cuts. It is important to note the significance of overlapping areas in this method, as they ensure that proliferation zones are effectively detected. Without this overlap, some proliferation zones might go undetected because they could straddle two cut matrices.The successively obtained temporary proliferation indices are stored in memory to obtain a series of localizable temporary proliferation indices (i.e., associatable with a plurality of elements in the first matrix Mp and the second matrix Mt, and thus ultimately, with a region of the image). In the end, the highest temporary proliferation indices (or those above a predetermined ceiling value) are retained as representative of candidate areas of malignant cell proliferation. These indices can then be compared to thresholds to determine the degree of malignant cell proliferation and, potentially, the associated cancer phases. These comparison operations are not the focus of the technique described here, with the exception of the threshold comparison, which allows for the differentiation of areas of interest from other areas.Taking the purely illustrative example of the first matrix Mp and second matrix Mt, both of which have a size of 32*64 cells (elements) respectively, a first horizontal split is performed, yielding three resulting level 0 matrices of size 32*32 cells (elements). In this example, there is a 16-column overlap between the matrices. For each of the two groups of resulting level 0 matrices (there is one group of three resulting level 0 matrices for the positively marked cells and one group of three resulting level 0 matrices for negatively labeled cells), a summation of the values ​​present in the cells of these matrices is performed and a temporary proliferation index is calculated based on these sums for each pair of resulting level 0 matrices. At the next iteration, the pairs are again cut: for example, for a current level 0 pair (i.e. the resulting level 0 matrices of size 32*32), we obtain two groups of three resulting level 1 matrices of size 16*32 and the previous procedure is repeated, to calculate temporary proliferation indices, until no lower level cutting is possible.Temporary proliferation indices are retained at each segmentation, allowing the determination of the highest temporary proliferation indices among all calculated indices and the retention of only the three highest indices (corresponding to three different regions of the original image). The advantage of this procedure is that even without parallel computing resources, it allows for rapid progress toward determining the proliferation index. With such resources available, this technique has enabled very rapid progress, even for very large images.

[0075] In a second embodiment, described in relation to [Fig. 3], this step of determining areas of malignant cell proliferation is implemented by performing a plurality of localized expansions. More specifically, in this embodiment, one or more area expansions are performed around one or more elements of the two matrices (the area within the first matrix Mp and the second matrix Mt). The starting point of this method is to identify, within one and / or the other of these matrices, a predetermined number of elements (for example, 3), these elements being those whose values ​​(i.e., either the number of negatively labeled cells or the number of positively labeled cells) are the highest (or located above a predetermined ceiling). In other words, the maximum values ​​(or values ​​close to the maxima) within the matrices are sought, for example, the 3 maximum values.According to this technique, these maximum values ​​are searched for in such a way that a predetermined minimum distance is maintained between the locations (within the matrix) of these values. For example, a distance between 5 and 20 can be imposed in the search. This distance depends primarily on the size of the matrix and therefore on the size of the original image before segmentation into tiles (in the example illustrated in [Fig. 3], this distance is not respected for the sake of clarity). This avoids selecting matrix elements that are too close to each other and that are potentially members of the same hotspot. The distance separating two elements (cells) of the matrix is ​​calculated by calculating the square root. The square of the sum of the square of the difference in indices and the square of the difference in ordinates of the two elements. Once the predetermined number of elements is obtained, an iterative expansion method is implemented around these elements. For example, for an original matrix Mp and an element mpij of this matrix, whose value is vpij (this value being one of the three maximum values ​​obtained), an expansion is performed on all or part of the elements whose index values ​​are adjacent to the element mpij (i.e., a kernel of all or part of the index values ​​i-1, i+1, j-1, j+1). These expansion iterations are illustrated by dashed lines of successively smaller sizes to illustrate only two iterations, and only for a purely illustrative example of a matrix Mp. The expansion can be performed element by element or by group of elements, as shown in [Fig. 3].In this example, the expansion is performed in groups of four elements (a group to which element mpij belongs). As in the previous implementation, this expansion is performed using parallel computing resources where such resources exist. Thus, the four possible expansions from element mpij are calculated in parallel. As before, the expansions are performed in a "binary" fashion: the same locations of the matrix elements of the positively labeled and negatively labeled cells are used for the expansion. The expansion includes calculating the sum of the values ​​of contiguous matrix elements: for example, the sum of (vpij, vpi+lj, vpij+1, vpi + lj+1), and includes calculating a temporary proliferation index.After calculating the four expansions of the element mpij, the elements yielding the highest temporary proliferation index are retained and form a "core" of elements for the next expansion iteration. This next expansion is performed according to the same principles as described above, with expansion parameters that may be identical or different (for example, the expansion can be performed by selecting either three additional elements from the matrix that are adjacent to one of the four elements forming the core, or by selecting a complete group of four new elements from the matrix, or by combining these expansion calculations). Regardless of the parameters chosen, this expansion is again preferentially performed using parallel calculations when available.Furthermore, as with the previous example, the "kernels" can be ranked in descending order of their temporary proliferation indices and can also be expanded, even if these are considered less promising. Finally, the expansion process is stopped when the temporary proliferation index decreases successively over a predetermined number of iterations: a temporary hotspot is then considered to have been previously determined, along with the coordinates of the matrix elements used. The values ​​assigned to it for determining this temporary hotspot are retained.

[0076] At the end of all the expansion phases, carried out in parallel or not, a plurality of temporary hotspots is obtained, and the temporary hotspots whose proliferation indices are the highest are retained.

[0077] As explained previously, these techniques demonstrate greater efficiency when the computational steps are implemented in parallel, that is, when the summation calculations are performed in parallel using suitable computing resources, such as parallelizable graphics card processors, whose cores each receive, at each iteration, a resulting matrix (segments) to calculate the sums of the cells present in these matrices. In the absence of such parallel computing resources, these calculations are performed in descending order of temporary proliferation indices, iteratively or recursively, so as to attempt to converge rapidly towards one or more proliferation zones when they exist.Other parallel computation techniques derived from the characteristics of parallelizable matrix computations as described above can, of course, be implemented without departing from the scope of this document. Furthermore, the two techniques presented earlier can, of course, be combined: it is possible, for example, to use the first technique to perform partitioning, and then, based on the partitioning, to use the second technique to perform grouping by kernel; the size of the kernels is modified, but the kernel grouping technique is similar. It is also possible, without exceeding the scope of this document, to generate intermediate matrices at each iteration, these intermediate matrices representing the sum of certain cell values, and to work iteratively on these intermediate matrices to identify a proliferation zone.It should also be noted that, according to this disclosure, it is possible to vary all parameters, whether they are partitioning or grouping parameters, iteratively. This means, for example, that parallel computing methods (whether those exemplified previously or others) can be iteratively implemented by varying parameters (partition size, kernel size, grouping size) until the proliferation zones with the highest proliferation indices are identified, further increasing the advantages of using parallel computing resources.

[0078] Other features and advantages

[0079] Optionally, the method may also include a step of displaying data representative of the number of positively labeled cells, corresponding to each tile of the original histopathological image, since a matrix is ​​available in which each element (box, cell) corresponds to a specific tile of the original image. The value of each element in the matrix represents the number of positive cells detected in the corresponding tile. Such a display step may include:

[0080] - a data association step in which the matrix is ​​used to establish a spatial association between the original histopathological image and the number of positive cells in each tile: each row and each column of the matrix corresponds to a specific tile and the values ​​in the matrix provide quantitative information on the positivity of the cells in those tiles.

[0081] - a coordinate mapping step, in which the matrix elements are referred to the coordinates of the original histopathological image: the row and column indices of each element correspond to the row and column indices of the original image, which allows a direct association to be established between the number of positive cells and the location of the corresponding tiles of the original histopathological image;

[0082] - a step of displaying the positive cell counts on the original image the original histopathological image, for example in the form of annotations (such as the number of positively labeled cells) and / or in the form of a color map (on the annotation, or on the surface of the tile): for example, the use of a color scale where different colors represent different ranges of positive cell counts can provide a quick and intuitive overview of cell proliferation patterns in the original histopathological image.

[0083] A simplified architecture of an electronic detection device capable of performing all or part of the processing described above is presented in relation to [Fig. 4]. An electronic detection device comprises a first electronic module including a memory 41, a processing unit 42 equipped, for example, with a microprocessor, and controlled by a computer program 43. In at least one embodiment, the present technique is implemented in the form of a set of programs installed partly or entirely on the electronic detection device. In at least one other embodiment, the present technique is implemented in the form of a dedicated component (CpX) capable of processing data from the processing units and installed partly or entirely on the processing device.Furthermore, the device also includes communication means (CIE) in the form of network components (Wi-Fi, 3G / 4G / 5G, wired, RFID / NFC, Bluetooth, BLE, LPWAN, VLC, etc.) that allow the device to receive data (I) from entities connected to one or more communication networks and to transmit processed data (T) to such entities. Such a device also includes parallelizable computing means (MCP), which are used to perform matrix calculations, such as... as previously explained, to search for and detect, in the Mp and Mt matrices, the presence of one or more proliferation hotspots within the representative image of the tissue section.

[0084] Such a device comprises, depending on the embodiments, a module for cutting the image of the tissue section into a set of tiles of predetermined size; a module for determining the data representing the number of malignant cells and the data representing the total number of cells, these data being inserted respectively into the first matrix Mp and into the second matrix Mt, each element of which is associated with a tile position within the image representing the tissue section; and a module for detecting the presence of malignant cell proliferation within said histopathological image which implements calculations of grouping data from the first matrix Mp and the second matrix Mt, according to an algorithm for distributing calculations in parallel.

[0085] As explained previously, these modules are implemented through components, software or hardware.

Claims

1. Demands A method for detecting the presence of at least one data point representative of malignant cell proliferation within a histopathological image, said method being implemented within an electronic processing device, the method comprising the following steps: - cutting (SOI) of the representative image of the tissue section, delivering a set of tiles of predetermined size, each tile having a position within the representative image of the tissue section; - determination (SO2), for each tile of the set of tiles, of a data representative of a number of malignant cells and of a data representative of a total number of cells, these data being inserted respectively into a first matrix Mp and into a second matrix Mt of which each element is associated with a tile position within the image representative of the tissue section; - detection (S03) of the presence of malignant cell proliferation within said histopathological image, said detection implementing calculations of data groupings of the first matrix Mp and the second matrix Mt, according to an algorithm for distributing calculations in parallel, comprising at least one detection iteration comprising the following steps: - cutting a current matrix of positively labeled cells into at least three resulting matrices of positively labeled cells, each resulting matrix including an area of ​​overlap with one of the other at least two remaining resulting matrices; - identical partitioning of a current cell count matrix, delivering at least three resulting cell count matrices; - obtaining, for each pair of resulting matrices of positively labeled cells and resulting matrices of the cell count, a temporary proliferation index, which is recorded within a data structure; each pair of resulting matrices forming the current matrices of the next iteration of the detection step (S03) until a predetermined stopping parameter is reached.

2. A method according to claim 1, characterized in that the determination step (S02) comprises, for a current tile of the set of tiles, a step of determining the data representative of a number of malignant cells (vp) of the current tile using a first neural network, previously trained to identify malignant cells.

3. A method according to any one of claims 1 to 2, characterized in that the determination step (S02) comprises, for a current tile of the set of tiles, a step of determining the data representative of a total number of cells (vt) of the current tile using a second neural network, previously trained in cell identification.

4. The method according to claim 3, characterized in that the first and second neural networks form a single neural network.

5. A method according to claim 1, characterized in that the detection step further comprises, subsequent to the detection iterations, a step of ranking, in descending order, the temporary proliferation indices so that the highest proliferation indices are associated with the elements of the resulting matrices which identify said less a representative data of malignant cell proliferation within the histopathological image.

6. A method according to any one of claims 1 to 4, characterized in that the detection step (S03) of the presence of malignant cell proliferation within said histopathological image comprises at least one step of determining, from at least one element, the first matrix Mp, having the highest value, of at least one group of contiguous elements whose values ​​allow the calculation of a temporary proliferation index of a value greater than those of other groups of contiguous elements, said group constituting a core for the iterative search for a proliferation zone within said histopathological image.

7. A method according to claim 6, characterized in that the detection step further comprises, subsequent to the step of determining said at least one core of elements, at least one iteration of an expansion step of said at least one core so as to maximize the temporary proliferation index obtained in selecting contiguous elements of said first matrix Mp and corresponding contiguous elements of the second matrix Mt.

8. A method according to any one of claims 6 and 7, characterized in that the kernel expansion iterations are implemented using parallelizable computing means.

9. Electronic device for detecting the presence of at least one data point representative of malignant cell proliferation within a histopathological image, the device comprising modules for: - cutting the image representative of the tissue section, delivering a set of tiles of predetermined size, each tile having a position within the image representative of the tissue section; - determining, for each tile in the set of tiles, a data point representative of a number of malignant cells and a data point representative of a total number of cells, these data points being inserted respectively into a first matrix Mp and into a second matrix Mt, each element of which is associated with a tile position within the image representative of the tissue section;- detection of the presence of malignant cell proliferation within said histopathological image, said detection implementing calculations of data groupings of the first matrix Mp and the second matrix Mt, according to an algorithm of parallel calculation distribution, which iteratively implements the following means of: - cutting a current matrix of positively labeled cells into at least three resulting matrices of positively labeled cells, each resulting matrix including an area of ​​overlap with another of the at least two remaining resulting matrices; - identical cutting of a current matrix of the number of cells, delivering at least three resulting matrices of the number of cells;- obtaining, for each pair of resulting matrices of positively marked cells and resulting matrices of the number of cells, a temporary proliferation index, which is recorded within a data structure; each pair of resulting matrices forming the current matrices of the next iteration of the detection step (S03) until a predetermined stopping parameter is reached.

10. Computer program comprising instructions for carrying out the method according to any one of claims 1 to 8, when said instructions are executed by a processor of a computer processing circuit.

11. Data carrier on which a computer program according to claim 10 is recorded.

Images