WO2024110407A1

WO2024110407A1 - Method for automatically segmenting an organ on a three-dimensional medical image

Info

Publication number: WO2024110407A1
Application number: PCT/EP2023/082439
Authority: WO
Inventors: Paul WEGIEL; Eric GAUDARD SURLIUGA; Michael Baumann; Antoine Leroy
Original assignee: Koelis
Priority date: 2022-11-21
Filing date: 2023-11-20
Publication date: 2024-05-30
Also published as: FR3142275B1; FR3142275A1

Abstract

The invention relates to a method for segmenting a three-dimensional image of at least one organ of a patient, comprising at least the steps of: extracting from the three-dimensional image of the organ a probability map which assigns to each voxel of the image a probability of belonging to a contour of the organ on the image; extracting from the probability map at least one target point based on the probability of belonging to the contour of the organ on the image, each target point being extracted from the probability map by applying to the probability map at least one threshold for the probability of belonging to the contour of the organ on the image; and based on each target point extracted in the preceding step, deforming a deformable model of this organ in order to obtain the three-dimensional segmentation of the organ on the image.

Description

Method for automatic segmentation of an organ on a three-dimensional medical image The present invention relates to the field of medical imaging for purposes for example of diagnosis, such as disease screening, for purposes for example of treatment, such as as the guidance of surgical instruments, etc. The invention relates more particularly to the segmentation of organs on three-dimensional medical images, that is to say the determination of the contour of the organ in such an image. BACKGROUND OF THE INVENTION Several technologies exist for obtaining three-dimensional images of a patient's organ. One of the most used is ultrasound, which allows images to be obtained quickly, possibly intraoperatively, using relatively inexpensive equipment compared to other technologies such as magnetic resonance imaging or MRI. Once the image has been obtained, it may prove useful for the practitioner to identify the outline of the organ to be analyzed in it. This is particularly the case if the outline of the organ is subsequently required to plan an operating procedure, or to merge different images together. This contour identification operation can be carried out manually, with tools allowing the contour of the organ to be drawn directly on the image. However, this is time-consuming, boring, and fails to meet two essential needs: having high precision and reproducibility. Image segmentation algorithms are known that exploit image properties such as intensity discontinuities in the image (Edge-Based Segmentation method based on gradients) or similarities between images. image elements (pixels for two-dimensional images, voxels for three-dimensional images; Region-Based Segmentation method). However, these algorithms are rarely effective on the entire contour of the organ present in the image. Thus, in the case of prostate ultrasound, it is difficult to distinguish the outline of the organ from the surrounding tissues, especially for the lower part of the prostate, mainly due to the low contrast between the prostate and non-prostatic areas. -prostatic, the granularity of the ultrasound images which may present speckles and/or speckles and the short range of gray levels available. OBJECT OF THE INVENTION The invention aims in particular to improve the segmentation of medical images. SUMMARY OF THE INVENTION For this purpose, according to the invention, a method of segmenting a three-dimensional image of at least one organ of a patient is provided, comprising at least the steps of: - extracting from the three-dimensional image of said organ, a probability map which assigns to each voxel of the image a probability of belonging to a contour of the organ on the image; - extract from said probability map, at least one target point based on said probability of belonging to the contour of the organ on the image, each target point being extracted from the probability map by applying to the probability map has at least one probability threshold of belonging to the outline of the organ on the image; - from each target point extracted in the previous step, deform a deformable model of this organ to obtain the three-dimensional segmentation of said organ on the image. The model is said to be deformable in that it can be applied to an image and deformed from at least one target point to correspond at all points to the contour of the organ sought in the image. We thus begin by first searching for at least one target point whose belonging to the contour of the organ present in the image has a probability considered sufficient and the deformable model is applied to the image at hand. from this target point. Thus, the target points whose belonging to the contour is most probable are considered to be exact and the definition of the contour is completed using the anatomical data of the deformable model. We therefore do not take into account the least probable points to apply the deformable model because this could introduce an error in the segmentation, by deforming the deformable model from target points whose probability of belonging to the contour of the organ would be too weak, and therefore for which the risk that these points would not truly belong to the contour of the organ would be too high. This combination significantly improves the segmentation achieved, especially as the number of target points retained by the process is high. In particular, with the invention, only the highest probability values are retained to determine the target points. Thus with the invention, the most doubtful areas of belonging to the organ are removed from the probability map instead of trying to adjust them. The probability map is three-dimensional. The probability map is not only defined by two unique values, one of which would represent belonging to the contour of the organ and the other would represent the absence of belonging to the contour of the organ. Usually, more than a single target point is extracted from the probability map. Usually, a cloud of target points is thus extracted from the probability map. We therefore understand that the threshold makes it possible to exclude all the voxels from the image for which the probability of belonging to the contour of the organ is too low. Optionally, the probability map is extracted using a model distinct from the deformable model. Optionally, the distinct model is defined by an at least partly automatic segmentation method. Optionally, this distinct model is defined by at least one artificial intelligence (or a neural network) trained from segmented images of the organ of at least other people. Optionally, the at least one threshold is global. Optionally, at least one threshold is local. Optionally, the at least one threshold has a value independent of the organ and/or the patient and/or the three-dimensional image and/or the probability map. Optionally, the at least one threshold has a value depending on the organ and/or the patient and/or the three-dimensional image and/or the probability map. Optionally at least one threshold is dynamic. Optionally the at least one dynamic threshold is a local threshold making it possible to select only local probability maxima from the probability map. Optionally, the value of at least one threshold can be modulated manually. Optionally the probability map is extracted via a model distinct from the deformable model, and in which the value of said at least one threshold can be modulated automatically if at least one characteristic of the three-dimensional image considered differs from those images from which the separate model was designed. Optionally, the deformable model is a statistical model of the shape of the organ to be segmented. A statistical shape model is a model which, as its name suggests, is defined from different shapes of the organ considered according to the frequency of these shapes in the population. Optionally, the deformable model has an initial shape defined from the average of the shapes of the same organ considered but of patients other than the targeted patient. Optionally, the initial shape of the deformable model is determined by an imaging technology other than ultrasound. Optionally, the deformable model has an initial shape combined with a segmentation of the same organ considered of this same patient, segmentation obtained from an image acquired prior to an acquisition of the image currently being segmented. Optionally, the segmentation obtained during a previous examination is determined by an imaging technology other than ultrasound. Optionally at least one target point can be added manually to said at least one target point extracted from the probability map. Optionally, if there are a number of target points lower than a predetermined threshold, the process is interrupted. Optionally, in the event of an interruption, the probability map is abandoned and the deformable model is deformed from manually added target points. Optionally, the deformable model is first pre-deformed. Optionally, the deformable model is pre-deformed by being globally pre-aligned with the target points. Optionally, the deformable model is deformed based on only part of the target points used in the pre-deformation step. Optionally: - from each target point extracted from the probability map, we pre-deform the deformable model, - then we select only part of the target points extracted from the probability map to form a subgroup of points -targets having a higher probability of belonging to the outline of the organ on the image than the target points not forming part of the subgroup, - from each target point of the subgroup, we deforms the deformable model. Optionally, the deformation of the deformable model is carried out by an iterative approach. Optionally, at each iteration, each target point is paired with the closest point of the deformable model. Optionally, at each iteration, the deformed model is first deformed globally via an affine transformation (and/or scaled) then deformed locally. Optionally the de- Formation of the deformable model is carried out iteratively by alternating at least one scaling phase with an elastic deformation phase. Optionally, during the scaling phase we also carry out a global deformation, and in which at each iteration, an elastic deformation field resulting from the elastic deformation is decomposed into an affine component and a non-residual deformation. -linear, the affine component being used for the global deformation during the following iteration. Optionally, the iterations are stopped when the distances between the target points and corresponding points of the deformable model are less than a predetermined threshold. Optionally, we estimate the quality of the segmentation obtained. Optionally, the quality of the segmentation obtained is estimated by evaluating a spatial correlation between the probability map and said segmentation. The invention also relates to an installation comprising at least one image-taking device and means for segmenting at least one image acquired by said device according to the method as mentioned above. The invention also relates to a computer program comprising instructions which cause an installation as mentioned above to execute the method as mentioned above. The invention also relates to a computer-readable recording medium, on which the computer program as mentioned above is recorded. Other characteristics and advantages of the invention will emerge on reading the following description of particular and non-limiting implementations of the invention. BRIEF DESCRIPTION OF THE DRAWINGS Reference will be made to the appended drawings, including: [Fig. 1] Figure 1 is a flowchart of a first mode of implementation of the method of the invention; [Fig. 2] Figure 2 is a flowchart of a second mode of implementation of the method of the invention; [Fig. 3] Figure 3 is a view of an ultrasound image of a prostate, on which target points are searched for; [Fig. 4] Figure 4 is a view of the same ultrasound image on which target points have been selected; [Fig. 5] Figure 5 is a view of the same ultrasound image on which a contour has been drawn; [Fig. 6] Figure 6 is a schematic view of an ultrasound device for implementing the method of the invention. DETAILED DESCRIPTION OF THE INVENTION The invention is described here in application to the segmentation of an ultrasound image of a patient's prostate. The invention is however applicable to the image of any organ and/or any patient (animal or human) and/or to any technology for image acquisition of such an organ. The invention is a method for segmenting a three-dimensional image. In the figures, for reasons of simplicity, the images represented are two-dimensional images. With reference to Figures 1 and 2, in the two implementations described here, the segmentation process is based on an Anatomical Model ^^ and an Artificial Intelligence Model ^^^. The Anatomical Model ^^ is developed from at least one image bank ^^ _^ including ^ images of the prostate. The ^ images are optionally images acquired with a technology different from that used for the acquisition of images which must subsequently be segmented and for example by MRI technology. Each image ^ is then segmented. The segmentation is for example carried out manually during human segmentation ^^ _^ . Optionally this segmentation is carried out by a practitioner or another qualified person. Then these segmentations are then processed by computer (C) to calculate an average of said segmentations (denoted average form ^^ subsequently) and deduce the Anatomical Model ^^. The Anatomical Model ^^ is a statistical model of deformable shape which is thus obtained by a statistical study of an average shape of the prostate and its variations. Note that the design of an anatomical model is known in itself and will not be further described here. The Artificial Intelligence Model ^^^ is developed from at least one image bank ^^ _^ comprising K images of the prostate. The image bank ^^ _^ is optionally different from the image bank ^^ _^ used to create the Anatomical Model ^^. Each image ^ is then segmented. The segmentation is for example carried out manually during human segmentation ^^ _^ . Optionally this segmentation is carried out by a practitioner or another qualified person. Preferably, the development of the Artificial Intelligence Model ^^^ is carried out from segmented images whose imaging modality corresponds to that of the images that we wishes to segment subsequently. Thus the ^ images here are images captured with the same technology as that used for the acquisition of the images which must be segmented subsequently, here ultrasound. These segmentations from ^^ _^ train artificial intelligence ^^ to train it to estimate for each voxel of a given image the probability that this voxel belongs to the contour of the prostate present in said image. We thus obtain the artificial intelligence model ^^^. It should be noted that the design and training of artificial intelligence to obtain an artificial intelligence model are known in themselves and will not be further described here. For this purpose, the artificial intelligence model can be obtained via automatic learning (or “machine learning” in English) and for example can be obtained via a network of neurons and for example through deep learning. First implementation To segment an image of a patient's prostate using the method according to the first mode of implementation shown in Figure 1, it is necessary to acquire an image ^^ of the patient's prostate, transmit the image ^^ to a computer system applying the Artificial Intelligence Model ^^^ to the image ^^ received as input. It will be understood that the neural network can be implemented on the same computer system or on a different system. We thus obtain as an output a prediction, in the form of a three-dimensional probability map ^^ in which each voxel of position ^, ^, ^ in the image is assigned a probability ^(^, ^ , ^) belonging to the contour of the prostate present in the image ^^. The probability map ^^ is thus a three-dimensional image. Advantageously, the probability of membership has a value V which is between two limits of a closed interval (for example between 0 and 1 or between 0 and 100). The probability map ^^ thus turns out to be precise because it is not only defined by binary values 0 or 1 but by a range of values V all included between two limits of a closed interval. This probability map ^^ can optionally be presented to an operator. Optionally, the probability of membership can be symbolized by a color code to simplify the understanding of the operator. For example, the greater the probability of belonging to the prostate, the warmer and closer its color is, and the lower the probability of belonging to the prostate, the colder the color. A zero probability of belonging to the prostate is thus symbolized by giving the associated voxel the color black and a certain probability of belonging to the prostate is thus symbolized by giving the associated voxel the color red. Given that certain voxels of the probability map ^^ can have a zero probability of belonging to the prostate (in particular the voxels internal to the prostate, it being recalled here that all the voxels of the image ^^ are concerned by the first segmentation step), this can lead to the fact that the probability map ^^ may not present fully closed contours. After having defined a global threshold ^ of probability of belonging (this global threshold ^ being predefined), corresponding to an acceptable probability of belonging, the computer system executes an algorithm for filtering the voxels of the probability map to select, during a selection operation ^, ^ target points each corresponding to a voxel of position ^, ^, ^ having a probability ^(^, ^, ^) greater than the global threshold ^. The threshold ^ is said to be global because it applies in the same way and with the same value for all voxels. Optionally, the threshold ^ is fixed, unlike a dynamic threshold whose value would be adapted to the probability map CP and/or to the image considered IP. This threshold ^ thus makes it possible to exclude all voxels for which the probability of belonging to the contour is too low. Consequently, the value of the threshold ^ is set for the intended application and the Artificial Intelligence Model ^^^ used at an optimized value. Preferably, we will modify the value of the threshold ^ according to the intended application and the Artificial Intelligence Model ^^^ used in order to set it at a new optimized value. The value of the threshold ^ may in particular be determined by a compromise between the desired precision and sensitivity of the model, established via the analysis of sensitivity-precision curves or ROC curves. The selection operation ^ is thus a filtering operation on the probability map ^^. In this way, we only retain target points presenting a very high probability of belonging to the contour of the prostate in the image ^^. We therefore understand that the target points are in three dimensions. We also understand that at the end of the selection operation ^, we obtain a cloud of target points. Depending on the image ^^, the number of target points thus extracted can be high. Therefore, at the end of the selection operation ^, certain groups of target points can already locally draw the contours of a prostate (in the form of at least one continuous surface and/or at least one surface discontinuous formed by points). However, given that other voxels of the probability map ^^ were discarded (because their probability ^(^, ^, ^) of belonging to the prostate in the image ^^ was too low and therefore lower than the threshold ^), these contours are necessarily partial. In no case, at the end of the selection operation ^, do we obtain a representation of a prostate whose contours are entirely closed. During a deformation operation ^, we apply the Anatomical Model ^^ to the image ^^ by deforming the Anatomical Model ^^ from the ^ target points retained at the end of the operation ^. We understand that the anatomical model ^^ thus makes it possible to determine the contour of the prostate on the image to be segmented based on the ^ target points. For example, the Anatomical Model ^^ draws one or more continuous surfaces between neighboring target points (by interpolating and/or connecting them) and completes said surfaces to draw a closed contour of the prostate by through the anatomical information it contains. The Anatomical Model ^^ only deforms from the target points. The Anatomical Model ^^ thus makes it possible to close the partial contour(s) which were defined by the cloud of target points. The deformation operation ^ can therefore be seen as a step of transformation of the Anatomical Model ^^ on the cloud of target points previously established. The deformation of the Anatomical Model ^^ can be carried out in a single phase or several phases. For example, the deformation of the Anatomical Model ^^ can be carried out by iteration. For example, at each iteration i: - each target point is paired with the closest point of the Anatomical Model ^^, - then the Anatomical Model ^^ is scaled (by scaling we mean that the Anatomical Model ^^ is as a whole enlarged or shrunk to better correspond to the target points – this scaling being particularly interesting for a patient whose organ in question is of much larger or much smaller dimensions than the average and therefore for which the organ considered is of much greater or much less important dimensions than the Anatomical Model ^^) and globally deformed via an affine transformation based on the distance between each target point considered and the point of the Anatomical Model ^^ which is paired with it (for example the transformation is determined so as to minimize the sum of said distances and/or for example the transformation is calculated with the Arun algorithm), - then the Anatomical Model ^ ^ is elastically deformed, under the action of external forces deforming the Anatomical Model ^^ towards the target points and internal forces ensuring the regularity of the surface of the Anatomical Model ^^ (an elastic deformation being well known from prior art, it will not be detailed further, it being understood that the external forces are obtained based on the distance between each target point considered and the point of the Anatomical Model ^^ which is paired with it and the internal forces are obtained based on the Anatomical Model ^^ before deformation and /or on a constraint of having a smooth deformed contour, the deformation being based on the sum of these external and internal forces to deduce a displacement of each of the vertices of the Anatomical Model ^^). The iterations i are stopped when the distances between the target points and the corresponding paired points of the Anatomical Model ^^ are, for all the target points, less than a predetermined threshold. This successively affine then elastic iterative approach allows the elastic deformations to progressively correct the errors of the previous affine transformation, and vice versa. We thus carry out an extrapolation of the contour from the Anatomical Model ^^. This iterative approach is particularly interesting in places where the Anatomical Model ^^ is very far from a target point or a group of target points and/or in places where there are few target points . This iterative approach makes it possible to generate a very good quality extrapolation of the contour from the Anatomical Model ^^ even in places where the Anatomical Model ^^ is very far from a target point or a group of target points and/or in places where there are few target points. We then obtain the segmentation ^ _^^ of the image ^^. Preferably, the deformation operation ^ takes into account at least one mesh consistency constraint when of the deformation of the Anatomical Model ^^: prevent the creation of a fold in the mesh, apply an elastic deformation or fluid mechanics constraint optionally according to a biomechanical model of the organ during the propagation of the deformation, … Optionally, at each iteration i and via Arun's algorithm, during the elastic deformation of the Anatomical Model ^^, we extract an affine transformation. For example, the elastic deformation field used during elastic deformation is decomposed using the Arun algorithm into an affine component and a non-linear residual deformation. It is then possible to use this affine component to scale said model during the following iteration i+1. This allows a coherent global deformation during iteration i+1 even in areas not covered by target points. Optionally, at least one of the internal forces is defined from at least one statistical study on the shape of the prostate (study identical or different from that which allowed the creation of the Anatomical Model ^^). An internal force thus makes it possible to further limit the risk of obtaining an anatomically improbable segmentation ^ _^^ . We understand that even if the Anatomical Model ^^ is deformed from ^ target points, the Anatomical Model ^^ can pass through only part of the n target points. It is possible to refine the selection of target points using a double filtering algorithm. Such double filtering is explained in relation to the second implementation mode below but can also be used with the first implementation mode. We can see in Figure 6 an example of a device usable for examining a patient within the framework of the method of the invention. This is an ultrasound device 10 comprising a housing 11 enclosing a computer system 12 connected to an ultrasound probe 13, a screen 14 and a computer network R. The computer system 12 is a computer which comprises at least one processor and memory. The memory here comprises the models ^^^ and ^^ and a computer program having instructions arranged to implement the method of the invention. Optionally the computer program has instructions arranged to implement the neural network itself implementing the artificial intelligence model ^^^. The ultrasound probe 13 allows the acquisition of images ^^ which will be processed by the computer system 12 and displayed on the screen 14. The network R is used for connection to a server containing for example reports, images of examinations of patients for which the practitioner or the organization within which he practices is responsible... Second implementation In the second mode of implementation shown in Figure 2, we have the Artificial Intelligence Model ^^^ and of the Anatomical Model ^^. We also have an old segmentation ^ _^^ ^{^^^} of the patient's prostate which is currently being examined (by old, we mean that it was carried out during a previous examination). During a combination operation ^^^^, the Anatomical Model ^^ is combined with the old segmentation ^ _^^ ^{^^^} , for example by replacing the average form ^^ contained in the ^^ by this segment. tation ^ _^^ ^{^^^} . Of course other combination possibilities are possible: for example taking an average of MS and this segmentation ^ _^^ ^{^^^} or even a weighted average of the average shape ^^ and this segmentation ^ ^^^ ^ _^ . We thus obtain a combined Anatomical Model ^^′. We note that the old segmentation ^ _^^ ^{^^^} may come from an ultrasound image or obtained via another technology such as an MRI image. As before, we start by acquiring an image ^^ of the patient's prostate. As before, the image ^^ is transmitted to a computer system in order to apply the artificial intelligence model on the image ^^ received as input. This makes it possible to obtain as an output a prediction in the form of a three-dimensional probability map ^^ in which each voxel of position ^, ^, ^ in the image is assigned a probability ^(^, ^, ^ ) belonging to the contour of the prostate. The probability map ^^ is thus a three-dimensional image. Advantageously, the probability of membership has a value V which is between two limits of an interval, the two limits being included (for example between 0 and 1 or between 0 and 100). The probability map ^^ thus turns out to be precise because it is not only defined by binary values 0 or 1 but by a range of values V all included between two limits of an interval, both limits being included. This probability map ^^ can optionally be presented to an operator. Optionally, the probability membership can be symbolized by a color code to simplify the understanding of the operator. For example, the greater the probability of belonging to the prostate, the warmer and closer the color, and the lower the probability of belonging to the prostate, the colder the color. A zero probability of belonging to the prostate is thus symbolized by giving the associated voxel the color black and a certain probability of belonging to the prostate is thus symbolized by giving the associated voxel the color red. Given that certain voxels of the probability map ^^ can have a zero probability of belonging to the prostate (in particular the voxels internal to the prostate, being reminded here that all the voxels of the image ^^ are concerned by the first segmentation step), this can lead to the fact that the probability map ^^ may not present fully closed contours. The computer system then executes an algorithm for filtering the voxels of the probability map to select, during a selection operation ^′, ^ target points for which a high degree of certainty of belonging to the contour has been assigned ( ie the probability ^(^, ^, ^) is high). For example, as for the first implementation, after having defined a fixed global threshold ^ of probability of membership (this global threshold ^ being predefined), corresponding to an acceptable probability of membership, the computer system executes an algorithm for filtering the voxels of the probability map to select, during a selection operation ^, ^ target points each corresponding to a voxel of position ^, ^, ^ having a higher probability ^(^, ^, ^) at the fixed global threshold ^. The inventors were able to observe that the difficulty of the segmentation task varies greatly from one image to another. This is why the use of a single fixed global threshold generally has the effect of over- or under-segmenting the outline of the organs. For this reason, the filtering algorithm preferably performs a second filtering of the filtered voxels, with a second threshold different from the first. As already noted, the second filtering is also applicable to the first implementation, so the following is indeed applicable to the first implementation. The second threshold is a dynamic threshold, ie a threshold specific to the image considered. The second threshold is optionally local, that is to say it is applied to only part of the CP probability map considered or its value depends on the area of the CP probability map to which it is applied, unlike a global threshold which has the same value over the entire CP probability map. For example, because the anterior part of the prostate is more difficult to segment than the posterior part of the prostate, the second threshold may have a lower value when applied in the area of the anterior part of the prostate than in the area of the posterior part of the prostate. The value of the second threshold is therefore adapted automatically (to each new probability map processed) and locally according to the zone(s) considered of a given probability map ^^. Depending on the application considered, this thresholding can exploit the gradient of the prediction to reject areas where the estimated contour is too diffuse or wide for us to be able to consider the segmented area as anatomically correct. During this second filtering step (part of the selection operation ^′), we will search for a reduced number of contour points, for example associated with local maxima of probability of presence of the contour of the organ. in these positions (figure 3 illustrates these local maxima). In practice, the program executed by the computer system proceeds as follows. We carry out a decomposition # of the image ^^ associated with the probability map ^^ into boxes (also called boxes) and for example into boxes of identical dimensions and for example into rectangular parallelepipeds or cubes. Optionally, each box has dimensions $^, $^, $^ with a resolution of $^ = $^ = $^ and for example $^ = $^ = $^ = & with α between 3 and 6 millimeters and by example between 5 and 5.5 millimeters. We say that the probability map considered has a local maximum in a given voxel '^(^ when there exists a neighborhood of '^(^ such that for any voxel in this neighborhood, the associated probability is less than or equal to that of '^(^. For a box, we thus select the voxel of maximum probability '^(^ within said box, to finally extract it as a target point. We repeat this process for a new box for extract a new target point, etc. We stop this process as soon as we manage to obtain ^ target points The number ^ (for example 100) is the number of target points which will be used to deform the model. , this number being predefined. The number ^ is predefined by the manufacturer and/or by the practitioner. Preferably, it is possible to consider that we have an insufficient number of target points, for example if the number of target points retained at the end of this second filtering is less than a predefined minimum threshold (for example by the manufacturer and/or the practitioner). The segmentation process is then interrupted to suggest to the operator to segment the organ without benefiting from the process of the invention. The selection operation ^′ is thus a filtering operation on the probability map ^^. In this way, we only retain target points presenting a very high probability of belonging to the contour of the prostate in the image ^^. We therefore understand that the target points are in three dimensions. We also understand that at the end of the selection operation ^′, we obtain a cloud of target points. Depending on the image ^^, the number of target points thus extracted can be high. Therefore, at the end of the selection operation ^′, certain groups of target points can already locally draw the contours of a prostate (in the form of at least one continuous surface and/or at least one minus a discontinuous surface formed by points). However, given that other voxels in the probability map ^^ were discarded (because their probability ^(^, ^, ^) of belonging to the prostate in the image ^^ was too low and therefore lower than the threshold ^), these contours are necessarily partial. In no case, at the end of the selection operation T′, do we obtain a representation of a prostate whose contours would be completely closed. Considering that we have obtained a sufficient number of target points (for example that ^ points have been obtained), the program begins the deformation operation ^' during which we apply the combined Anatomical Model ^^^^ to the image ^^ by deforming the combined Anatomical Model ^^^^ from the ^ target points retained at the end of the ^′ operation. We understand that the combined anatomical model thus makes it possible to determine the contour of the prostate from the target points. For example, the Combined Anatomical Model ^^^^ draws one or more continuous surfaces between neighboring target points (by interpolating and/or connecting them) and completes said surfaces to draw a closed contour of the prostate through the anatomical information it contains. The combined Anatomical Model ^^^^ only deforms from the target points. The combined Anatomical Model ^^^^ thus makes it possible to close the partial contour(s) which were defined by the cloud of target points. The deformation operation ^′ can therefore be seen as a step of transformation of the combined Anatomical Model ^^^^ on the cloud of target points. The deformation of the combined Anatomical Model ^^^^ can be carried out in a single phase or several phases. For example the deformation of the combined Anatomical Model ^^^^ can be carried out by iteration. For example, at each iteration i: - each target point is paired with the closest Combined Anatomical Model point ^^^^ e, - then the combined Anatomical Model ^^^^ is scaled (by scaling we mean that the combined Anatomical Model ^^^^ is as a whole enlarged or shrunk to better correspond to the target points – this scaling being particularly interesting for a patient whose organ considered is of much larger or much smaller dimensions than the average and therefore for whom the organ considered is of much larger or much smaller dimensions than the combined Anatomical Model ^^^^) and deformed globally via an affine transformation based on the distance between each target point considered and the point of the combined Anatomical Model ^^^^ which was paired with it (for example the transformation is determined so as to minimize the sum of said distances and/or for example the transformation is calculated with Arun's algorithm), - then the combined Anatomical Model ^^^^ is deformed elastically, under the action of external forces deforming the combined Anatomical Model ^^^^ towards the target points and internal forces ensuring the regularity of the surface of the combined Anatomical Model ^^^^ (an elastic deformation being well known from the prior art , it will not be detailed further, it being understood that the external forces are obtained based on the distance between each target point considered and the point of the combined Anatomical Model ^^^^ which is paired with it and the internal forces are obtained based on the combined Anatomical Model ^^^^ before deformation and/or on a constraint of having a smooth deformed contour, the deformation being based on the sum of these external and internal forces to deduce a displacement of each of the vertices of the combined Anatomical Model ^^^^). The iterations i are stopped when the distances between the target points and the corresponding paired points of the combined Anatomical Model ^^^^ are (for at least part of the target points and preferably for all the target points) less than one predetermined threshold. This successively affine then elastic iterative approach allows the elastic deformations to progressively correct the errors of the previous affine transformation, and vice versa. We thus carry out an extrapolation of the contour from the combined Anatomical Model ^^^^. This iterative approach is particularly interesting in places where the combined Anatomical Model ^^^^ is very far from a target point or a group of target points and/or in places where there are few points. targets. This iterative approach makes it possible to generate a very good quality contour extrapolation from the combined Anatomical Model ^^^^ even in places where the combined Anatomical Model ^^^^ is very far from a target point or of a grouping of target points and/or in places where there are few target points. This iterative approach is particularly interesting where the combined Anatomical Model ^^^^ is very far from a target point or a group of target points. Indeed, the combined Anatomical Model ^^^^ ^^ is thus deformed by iteration and in a localized manner to reach this or these target points and/or the area comprising this or these target points. This localized deformation is preferably of the same order of magnitude as the localized deformation carried out in the places where the Anatomical Model combined ^^^^ is close to a target point or a grouping of target points and is distorted in view of this target point or these target points. We then obtain the segmentation ^ _^^ ' of the image ^^. Preferably, the deformation operation ^ takes into account at least one mesh consistency constraint during the deformation of the combined Anatomical Model ^^^^: prevent the creation of a fold in the mesh, apply a constraint of elastic deformation or liquid mechanics during the propagation of the deformation, ... Optionally, at each iteration i and via the Arun algorithm, we extract during the elastic deformation of the Anatomical Model ^^, an affine updating transformation the scale. For example, the elastic deformation field used during elastic deformation is decomposed using the Arun algorithm into an affine component and a non-linear residual deformation. It is then possible to use this affine component to scale said model during the following iteration i+1. This allows a global deformation during iteration i+1 that is consistent even in areas not covered by target points. Optionally, at least one of the internal forces is defined from at least one statistical study on the shape of the prostate (study identical or different from that which allowed the creation of the Anatomical Model ^^). An internal force thus makes it possible to further limit the risk of obtaining an anatomically improbable segmentation ^ _^^ '. In Figure 5, we represent the segmentation ^ _^^ 'obtained using the combined Anatomical Model ^^^^ and the segmentation ^ _^^ which would have been obtained using only the Anatomical Model ^^. Optionally, the computer program preferably carries out an automatic comparison of the segmentation ^ _^^ (in the case of the first implementation), or of the segmentation ^ _^^ ' (in the case of the second implementation) ,with the old segmentation ^ _^^ ^{^^^} . Such a comparison (for example based on the Dice coefficient or the volume difference) makes it possible to warn the practitioner when the two segmentations of the same organ are not coherent with each other. For both implementations, it can be expected that the program allows a modulation of the influence of the prediction from the artificial intelligence model in the deformation of the anatomical model by varying the selectivity of the filtering process described upper. Using a cursor, a box with a value to complete (for example by a percentage) or a check box displayed on screen 14, the practitioner can adjust, in the segmentation obtained , the part of information predicted by the neural network and that completed by the anatomical model. If desired, the practitioner can therefore use only the deformable anatomical model. In a clinical routine, the practitioner can accept or reject the segmentation automatically proposed by the method of the invention whatever the implementation considered. The computer program thus leaves the practitioner with the possibility of modifying the segmentation proposed by the method by: - adding target points manually regardless of the number of target points selected by the method of the invention; - removing target points selected by the method of the invention. Then, the anatomical model is then deformed towards these target points. We have thus described a method making it possible to segment an image in three dimensions. Image segmentation makes it possible to obtain a contour which can be used to carry out, for example, image fusion or to plan an operating procedure. Of course, the invention is not limited to the embodiments described but encompasses any variant falling within the scope of the invention as defined by the claims. In particular, it is possible to combine the two modes of implementation in whole or in part. For example, we can introduce in the first mode of implementation double threshold filtering and/or the use of an old segmentation and/or the algorithm described comprising the steps of decomposing the probability map into a box. and look for a '^(^ for at least one of the boxes. Whatever the implementation considered, the filtering could be global or local. Whatever the implementation considered, the filtering could be fixed or dynamic If the filtering is dynamic, we can adapt the filtering so that for at least a first zone of the probability map, the value of the threshold depends on the value of at least one threshold of at least a second zone of the probability map. probability map close to or adjacent to the first. We could have a dynamic and global filtering and/or a. fixed and local filtering and/or fixed and global filtering and/or a dynamic and local threshold. The examination device may be different from that described by the technology used (MRI instead of ultrasound for example) or by the arrangement and/or choice of its components. The value of at least one threshold could be modulated manually, for example by the practitioner. The segmentation process may rely on another model than the ^^^ to establish the probability map. For example, in place of ^^^, the method could rely on an at least partly automatic segmentation method other than that of relying on the ^^^ model and for example any at least partly automatic segmentation method capable of generating localized information on the reliability of its result. In general, each of the steps of the process of the invention may be automatic or only partly automatic. Although here the statistical shape model has an average shape which is the average of the shapes of the same organ in a set of patients not including the patient considered, the statistical shape model could have an average shape which is the average of the forms of the same organ in a set of patients including the patient considered. Likewise, artificial intelligence could be trained from images of the same organ in a set of patients, including or not the patient in question. The model ^^ and/or the model ^^^ and/or the three-dimensional image to be segmented can be designed using statistical processing of segmented images whose image modality gerie differs from that of ultrasound and for example by MRI technology (T1 sequence and/or T2 sequence and/or DC sequence and/or ADC sequence etc.), by scanner technology (like PET Scan)…, Although here each image ^ and/or each image ^ is segmented manually during the pre-intervention phase, this segmentation can be carried out otherwise. For example for at least one image ^ and/or at least one image ^, the segmentation could be automatic or partly automatic (for example the segmentation is carried out automatically and then controlled and retouched by a human). Although here the image bank used to create the Artificial Intelligence Model is different from the Anatomical Model, the Artificial Intelligence Model can be established from at least one image bank used to establish the Anatomical Model. . The ^ images may be images captured with a technology different from that used for the acquisition of images which must subsequently be segmented such as MRI technology. The ^ images may be images captured with the same technology as that used for the acquisition of the images which must subsequently be segmented. In a preliminary step the practitioner will be able to manually place target points on the contour of the organ present in the image displayed on the screen 14 to initialize the placement of the anatomical model and the evolution of the nearest neighbors from the latter towards these indicated points. This initialization can be automated. Preferably, the target points used to initialize the placement of the deformable model will be different from those used by the present invention to deform the deformable model. The deformable model may have an initial shape defined by the shape of several organs (for example the average of said shapes) and/or by the shape of a single organ (and for example that of the patient obtained during a previous examination ) and/or by an analytical form (for example drawn by a practitioner or any other expert) and/or a combination of said aforementioned forms. In the case where the deformable model is linked to the shapes of several organs, we will therefore speak of a “statistical deformable shape model”. The shape of the patient's organ can be included or excluded to define the deformable model. We could also include only the patient's organ (whose shape varies from one image to another, for example) to create the “statistical deformable shape model”. For the two modes of implementation described above, the process may include one or more additional steps. For example, the method may include an additional step of pre-deformation of the anatomical model followed subsequently by the actual deformation step which has been described previously. For example, during the pre-deformation step, the anatomical model can be roughly deformed to quickly obtain a first deformation. The pre-deformation step is thus robust. Then during the deformation itself, we will continue to deform the anatomical model in a more precise manner as proposed above. Optionally, a first group of target points is thus extracted from the probability map by applying to the probability map with at least a first probability threshold of belonging to the contour of the organ on the image. This first threshold is then relatively unselective in order to be able to create a first group containing a large number of target points. The anatomical model is then pre-deformed using these target points. For example, the anatomical model is deformed so as to align itself with the different target points of the first group, each target point being paired with the nearest point of the anatomical model. This alignment is for example carried out using the Arun algorithm. Then, a second group of target points is extracted from the first group of target points by applying at least a second probability threshold of belonging to the outline of the organ on the image to the first group. This first threshold is more selective than the first threshold. This second threshold is thus defined so as to retain the points having a high probability of belonging to the contour. We then find ourselves in the configuration described with regard to the two embodiments and we can then proceed to the deformation step itself which was described previously (with regard to the two embodiments) based on this second group of target points. This deformation is thus more local and precise. For example, the method includes an additional step of automatic estimation of the quality of the segmentation obtained. For example, such a step could make it possible to provide an indicator of reliability of the result obtained to the operator and for example a quality score Q. For example when Q is below a first predetermined threshold, the segmentation obtained ^ _^^ or ^ _^^ ′ is automatically discarded and/or when Q is between this first predetermined threshold and a second predetermined threshold, the operator is automatically warned that the quality of the segmentation obtained is doubtful. For example Q is calculated by evaluating the spatial correlation between the probability map CP and the segmentation obtained ^ _^^ or ^ _^^ ′. For example, an automatic estimation of the quality of the segmentation ^ _^^ or ^ _^^ ′ is carried out by calculating a quality score Q such that: Q=Q0+α×r with - r the Pearson correlation coefficient, and - Q0and α estimated by a linear regression model, where the explained variable is the Dice coefficient between the segmentation ^ _^^ or ^ _^^ ′ and a segmentation obtained manually, and the explanatory variable is this correlation r. The Pearson correlation coefficient r is optionally calculated between the probability map ^^ and an ideal probability map ^ ⁺ ^ obtained from the final segmentation ^ _^^ or ^ _^^ ′. This ideal probability map ^ ⁺ ^, very strongly correlated with the final segmentation ^ _^^ or ^ _^^ ′ from which it comes, is of the same size as the probability map ^^. It allows the calculation of the Pearson correlation coefficient r (representative of the spatial correlation between the probability map ^^ and the ideal probability map ^ ⁺ ^) without having to binarize the probability map ^^ with an arbitrary threshold which would cause information to be lost. The ideal probability map ^ ⁺ ^ is calculated from the following formula:

where p(i) is a probability value of a voxel i of the ideal probability map ^ ⁺ ^ and δ(i,j) the Euclidean distance between the point associated with voxel i (usually the center of said voxel) and the point associated with voxel j (usually the center of said voxel) which is the contour point of ^ _^^ or ^ _^^ ′ closest to the point associated with voxel i. We normalize the distance δ(i,j) by the maximum Euclidean distance δmax of the voxels of this probability map ^ ⁺ ^.

Claims

CLAIMS 1. Method for segmenting a three-dimensional image of at least one organ of a patient comprising at least the steps of: - extracting from the three-dimensional image of said organ, a probability map which assigns to each voxel of the image a probability of belonging to a contour of the organ on the image; - extract from said probability map, at least one target point based on said probability of belonging to the contour of the organ on the image, each target point being extracted from the probability map by applying to the probability map has at least one probability threshold of belonging to the outline of the organ on the image; - from each target point extracted in the previous step, deform a deformable model of this organ to obtain the three-dimensional segmentation of said organ on the image. 2. Method according to claim 1, in which the probability map is extracted via a model distinct from the deformable model. 3. Method according to claim 2, in which the distinct model is defined by an at least partly automatic segmentation method. 4. Method according to claim 3, in which the distinct model is defined by at least one artificial intelligence trained from segmented images of the organ of at least other people. 5. Method according to one of the preceding claims, in which the at least one threshold is global or local. 6. Method according to one of claims 1 to 5, in which the at least one threshold has a value independent, or dependent, of the organ and/or the patient and/or the three-dimensional image and /or the probability map. 7. Method according to one of claims 1 to 6, in which at least one threshold is dynamic. 8. Method according to claim 7, in which the at least one dynamic threshold is a local threshold making it possible to select only local probability maxima from the probability map. 9. Method according to one of the preceding claims, in which the deformable model is a statistical model of the shape of the organ to be segmented. 10. Method according to one of the preceding claims, in which the deformable model has an initial shape defined from the average of the shapes of the same organ considered but of patients other than the targeted patient. 11. Method according to one of the preceding claims, in which the initial shape of the deformable model is determined by an imaging technology other than ultrasound. 12. Method according to one of the preceding claims, in which the deformable model has an initial shape combined with a segmentation of the same organ considered of this same patient, segmentation obtained from an image acquired prior to an acquisition of the The image currently being segmented. 13. Method according to one of the preceding claims, in which: - from each target point extracted from the probability map, the deformable model is pre-deformed, - then we select only part of the target points extracted from the probability map to form a subgroup of target points having a higher probability of belonging to the contour of the organ on the image than the target points not being part of the subgroup, - from each target point of the subgroup, we deform the deformable model. 14. Method according to one of the preceding claims, in which the deformation of the deformable model is carried out iteratively by alternating at least one scaling phase with an elastic deformation phase. 15. Method according to claim 14, in which during the scaling phase a global deformability is also carried out, and in which at each iteration, an elastic deformation field resulting from the elastic deformation is decomposed into an affine component and a non-linear residual deformation, the affine component being used for the global deformation during the following iteration. 16. Method according to one of the preceding claims, in which the quality of the segmentation obtained is estimated by evaluating a spatial correlation between the probability map and said segmentation. 17. Method according to one of the preceding claims, in which if there is a number of target points less than a predetermined threshold, the method is interrupted. 18. Installation comprising at least one image capturing device and means for segmenting at least one image acquired by said device according to the method according to one of the preceding claims. 19. Computer program comprising instructions which cause a processor of an installation according to claim 18 to execute the method according to one of claims 1 to 17. 20. Computer-readable recording medium on which the computer program according to claim 19 is recorded.