Method for determining an ablation region based on deep learning

Abstract

The invention relates to a method for evaluating, post-treatment, the ablation of a portion of an anatomical area of ​​interest (130) in an individual (110), comprising at least one lesion (165). The evaluation method includes, in particular, a step of automatically determining the contour of the ablation zone using a machine learning method, such as a neural network, which analyzes the post-treatment image of the anatomical area of ​​interest in the individual. This machine learning method is pre-loaded during a training phase using a database containing several post-operative medical images of an identical anatomical area of ​​interest from a group of patients, each medical image in the database being associated with an ablation zone of the anatomical area of ​​interest in that patient.The invention also relates to an electronic device (150) comprising a processor and a computer memory that stores the instructions for said evaluation method.

Description

TECHNICAL FIELD OF THE INVENTION

[0001] The field of the invention is that of the evaluation of a medical intervention.

[0002] More specifically, the invention relates to a method for post-treatment evaluation of a lesion ablation region and prediction of an associated risk of recurrence.

[0003] The invention finds applications, in particular, for evaluating minimally invasive medical interventions by predicting the risk of recurrence of a lesion, such as a tumor or metastasis. Such a minimally invasive medical intervention corresponds, for example, to percutaneous ablation of a lesion, such as a tumor in the liver, lung, kidney, or any other organ. Percutaneous ablation generally involves using imaging to guide the insertion of one or more needles through the skin to reach and destroy a lesion. STATE OF THE ART

[0004] It is known from the prior art of techniques that allow for the evaluation of the effectiveness of an intervention and the prediction of a risk of recurrence of a lesion.

[0005] One such technique involves, for example, determining the extent to which the ablated area overlaps the lesion after its removal. By comparing the volume of the ablated area to the volume of the lesion, it is possible to determine the ablation margins. In practice, it is generally recommended to have ablation margins of at least five millimeters.

[0006] To determine these margins, the volume of the lesion is usually determined during the planning of the intervention and compared to the volume of the ablation area which is segmented by an operator on at least one postoperative image.

[0007] The major drawback is that the volume of the ablation zone is generally determined with little precision, often depending on the surgeon who performed the segmentation. Furthermore, the quality of postoperative images is frequently poor, which contributes to uncertainties in the segmentation. Consequently, establishing a correlation between ablation margins and the risk of lesion recurrence is difficult.

[0008] In order to improve upon previous art techniques, it is known to use methods of automatic segmentation of the ablation region.

[0009] Such a technique is described for example in the scientific publication by Zhang et al, entitled "Detection and Monitoring of Thermal Lesions Induced by Microwave Ablation Using Ultrasound Imaging and Convolutional Neural Networks", published in September 2019. The segmentation method described in this publication makes it possible to calculate the margins of the ablation region by segmenting a pre-operative ultrasound image and a post-operative ultrasound image.

[0010] However, the segmentation method described in this publication cannot predict the risk of recurrence because the accuracy of the automatic segmentation of the ablation region is low. Firstly, the segmentation method is limited to a fixed-size subsampling matrix of an image of the ablation region, typically 4 mm², thus restricting the method's use to small areas. Furthermore, the position of the ablation region must be known to determine the position of the subsampling matrix, making the method difficult to use in the absence of consistent reference points within the subsampling matrix in the preoperative and postoperative images.Finally, due to the nature and quality of two-dimensional ultrasound images, the anatomy of interest can be difficult to define, making the segmentation of the region imprecise, leading in particular to erroneous observations where the ablation region as segmented does not encompass the lesion that is the object of the ablation.

[0011] Furthermore, automatic segmentation methods yield consistent results for homogeneous regions, such as a bone, blood vessels, or a lesion, or when the image comprises a known number of regions. In the case of ablation regions, the segmentation results obtained are inconsistent because ablation regions are highly complex, typically composed of various materials such as gas, necrotic cells, healthy cells, residual contrast agent, calcification, etc. Moreover, segmentation is generally performed on medical images that are typically blurry and have low contrast, making automatic image segmentation challenging.

[0012] The scientific publication "Interactive Volumetry Of Liver Ablation Zones", Egger Jan et al., vol. 5, no. 1 20 October 2015 (2015-10-20), XP055772148, DOI: 10.1038 / srep15373, concerns a method of analyzing a recurrent tumor by monitoring over time the alteration of tissues around the tumor ablation region.

[0013] None of the current systems can simultaneously meet all the required needs, namely to offer a technique that allows for the fine evaluation of an ablation treatment by segmenting an ablation region, particularly a heterogeneous one, with better accuracy, especially from a medical image that is not very clear and / or has low contrast. DESCRIPTION OF THE INVENTION

[0014] The present invention aims to remedy all or part of the drawbacks of the prior art mentioned above.

[0015] To this end, the invention relates to a method for post-treatment evaluation of an ablation of a part of an anatomy of interest of an individual, the anatomy of interest comprising at least one lesion.

[0016] Ablation is performed percutaneously or minimally invasively, generally involving the insertion of at least one needle through the skin to reach and destroy a lesion. Several ablation techniques are possible: radiofrequency, microwave, electroporation, laser, cryotherapy, ultrasound, etc.

[0017] The anatomy of interest may be a liver, a lung, a kidney, or any other organ that may have an injury.

[0018] According to the invention, the post-processing evaluation method comprises the following steps: acquisition of a post-operative medical image of the anatomy of interest of the individual; automatic determination of a contour of the ablation region by a machine learning method, of the neural network type, analyzing the post-processing image of the anatomy of interest of the individual, said machine learning method being previously trained during a so-called training phase on a database containing a plurality of medical images of an identical anatomy of interest of a set of patients, each medical image in the database being associated with an ablation region of the anatomy of interest of said patient.

[0019] Thus, the ablation region is automatically segmented in the postoperative medical image based on prior training from multiple medical images segmented by at least one operator, preferably several operators. It is worth noting that this automatic segmentation by the neural network eliminates the need for an operator experienced in medical image analysis. Furthermore, the segmentation obtained by the neural network is generally more accurate, particularly in the case of three-dimensional image analysis. Finally, automatic segmentation using this method also provides better quality in the case of images with low contrast and / or sharpness.

[0020] In particular embodiments of the invention, the training phase includes a preliminary training step on medical images of an identical anatomy of interest including an unablated lesion.

[0021] Thus, the neural network segments the ablation region more effectively in the individual's postoperative medical image. This surprising effect can be explained by the similarity in shape and position between the lesion and the ablation region. Furthermore, it is worth noting that the number of accessible medical images showing an unablated lesion in a given anatomy of interest is generally greater than the number of accessible medical images acquired after the lesion has been ablated.

[0022] In other particular embodiments of the invention, the post-processing evaluation method also includes a registration step of the post-operative image and a medical image of the anatomy of interest of the individual, acquired before the surgical treatment, called the pre-operative medical image, the registered pre-operative medical image and the post-operative medical image forming a pair of medical images of the anatomy of interest of the individual,

[0023] Thus, an analysis of the position of the ablation region relative to the lesion can be performed.

[0024] A pre-operative medical image is a medical image acquired before ablation treatment, and a post-operative medical image is a medical image acquired after ablation treatment.

[0025] In particular embodiments of the invention, the post-treatment assessment method also includes a step of assessing the risk of recurrence based on a relative characteristic between the ablation region and the lesion, between the ablation region and the anatomy of interest, or between the lesion and the anatomy of interest.

[0026] Thus, the post-treatment evaluation method according to the invention offers medical personnel a better view of the ablation treatment applied to the individual, allowing them to assess the need for additional treatment if the risk of recurrence is proven.

[0027] The risk of recidivism generally takes the form of a binary value, which can be equal to 0 or 1, for example. A positive value is understood when the risk of recidivism is proven, and a negative value is understood when the risk of recidivism is low.

[0028] However, the risk of recidivism can also take the form of a probability between 0 and 1. We will then understand that the risk of recidivism is proven when the value of the risk is greater than a threshold value, for example equal to 0.5.

[0029] In particular embodiments of the invention, when the risk of recurrence is proven, the post-treatment evaluation method also includes a step of determining the position of the recurrence based on a relative characteristic between the ablation region and the lesion, between the ablation region and the anatomy of interest, or between the lesion and the anatomy of interest.

[0030] In particular embodiments of the invention, the risk of recurrence is assessed by taking into account an ablation margin between the ablation region and the lesion.

[0031] For example, the ablation margins are equal to or greater than 5 mm for the recurrence risk value to be negative.

[0032] The ablation margin is generally defined as the smallest distance between the ablation region and the lesion.

[0033] In particular embodiments of the invention, the risk of recurrence is assessed by taking into account a distance between a center of mass of the lesion and a center of mass of the ablation region.

[0034] The reference value for the center of mass depends on the ablation margins. If the ablation margins are equal to 10 mm and the reference value for the ablation margins is 5 mm, the distance between the centers of mass of the lesion and the centers of mass of the ablation region must be less than or equal to 5 mm.

[0035] In particular embodiments of the invention, the risk of recurrence is assessed by taking into account the regularity and sharpness of the edges of the ablation area in relation to the surrounding healthy tissue.

[0036] Healthy surrounding tissue is understood to mean healthy tissue of the anatomy of interest located within the frame of the cropped medical image.

[0037] In particular embodiments of the invention, the risk of recurrence is assessed by taking into account the ratio between the volume of the lesion and the volume of the ablation area.

[0038] In particular embodiments of the invention, the risk of recurrence is assessed by taking into account the position of the lesion in relation to the center of the anatomy of interest.

[0039] In particular embodiments of the invention, the post-treatment evaluation method also includes a step of segmenting the lesion in the pre-operative medical image of the anatomy of interest of the individual.

[0040] In particular embodiments of the invention, the post-treatment evaluation method also includes a step of detecting the lesion in the pre-operative medical image of the anatomy of interest of the individual.

[0041] In particular embodiments of the invention, all or part of the medical images in the database are cropped around the ablation region comprising at least one lesion, the cropping of the images being carried out according to a common frame of predetermined dimensions, the set of centers of the ablation region of the cropped medical images in the database forming a constellation of distinct points within the common frame.

[0042] Thus, by distributing the positions of the ablation regions within the common frame, it is possible to reduce the prediction errors of the machine learning method. If all the ablation regions were in the same position within the common frame, the machine learning method would primarily consider postoperative images with an ablation region in that particular position, leading to prediction errors when the ablation region was in a different position.

[0043] In particular embodiments of the invention, for all medical images in the database, the part of the individual's body included in said image is divided into a plurality of elementary units of a single size, the number of elementary units being divided into two almost equal parts between the part of the human body delimited by the ablation region and the rest of the part of the individual's body included in the image.

[0044] It should be emphasized that the equivalent distribution between the elementary units corresponding to an ablation region and the elementary units to a non-ablation region can be analyzed at the level of an image or globally at the level of all images.

[0045] The elementary units are generally called pixels in the context of two-dimensional images or voxels in the context of three-dimensional images.

[0046] We will understand by almost equal parts when the two sets of elementary units are made up of the same number of elementary units or when the difference in the number of elementary units of each of the two sets is for example less than 5% of the number of elementary units of the two sets.

[0047] In particular embodiments of the invention, the post-operative medical image database includes at least one pre-operative medical image comprising at least one non-ablated lesion.

[0048] Thus, the machine learning method is better trained.

[0049] In particular embodiments of the invention, the post-treatment evaluation process also includes a step of determining a complementary ablation mask when the risk of recurrence is proven.

[0050] Thus, a treatment plan is proposed with the aim of eliminating the risk of recurrence through the post-treatment evaluation process. It should be emphasized that this treatment plan is non-binding and may or may not be followed by medical staff.

[0051] In particular embodiments of the invention, the post-treatment evaluation method also includes a step of planning a trajectory of a medical instrument to a target point in an ablation region defined by the complementary ablation mask.

[0052] It should be emphasized that this planning stage is carried out prior to the additional treatment.

[0053] In particular embodiments of the invention, the post-treatment evaluation method also includes a step to assist in tracking the trajectory planned by an operator of the medical device.

[0054] In particular embodiments of the invention, the planned trajectory and / or a guidance indication is displayed in real time on a screen of an augmented reality device.

[0055] In particular embodiments of the invention, medical images are three-dimensional images.

[0056] It should be noted that a three-dimensional image can correspond to a collection of two-dimensional images taken at generally regular intervals along a predefined axis.

[0057] In particular embodiments of the invention, each post-operative image is acquired using the same image acquisition technique.

[0058] In other words, the technique used to acquire the post-operative medical image is identical to that used to acquire the post-operative medical images from the training database of the machine learning method.

[0059] The invention also relates to an electronic device comprising a processor and computer memory storing instructions for a process according to any one of the preceding embodiments.

[0060] Such an electronic device could be, for example, a control device, a navigation system, a robotic device, or an augmented reality device. The control device could be, in particular, a computer located in the operating room or a remote server. BRIEF DESCRIPTION OF THE FIGURES

[0061] Other advantages, purposes and particular features of the present invention will become apparent from the following non-limiting description of at least one particular embodiment of the devices and methods of the present invention, with reference to the accompanying drawings, in which: there figure 1 is a schematic view of a medical intervention; the figure 2 is a synoptic view of a post-treatment evaluation process for medical intervention figure 1 ; there figure 3 is an example of a three-dimensional medical image in which an ablation region is highlighted, used during the procedure illustrated in figure 2 ; there figure 4 is an example of four medical images, each including an ablation mask manually delineated by an operator and an ablation mask predicted by a neural network of the procedure illustrated in figure 2 ; there figure 5is an example of a medical image in which automatic segmentation of a lesion and an ablation region is performed during a step of the procedure illustrated in figure 2 ; there figure 6 illustrates an example of additional ablation such as may be proposed by the procedure illustrated in figure 2 . DETAILED DESCRIPTION OF THE INVENTION

[0062] The present description is given by way of non-limiting attribution, each feature of an embodiment being able to be advantageously combined with any other feature of any other embodiment.

[0063] It should be noted from the outset that the figures are not to scale. Example of a particular implementation method

[0064] There figure 1This is a schematic view of a medical procedure in which an individual 110 lying on a table 115 is treated using a medical instrument 120. In this non-limiting example of the invention, the medical procedure consists of removing a lesion 165 in an anatomical structure of interest 130, which is here the liver of the individual 110, using the medical instrument 120, which in this case is a semi-rigid needle. The medical procedure is a percutaneous procedure in which the body of the individual 110 is not opened.

[0065] The manipulation of the medical instrument 120 by an operator 140 can be advantageously guided by means of a guidance device which in this non-limiting example of the invention is an augmented reality device such as a headset 150 worn by the operator 140. The medical instrument 120 can also be associated with a robotic medical device 125.

[0066] The helmet 150 includes a translucent screen 155 allowing the operator to see normally. An image is superimposed on the screen 155 to display reference points that guide the operator 140 in manipulating the medical instrument 120 to treat, by ablation, a region 160 called the ablation zone around the lesion 165 identified in the anatomy of interest 130. These reference points may include an ablation mask that has been previously estimated on a medical image 170 of the anatomy of interest 130 acquired before the operation. The medical image 170 will be referred to hereafter as the preoperative medical image 170.

[0067] When the operation is completed, an evaluation of the operative treatment is performed using a post-treatment ablation evaluation method 200, as illustrated synoptically in figure 2and whose instructions are stored in a computer memory 180 of an electronic control device 181 connected to the headset 155, either by cable or wirelessly. The post-treatment evaluation method 200, whose instructions are processed by a computer processor 182 of the electronic control device 181, makes it possible, in particular, to determine an ablation region and an associated risk of recurrence, in order to verify whether the surgical treatment performed during the operation is sufficient or whether it is preferable to continue the treatment by, for example, performing a further ablation.

[0068] It should be noted that the electronic device 181 can be advantageously integrated into the helmet 150.

[0069] The post-processing evaluation procedure 200 includes a first step 210 of acquiring a post-operative medical image of the anatomy of interest 130.

[0070] It should be noted that pre-operative and post-operative medical images are preferably acquired by computed tomography (CT scan). Alternatively, they can be acquired by magnetic resonance imaging (MRI).

[0071] Preferably, the technique used to acquire the pre-operative medical image and the post-operative medical image is similar, or even identical.

[0072] However, the technique used to acquire the post-operative medical image may be distinct from the technique used to acquire the pre-operative medical image.

[0073] Pre-operative and post-operative medical images are advantageously represented in this non-limiting example of the invention as three-dimensionally acquired images. In practice, each three-dimensionally acquired medical image generally corresponds to a collection of two-dimensional medical images, each corresponding to a cross-section of the anatomy of interest, taken at regular intervals along a predetermined axis. A three-dimensional representation of the anatomy of interest can be reconstructed from this collection of two-dimensional medical images. Thus, a three-dimensional image will be understood to refer both to a collection of medical images and to a three-dimensional representation. A voxel will be understood to be an elementary unit relating to the resolution of the three-dimensional image.

[0074] Alternatively, pre-operative and post-operative medical images are each acquired in two dimensions. The basic unit for the resolution of the two-dimensional image is then commonly called a pixel.

[0075] Pre- and post-operative medical images include the entire anatomy of interest or are cropped around the ablation area within a predefined frame. In a three-dimensional image, the frame surrounding the ablation area corresponds to a cube, while in a two-dimensional image, the frame corresponds to a square.

[0076] The frame surrounding the ablation area, also known by the English term " bounding boxThe frame can be automatically generated around the ablation area following an action by the operator. Such an action could, for example, involve the operator indicating a point on the post-operative medical image that belongs to the ablation area, and the frame is then generated around that point. For instance, in minimally invasive ablation treatment for small lesions—for example, lesions of approximately 5 cm + / - 10% in diameter, or preferably approximately 3 cm + / - 10% in diameter—each edge of the cube or each side of the square measures between 5 and 10 cm.

[0077] There figure 3is an illustration of a three-dimensional medical image 300 in which an ablation region 310 is surrounded by a frame 320. The frame 320 is cubic and corresponds to squares 330 on the cross-sectional views 340, 350 and 360, respectively according to the sagittal view, the axial view and the coronal view.

[0078] The post-operative medical image of the anatomy of interest 130 is then analyzed by a neural network, which is a machine learning method, during a second step 220 in order to automatically segment the ablation region in the post-operative medical image of the anatomy of interest 130 of the individual 110.

[0079] To this end, the neural network was first trained on a database of medical images of an identical anatomy of interest, in this case, a liver, from a set of patients during a preliminary training phase. Each medical image in the database includes an anatomy of interest with a function identical to that of the anatomy of interest.

[0080] Advantageously, the post-operative medical image of the anatomy of interest 130 of individual 110 is acquired in the same modality as that of the medical images in the neural network training database.

[0081] When the postoperative medical image of the anatomy of interest 130 of individual 110 is cropped, the dimensions of the cube or square of this postoperative medical image are advantageously identical to those of the cubes or squares used to train the neural network. In other words, the medical images in the database have the same dimensions as the cropped postoperative medical image.

[0082] To train the neural network, the ablation region of each postoperative image in the database, where the lesion was ablated, was first segmented by at least two operators. This was done to increase the relevance of the training and, consequently, the analysis results obtained by the neural network. Indeed, it can be difficult for an operator to delineate an ablation region, especially when the image contrast is low, as can be seen, for example, in the four previously segmented postoperative medical images. figure 4The use of multiple operators annotating medical images thus improves the identification of the ablation region. In this non-limiting example of the invention, the ablation region associated with the registered postoperative image corresponds to the union of the ablation regions proposed by the operators. As an alternative example, the ablation region associated with the registered postoperative medical image could correspond to the intersection, a consensus, or an adjudication of the ablation regions proposed by the operators. Furthermore, the neural network is trained to classify the voxels of a medical image into ablation or non-ablation regions.

[0083] Alternatively, training can be performed using a single expert annotator to delineate ablation regions in medical images. The operator's experience is then crucial so that the neural network can produce well-defined ablation regions.

[0084] It is also worth noting that it is preferable for the database images to contain an equal number of voxels belonging to the ablation region and voxels belonging to a non-ablation region. This proportion is calculated based on the voxel classification determined manually by operators.

[0085] In other words, the part of the individual's body included in each image in the database is divided into a plurality of uniquely sized elementary units, with the number of elementary units being divided into two almost equal parts between the part of the human body delimited by the ablation region and the rest of the individual's body part included in the image.

[0086] Furthermore, it is also preferable that the ablation region not always be centered within the frame in every medical image in the training database. If the ablation region were predominantly centered, a bias would be introduced into the neural network, which would learn that the ablation region is primarily centered, which is not necessarily the case, particularly if an operator makes a frame positioning error. Therefore, a bounded random variable is advantageously added to the frame positions to mitigate this bias of centering the lesion within the frame.

[0087] The entire set of centers in the ablation region of the cropped medical images in the database thus forms a constellation of distinct points within the common frame.

[0088] In order to limit the production of false positives by the neural network, it is also preferable that the database include medical images containing at least one non-ablated lesion.

[0089] Phase 290 of neural network training is generally carried out in several stages: a training step 291; a validation step 292; a testing step 293.

[0090] The medical image database is thus partitioned into three databases containing distinct medical images. These three databases are referred to as the training database, validation database, and test database, respectively. In this non-limiting example of the invention, 60 to 98% of the medical images in the medical image database are grouped in the training database, 1 to 20% in the validation database, and 1 to 20% in the test database. The percentages, which are generally functions of the number of images in the medical image database, are given here for illustrative purposes only.

[0091] The first two steps, 291 and 292, of phase 290 of neural network training are main steps that can be repeated several times. The third testing step is optional.

[0092] During the first step 291 of the training phase 290, a weight W and a bias b for each neuron in the neural network are determined from the medical images in the training base.

[0093] It should be emphasized that the training database may advantageously include medical images comprising at least one non-ablated lesion.

[0094] The second step 292 of the training phase 290 allows the weight W and the bias b previously determined for each neuron of the neural network to be validated from the medical images of the validation database, in order to verify the results of the neural network, in particular the prediction error, i.e. by comparing for medical image of the validation database the ablation region obtained with the ablation region segmented in the medical image extracted from the training database.

[0095] In case the prediction error is too large at the end of this second step, the two training steps 291 and validation steps 292 are implemented again to retrain the neural network by reusing the same medical images, in order to refine the values ​​of the W weights and the b biases of each neuron.

[0096] Alternatively, during the neural network retraining, the first step 291 uses resampling of medical images, considering for training the medical images from the training database and a portion of the medical images from the validation database. The remaining medical images from the validation database are then used to validate the W weights and b biases obtained at the end of the first retraining step.

[0097] It should be emphasized that the neural network can be retrained as many times as necessary until the prediction error is acceptable, i.e., less than a predetermined value.

[0098] When both steps 291 and 292 of the training phase 290 are implemented at least once, the final performance of the neural network can be tested in a possible third test step 293 using the medical images from the test set. These medical images, distinct from the medical images in the training and validation sets, allow verification that the neural network, as configured with the parameters W and b for each neuron, can segment the ablation region with good accuracy in all situations the neural network is likely to encounter. However, unlike the validation step 292, this possible third test step 293 does not result in a new training cycle for the neural network.

[0099] It should be noted that the images used in test step 293 are generally carefully selected to cover different positions and sizes of the ablation region in the anatomy of interest in order to best test the prediction capabilities of the neural network.

[0100] From the postoperative medical image of the anatomy of interest 130, the neural network classifies each voxel as an ablation or non-ablation region. This prediction can take the form of an ablation mask superimposed on the postoperative medical image of the anatomy of interest 130. The ablation mask is generally registered to the voxels belonging to the ablation region predicted by the neural network. It should be noted that the ablation mask is usually delimited by a surface or a contour in the context of a two-dimensional image.

[0101] Advantageously, before the training phase 290, the neural network may have been previously trained, during a so-called pre-training phase 295, on a second database of medical images including medical images showing a lesion of an anatomy of interest of the same type as that of individual 110. This pre-training phase 295 allows for better segmentation of the ablation region in the post-operative medical image of individual 110, by cleverly using the similarity in shape of a lesion and an ablation region, or even in position within the anatomy of interest.

[0102] It is worth noting that the medical images in the second database may have been advantageously pre-segmented by at least one operator. The learning process during phase 295 is similar to that performed during phase 290.

[0103] The post-operative medical image is registered with the pre-operative medical image 170 during a third step 230 of the procedure 200 illustrated in figure 2 The registration process, which establishes the correspondences between anatomical points in the two medical images, is performed using a method known to those skilled in the art. Registration can be performed rigidly, meaning that all points in the images are transformed in the same way, or non-rigidly, meaning that each point in the images can undergo a specific transformation.

[0104] An assessment of the risk of relapse is then carried out during a fourth step 240 comprising four sub-steps 241, 242, 243 and 244.

[0105] During a possible substep 241, the lesion 165 is detected in the pre-operative medical image of the anatomy of interest 130 of the individual 110. This detection can be performed automatically or manually by an operator.

[0106] During substep 242, a segmentation of lesion 165 is performed automatically on the pre-operative medical image of the anatomy of interest 130 of individual 110. Alternatively, the segmentation is performed manually by an operator.

[0107] The lesion is automatically segmented using methods known to a person skilled in the art. For example, segmentation is performed using a histogram-based method of the image, such as the Otsu method, or by a deep learning method, more commonly referred to by the English term " deep learning "

[0108] This substep 242 of segmentation is illustrated by the figure 5which presents a pre-operative medical image 500 in which an automatic segmentation based on a deep learning method is performed to determine the three-dimensional location of the lesion 510 and the ablation region 520. An equivalent result can be obtained by a neural network distinct from the neural network used during step 230.

[0109] An ablation margin is then determined between the lesion segmentation and the previously established ablation mask during substep 243. The ablation margin corresponds to the minimum margin, that is, the minimum distance, measured between the lesion segmentation and the ablation mask. In other words, the ablation margin corresponds to the smallest calculated distance between a point on the lesion and a point in the ablation region and is calculated for all points on the lesion.

[0110] Determining the ablation margins ensures that the ablation area fully covers the lesion.

[0111] Based on the ablation margin value determined in substep 243, a prediction of the risk of recurrence, or even a determination of the location of the recurrence, can be assessed in substep 244 by comparing the calculated ablation margin with reference values ​​for ablation margins associated with a recurrence status, stored in a database. The recurrence status indicates whether a recurrence was observed after the operation or not, possibly with an associated recurrence date. For example, the risk of recurrence can be considered zero when the ablation margins are equal to or greater than 5 mm.

[0112] It should be noted that the prediction of a risk of recurrence of the lesion generally takes the form of a binary value equal for example to 0 (zero or negative risk) or 1 (proven or positive risk).

[0113] In addition to, or as an alternative to, other predictors of recurrence risk besides ablation margins can be used. Recurrence risk and recurrence location can be estimated by weighting some or all of these different predictors. For example, recurrence risk predictors are based on relative characteristics that may include: an ablation margin; a distance between the surface of the lesion, or part of the surface of the lesion, and the ablation region; a distance between the centers of mass of the lesion and the centers of mass of the ablation region; a distance between the surface of the lesion, or part of the surface of the lesion, and the ablation region and the distance between the centers of mass of the lesion and the centers of mass of the ablation region, taking into account the proximity of the capsule of the anatomy of interest, particularly in the case of subcapsular lesions; a regularity of the edges of the ablation region relative to the surrounding healthy tissue; a sharpness of the edges of the ablation region relative to the surrounding healthy tissue; a ratio between the volume of the lesion and the volume of the ablation region; a position of the lesion in the anatomy of interest.

[0114] The reference value for the center of mass depends on the ablation margins. If the ablation margins are equal to 10 mm and the reference value for the ablation margins is 5 mm, the distance between the centers of mass of the lesion and the centers of mass of the ablation region must be less than or equal to 5 mm.

[0115] It is worth noting that the risk of recidivism can advantageously be a continuous value between 0 and 1, rather than a binary value, in order to take into account several risk predictors. A weighting can thus be applied between different risk predictors to obtain a value between 0 and 1. The risk of recidivism is then considered proven when it exceeds a predetermined threshold value, for example, 0.5.

[0116] If the risk is confirmed, the location of the recurrence can be determined to estimate a supplementary ablation mask during a fifth step (250) of the post-treatment evaluation procedure (200). The supplementary ablation mask is considered, for example, to perform additional ablation in an area where the ablation margin is less than a given value, for example, five millimeters.

[0117] During step 250, an automatic identification of a region where the ablation margin is less than a threshold value, for example less than five millimeters, can be performed.

[0118] There figure 6This illustrates an example of additional ablation following treatment of a lesion 600 with a risk of recurrence. The evaluation procedure 200 identified areas where the ablation margins 605 were insufficient between the lesion 600 and the ablation area 610, and generated an additional ablation mask 620. It should be noted that the generation of the additional ablation mask 620 is performed by attempting to limit the ablation area as much as possible. Furthermore, target points 630 to be reached by an ablation needle can then be defined in the mask 620.

[0119] The process 200 may also include a step 260 of planning a trajectory to be followed by the medical instrument 120 associated with either the ablation mask or the complementary ablation mask, in order to guide the operator during the manipulation of the medical instrument 120 during a step 270 of guiding the medical instrument 120 along the planned trajectory.

[0120] An example of a planning method is described in French patent application no. 1914780 entitled " Automated trajectory planning method for a medical intervention "

[0121] It should be emphasized that the guidance in this non-limiting example of the invention is of the visual type by displaying the planned trajectory and / or a guidance indication on the screen 155 of the helmet 150.

[0122] Alternatively, the guidance of the medical instrument 120 can be carried out via a navigation system providing position and orientation information for the medical instrument 120. This can be mechanical guidance via a robotic device coupled to such a navigation system.

[0123] It should be noted that steps 230 to 260 can be repeated until the risk of recurrence is zero or almost zero, or until the ablation margins are sufficient.

Claims

1. Method (200) for the post-treatment evaluation of an ablation of a portion of an anatomical structure of interest (130) of an individual (110), the anatomical structure of interest comprising at least one lesion (165, 510, 600), the ablation of the portion of the anatomical structure of interest being delimited by an ablation region (160), the evaluation method comprising the steps of: • acquiring (210) a post-operative medical image of the anatomical structure of interest of the individual, comprising all or part of the ablation region; • automatically determining a contour of the ablation region using a machine learning method of the neural network type, analysing the post-treatment image of the anatomical structure of interest of the individual, said machine learning method being previously trained in a phase (290), referred to as a training phase, using a database comprising a plurality of post-operative medical images of an identical anatomical structure of interest from a set of patients, each medical image in the database being associated with an ablation region of the anatomical structure of interest of said patient. characterised in that the training phase comprises a prior step of training using medical images of an identical anatomical structure of interest comprising a non-ablated lesion.

2. Post-treatment evaluation method according to the preceding claim, further comprising a step of registering (220) the post-operative image and a medical image (170, 500) of the anatomical structure of interest of the individual acquired before the surgical treatment, referred to as a pre-operative medical image.

3. Post-treatment evaluation method according to any one of the preceding claims, further comprising a step of evaluating a risk of recurrence based on a relative characteristic between the ablation region and the lesion, between the ablation region and the anatomical structure of interest or between the lesion and the anatomical structure of interest.

4. Post-treatment evaluation method according to the preceding claim, further comprising, when the risk of recurrence is demonstrated, a step of determining the position of the recurrence according to a relative characteristic between the ablation region and the lesion, between the ablation region and the anatomical structure of interest or between the lesion and the anatomical structure of interest.

5. Post-treatment evaluation method according to any one of Claims 3 to 4, wherein a relative characteristic is an ablation margin between the ablation region and the lesion.

6. Post-treatment evaluation method according to any one of Claims 3 to 5, wherein a relative characteristic is a centre of mass of the lesion and a centre of mass of the ablation region.

7. Post-treatment evaluation method according to any one of Claims 3 to 6, wherein a relative characteristic is the regularity and sharpness of the edges of the ablation region in relation to the surrounding healthy tissue.

8. Post-treatment evaluation method according to any one of Claims 3 to 7, wherein a relative characteristic is the ratio of the volume of the lesion to the volume of the ablation region.

9. Post-treatment evaluation method according to any one of Claims 3 to 8, wherein a relative characteristic is a position of the lesion in relation to the centre of the anatomical structure of interest.

10. Post-treatment evaluation method according to any one of Claims 2 to 9, further comprising a step of segmenting the lesion in the pre-operative medical image of the anatomical structure of interest of the individual.

11. Post-treatment evaluation method according to any one of Claims 2 to 10, further comprising a step of detecting the lesion in the pre-operative medical image of the anatomical structure of interest of the individual.

12. Post-treatment evaluation method according to any one of Claims 2 to 11, wherein all or some of the training post-operative medical images of the database are cropped around the ablation region comprising at least one lesion, the cropping of the images being carried out using a common frame of predetermined size, the set of the centres of the ablation region in the cropped post-operative medical images forming a constellation of distinct points inside the common frame.

13. Post-treatment evaluation method according to any one of Claims 2 to 12, wherein for the set of post-operative medical images in the database, the part of the individual's body included in the image is divided into a plurality of elementary units of a single size, the number of elementary units being divided into almost equal parts between the part of the human body delimited by the ablation region and the rest of the part of the individual's body included in the image.

14. Post-treatment evaluation method according to any one of Claims 2 to 13, wherein the post-operative medical image database comprises at least one pre-operative medical image comprising at least one non-ablated lesion.

15. Post-treatment evaluation method according to any one of Claims 3 to 14, further comprising a step of determining a supplementary ablation mask when the risk of recurrence is demonstrated.

16. Post-treatment evaluation method according to the preceding claim, further comprising a step of proposing a path to be followed by a medical instrument to a target point of the supplementary ablation mask.

17. Post-treatment evaluation method according to any one of the preceding claims, wherein the medical images are three-dimensional images.

18. Post-treatment evaluation method according to any one of Claims 2 to 17, wherein each post-operative image is acquired using the same image acquisition technique.

19. Electronic device (150) comprising a processor and a computer memory storing instructions of a post-treatment evaluation method according to any one of the preceding claims.

20. Electronic device according to the preceding claim, which may be a control device, a navigation system, a robotic device or an augmented reality device.

Images