Method for predicting lesion recurrence by image analysis

Through neural network analysis of preoperative and postoperative medical images, combined with machine learning and additional ablation mask technology, the accuracy of the assessment of lesion recurrence risk after minimally invasive medical intervention in the existing technology is solved, providing more accurate recurrence risk Evaluation and treatment recommendations have improved treatment effectiveness.

CN113993475BActive Publication Date: 2024-11-15QUANTUM SURGICAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202180003922.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2020-05-20
Filing Date
2021-05-20
Publication Date
2024-11-15
Estimated Expiration
2041-05-20

AI Technical Summary

Technical Problem

The prior art is difficult to accurately assess the risk of lesion recurrence after minimally invasive medical intervention, especially in medical images with unclear and/or poor contrast, and the accuracy of the automatic segmentation method makes it difficult to establish the risk of ablation margins and lesion recurrence. correlation between.

Method used

Using a neural network-based machine learning method, the preoperative and postoperative medical images are readjusted and analyzed, and the paired images in the training database are used to predict the risk of lesion recurrence, combined with an additional ablation mask to determine the location of recurrence and risk factors. Provides a more accurate risk assessment of recurrence.

Benefits of technology

It provides more accurate lesion recurrence risk estimates without the need for experienced operators, helps medical staff to evaluate recurrence risk and propose additional treatment recommendations, and improves post-treatment evaluation accuracy and treatment effectiveness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113993475B_ABST
    Figure CN113993475B_ABST
Patent Text Reader

Abstract

The present invention relates to a post-treatment evaluation method for ablation of a portion (160) of a target anatomical structure (130) of an individual (110), the target anatomical structure comprising at least one lesion (165). The post-treatment evaluation method specifically comprises the following steps: based on an analysis of a pair of pre-operative (170) and post-operative medical images of the target anatomical structure of the individual, using a machine learning method based on a neural network, automatically evaluating the risk of lesion recurrence of the target anatomical structure of the individual; the method is previously trained in a so-called training phase on a database containing multiple pairs of medical images of the same anatomical structure of multiple individuals, each pair of medical images in the database being associated with the recurrence status of the lesion of the target anatomical structure of the patient. The present invention also relates to an electronic device (150), comprising a processor and a computer memory storing instructions for the evaluation method.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Technical Field of the Invention

[0002] The field of the invention is that of evaluating medical interventions.

[0003] The present invention more particularly relates to a method for assessing the risk of recurrence of an ablated lesion after treatment.

[0004] The present invention is particularly suitable for evaluating minimally invasive medical interventions by predicting the risk of recurrence of lesions, such as tumors or metastases. Such minimally invasive medical interventions correspond, for example, to percutaneous ablation of lesions, such as tumors in the liver, lungs, kidneys or any other organ. Percutaneous ablation generally involves using images to guide one or more needles to be inserted and penetrate the skin to contact and destroy the lesion.

[0005] Description of the Prior Art

[0006] The prior art discloses techniques for assessing the effectiveness of interventions and for predicting the risk of lesion recurrence.

[0007] For example, this technique involves determining, after a lesion has been ablated, the extent to which the ablation zone covers the lesion. By comparing the volume of the ablation zone with the volume of the lesion, the ablation margin can be determined. In practice, an ablation margin of at least 5 mm is generally recommended.

[0008] To determine these margins, the volume of the lesion is generally determined when planning the intervention and compared to the volume of the ablation zone, which is segmented by the operator on at least one postoperative image.

[0009] The main disadvantage is that the volume of the ablation zone is usually determined inaccurately and often depends to some extent on the operator performing the segmentation. In addition, the quality of postoperative images is often poor, which introduces uncertainties in the segmentation. Therefore, the correlation between the ablation margin and the risk of lesion recurrence is difficult to establish.

[0010] In order to improve the prior art techniques, it is known to use methods for automatic segmentation of the ablation zone.

[0011] This technique is described, for example, in a scientific publication by Zhang et al., September 2019, entitled "Detection and Monitoring of Thermal Lesions Induced by Microwave Ablation Using Ultrasound Imaging and Convolutional Neural Networks". The segmentation method described in this publication makes it possible to calculate the edges of the ablation zone by segmenting the pre-operative ultrasound image and the post-operative ultrasound image.

[0012] However, the segmentation method described in this publication does not predict the risk of recurrence, because the accuracy of the automatic segmentation of the ablation zone is low. First, the segmentation method is limited by the subsampling matrix of the ablation zone image, which has a fixed size, generally equal to 4 mm. 2 , so the use of this method is limited to regions of small size. In addition, in order to set the position of the subsampling matrix, the position of the ablation zone must be known, which makes this method difficult to use without constant reference points within the subsampling matrix of pre- and postoperative images. Finally, due to the nature and quality of 2D ultrasound images, the target anatomical structure may be difficult to determine, making the segmentation of this area inaccurate, and in particular leading to incorrect observations, i.e., the segmented ablation zone does not cover the lesion to be ablated.

[0013] Furthermore, automatic segmentation methods give consistent results for homogeneous regions (that is, for example, for bones, vessels or lesions, or when the image contains a known number of regions). For ablation zones, the obtained segmentation results are inconsistent, because ablation zones are very complex regions, generally composed of various materials, such as gas, necrotic cells, healthy cells, residual contrast products, calcifications, etc. Furthermore, segmentation is generally performed on medical images, which are usually blurry and have poor contrast, making it difficult to segment images automatically.

[0014] Currently no single system meets all requirements simultaneously, in particular providing a technique that makes it possible to finely assess the risk of recurrence of previously ablated lesions, in particular based on unclear and / or poorly contrasted medical images, without the need for intervention by an experienced operator. SUMMARY OF THE INVENTION

[0016] The object of the present invention is to remedy all or part of the above-mentioned disadvantages of the prior art.

[0017] To this end, the present invention is directed to a method for post-treatment evaluation of ablation of a portion of a target portion of an anatomical structure of an individual, the target anatomical structure comprising at least one lesion, the ablated portion of the target anatomical structure being referred to as the ablation zone.

[0018] Ablation is percutaneous or minimally invasive, which usually involves inserting at least one needle through the skin to reach and destroy the lesion. A variety of ablation techniques can be used: radiofrequency, microwave, electroporation, laser, cryotherapy, ultrasound, etc.

[0019] The target anatomy may be the liver, lung, kidney, or any other organ susceptible to lesions.

[0020] According to the present invention, the post-treatment evaluation method comprises the following steps:

[0021] - Acquire post-operative medical images of target anatomical structures of an individual;

[0022] - realigning the postoperative image and the medical image of the individual target anatomical structure obtained before the surgical treatment, referred to as the preoperative medical image, the realigned preoperative medical image and the realigned postoperative medical image constituting a pair of medical images of the individual target anatomical structure;

[0023] - Assessing the risk of recurrence of a lesion of a target anatomical structure of an individual using a neural network-based machine learning method, analyzing a pair of medical images of the target anatomical structure of the individual, wherein the machine learning method has been previously trained in a so-called training phase on a database of multiple pairs of medical images of the same target anatomical structure of a group of patients, wherein each pair of medical images in the database is associated with a recurrence status of a lesion of the target anatomical structure of the patient.

[0024] Thus, by performing a very large-scale analysis of paired medical images of a specific anatomical structure (e.g., the liver, when the anatomical structure to be treated is the liver), it is possible to predict the risk of lesion recurrence based on previous clinical data. Such an analysis makes it possible to determine a more precise estimate of the risk of recurrence of the lesion to be ablated, without requiring the presence of an experienced operator. Thus, the post-treatment evaluation method according to the present invention can provide medical personnel with a better view of the treatment applied to an individual, enabling them to assess the necessity of additional treatment in cases where the risk of recurrence is significant.

[0025] Generally, when recurrence is observed on the postoperative examination day, the recurrence status is called a positive value; or when no recurrence is observed on the postoperative examination day, it is called a negative value.

[0026] The risk of recurrence for its part generally takes the form of a probability between 0 and 1.

[0027] Advantageously, the risk of recurrence is assessed at a predetermined date after treatment and when the recurrence status is positive, each pair of medical images in the database is also associated with a recurrence date.

[0028] The date may be, for example, 1 month, 3 months, 6 months, 1 year, 2 years, 5 years, or even 10 years after the ablation treatment.

[0029] The term "pre-operative medical image" is understood to refer to a medical image obtained before an ablation treatment, while the term "post-operative medical image" is understood to refer to a medical image obtained after an ablation treatment.

[0030] In some specific embodiments of the present invention, all or some training image pairs in the database are cropped around ablation zones including at least one lesion after re-adjustment, and the images are cropped within a common frame of a predetermined size, and the center set of ablation zones in the cropped image pairs forms a constellation of different points within the common frame.

[0031] Therefore, by assigning the location of the ablation zones in a common frame, it is possible to reduce the prediction error of the machine learning method. If all ablation zones are at the same location in the common frame, the machine learning method will mainly consider image pairs that have ablation zones at this specific location, thus leading to prediction errors if the ablation zones are at other locations.

[0032] In some specific embodiments of the present invention, for all pre-cropped medical image pairs in the database, the individual body parts contained in the image are divided into multiple basic units of a single size, and the number of basic units is divided into two almost equal parts between the human body part delineated by the ablation zone and the rest of the individual body part contained in the image.

[0033] In other words, the number of basic units is divided into two almost equal parts between the human body part delineated by the ablation zone and the rest of the individual body parts contained in the paired cropped images in the database.

[0034] It should be emphasized that the isodistribution between basic units corresponding to ablation zones and basic units corresponding to non-ablation zones can be analyzed at the level of one image or globally at the level of all images.

[0035] These basic units are often referred to as pixels in the context of two-dimensional images and voxels in the context of three-dimensional images.

[0036] Almost equal parts are to be understood as when both groups of basic units consist of the same number of basic units, or when the difference between the number of basic units in both groups is, for example, lower than 5% of the number of basic units in both groups.

[0037] In some specific embodiments of the present invention, the database of pairs of medical images comprises at least one pair of images without an ablation zone.

[0038] Therefore, machine learning methods are best trained with at least one pair of images without ablation performed.

[0039] In some specific embodiments of the present invention, the post-treatment evaluation method further comprises the step of determining an additional ablation mask when the risk of recurrence is greater than a predetermined threshold.

[0040] Therefore, a treatment recommendation was estimated using the post-treatment evaluation method to improve the risk of relapse. It should be emphasized that this additional treatment recommendation is not binding and may or may not be followed by the medical staff.

[0041] It should be emphasized that the additional ablation mask is a display of the area to be ablated during the additional treatment. When the risk of recurrence is greater than a threshold (e.g., 0.5), the ablation mask is generally generated after the following operations:

[0042] - Lesions and ablation areas were detected and segmented in pre-operative and post-operative images respectively;

[0043] - At least one location of recurrence was identified.

[0044] Furthermore, detection and segmentation can be performed automatically or manually by an operator.

[0045] In some specific embodiments of the present invention, the step of determining an additional ablation mask comprises the sub-step of segmenting the ablation zone in the post-operative image of the individual target anatomical structure.

[0046] Segmentation can be performed automatically or manually by an operator. It should be emphasized that the ablation zone is generally detected in the image beforehand, with at least one operator indicating its location in the image.

[0047] In some specific embodiments of the present invention, the step of determining an additional ablation mask comprises the sub-step of detecting the ablation zone in a post-operative image of the target anatomical structure of the individual.

[0048] In some specific embodiments of the invention, the step of determining an additional ablation mask comprises a sub-step of detecting lesions in pre-operative images of the individual target anatomical structure.

[0049] Therefore, segmentation can be performed around lesions detected in medical images.

[0050] In some specific embodiments of the invention, the step of determining an additional ablation mask comprises a sub-step of determining the location of the recurrence as a function of at least one risk prediction factor selected from the following list:

[0051] - ablation margins between ablation zones and lesions,

[0052] - the distance between the centroid of the lesion and the centroid of the ablation zone,

[0053] - Regularity and definition of the edges of the ablation zone relative to the surrounding healthy tissue,

[0054] - the ratio of the volume of the lesion to the volume of the ablation zone,

[0055] -Location of the lesion relative to the center of the target anatomy.

[0056] In some specific embodiments of the present invention, the medical image is a three-dimensional image.

[0057] It should be emphasized that, for example, a three-dimensional image may correspond to a collection of two-dimensional images taken at generally regular time intervals along a predetermined axis.

[0058] In some specific embodiments of the present invention, each pre-operative image is acquired using a first image acquisition technique, and each post-operative image is acquired using a second image acquisition technique, wherein the first technique and the second technique are the same or different.

[0059] In some specific embodiments of the present invention, the post-treatment evaluation method further comprises a step of proposing a trajectory of the medical instrument to a target point of the ablation zone defined by the additional ablation mask.

[0060] It should be emphasized that this step of proposing a trajectory, which can also be called the step of planning a trajectory, is carried out before additional treatment, especially before any medical action is taken on the individual. It should also be emphasized that this trajectory is not binding and may or may not be followed by the medical staff.

[0061] In some specific embodiments of the present invention, the post-treatment evaluation method further comprises the step of helping the operator to follow the planned trajectory of the medical instrument.

[0062] In some specific embodiments of the present invention, the planned trajectory and / or guidance instructions are displayed in real time on the screen of the augmented reality device.

[0063] The invention also relates to an electronic device comprising a processor and a computer memory storing instructions for the method according to any one of the above embodiments.

[0064] Such an electronic device may be, for example, a control device, a navigation system, a robotic device or an augmented reality device. The control device may in particular be a computer present in the operating room, or a remote server.

[0065] In other words, the present invention relates to an electronic device comprising a processor and a computer memory storing instructions for a post-treatment evaluation method for ablating a portion of a target anatomical structure of an individual, the target anatomical structure comprising at least one lesion, the ablated portion of the target anatomical structure being referred to as an ablation zone, the processor being configured, when executing the instructions, to facilitate:

[0066] - Acquire post-operative medical images of the individual's target anatomy, including all or part of the ablation zone;

[0067] - realigning the postoperative image and the medical image of the individual target anatomical structure obtained before the surgical treatment (referred to as the preoperative medical image), the realigned preoperative medical image and the realigned postoperative medical image constituting a pair of medical images of the individual target anatomical structure;

[0068] - Assessing the risk of recurrence of a lesion of a target anatomical structure for an individual using a neural network-based machine learning method, analyzing a pair of medical images of the target anatomical structure for the individual, wherein the machine learning method has been previously trained in a so-called training phase on a database of multiple pairs of medical images of the same target anatomical structure for a group of patients, wherein each pair of medical images in the database is associated with a recurrence status of the lesion of the target anatomical structure for the patient.

[0069] In some specific embodiments of the present invention, the processor is configured to assess the risk of recurrence at a predetermined date after surgery, and when the recurrence status is positive, each pair of medical images in the database is also associated with a recurrence date.

[0070] In some specific embodiments of the present invention, after rescaling, the processor is further configured to crop all or some of the image pairs in the database around the ablation zones contained in the post-operative images of all or some of the training image pairs, and the images are cropped within a common frame of a predetermined size, and the center set of ablation zones in the cropped image pairs forms a constellation of different points within the common frame.

[0071] In some specific embodiments of the present invention, for all pre-cropped medical image pairs in the database, the processor is further configured to divide the individual body parts contained in the image into a plurality of basic units of a single size, and the number of basic units is divided into two almost equal parts between the human body part delineated by the ablation zone and the rest of the individual body parts contained in the image.

[0072] In some specific embodiments of the present invention, the database of pairs of medical images comprises at least one pair of images without an ablation zone.

[0073] In some specific embodiments of the present invention, the processor is further configured to determine an additional ablation mask when the risk of recurrence is greater than a predetermined threshold.

[0074] In some specific embodiments of the present invention, the processor is further configured to detect ablation zones in post-operative images of the target anatomical structure of the individual when determining the additional ablation mask.

[0075] In some specific embodiments of the present invention, the processor is further configured to segment the ablation zone in the post-operative image of the individual target anatomical structure when determining the additional ablation mask.

[0076] In some specific embodiments of the present invention, the processor is further configured to detect lesions in pre-operative images of the target anatomical structure of the individual when determining the additional ablation mask.

[0077] In some specific embodiments of the present invention, the processor is further configured to segment lesions in pre-operative images of the individual's target anatomical structure when determining the additional ablation mask.

[0078] In some specific embodiments of the present invention, when determining the additional ablation mask, the processor is further configured to determine the location of the recurrence as a function of at least one risk prediction factor selected from the following list:

[0079] - ablation margins between ablation zones and lesions,

[0080] - the distance between the centroid of the lesion and the centroid of the ablation zone,

[0081] - Regularity and definition of the edges of the ablation zone relative to the surrounding healthy tissue,

[0082] - the ratio of the volume of the lesion to the volume of the ablation zone,

[0083] -Location of the lesion relative to the center of the target anatomy.

[0084] In some specific embodiments of the present invention, the medical image is a three-dimensional image.

[0085] In some specific embodiments of the present invention, each pre-operative image is obtained using a first image acquisition technique, and each post-operative image is obtained using a second image acquisition technique, wherein the first technique and the second technique are the same or different. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Other advantages, objects and specific features of the invention will become apparent from the following non-limiting description of at least one particular embodiment of the apparatus and method forming the subject of the invention, with reference to the accompanying drawings, in which:

[0088] - Figure 1 is a schematic diagram of medical intervention.

[0089] - Figure 2 yes Figure 1 An overview of post-treatment evaluation methods for medical interventions in.

[0090] - Figure 3 is an example of a 3D medical image with the ablation zone highlighted, which is used for Figure 2 The method shown.

[0091] - Figure 4 are examples of four medical images, each containing an ablation mask manually delineated by an operator and a Figure 2 Ablation masks predicted by the neural network of the shown method.

[0092] - Figure 5is an example of a medical image, where Figure 2 In one step of the method shown, lesions and ablation zones are automatically segmented.

[0093] - Figure 6 An example of additional ablation is shown, possibly by Figure 2 The method shown is proposed. DETAILED DESCRIPTION OF THE INVENTION

[0095] The present description is not restrictive and each feature of one embodiment can be combined in an advantageous manner with any other feature of any other embodiment.

[0096] From now on, it is important to note that these drawings are not made to scale.

[0097] Example of a specific implementation

[0098] Figure 1 1 is a schematic diagram of a medical intervention in which an individual 110 lying on a table 115 is being treated by a medical instrument 120. In a non-limiting example of the present invention, the medical intervention corresponds to ablation of a lesion 165 in a target anatomical structure 130, in this case the liver of the individual 110, by the medical instrument 120, in this case a semi-rigid needle. In this case, the medical intervention is a percutaneous procedure in which no incision is made on the body of the individual 110.

[0099] The operation of the medical instrument 120 by the operator 140 can be advantageously guided by a guiding device, which in a non-limiting example of the present invention is an augmented reality device, such as a helmet 150 worn by the operator 140. The medical instrument 120 can also be associated with a robotic medical device 125.

[0100] The helmet 150 comprises a semi-transparent screen 155, which enables the operator to see normally. The image is displayed in superimposed form on the screen 155 so as to display markings, making it possible to guide the operator 140 to operate the medical instrument 120 with the image so as to treat by ablation an area 160 around a lesion 165 determined in the target anatomical structure 130, called the ablation zone. In particular, the markings may comprise an ablation mask, which is estimated in advance on a medical image 170 of the target anatomical structure 130 obtained before the operation. The medical image 170 is hereinafter referred to as the preoperative medical image 170.

[0101] After the procedure is completed, the surgical treatment is evaluated using the method 200 for post-treatment evaluation of ablation, such as Figure 2As shown in the overview of FIG. 1 , its instructions are stored in a computer memory 180 in an electronic control device 181 connected to the helmet 155 by a cable or wireless technology. The post-treatment evaluation method 200, whose instructions are processed by a computer processor 182 of the electronic control device 181, specifically makes it possible to determine the risk of recurrence associated with the lesion to verify whether the surgical treatment performed during the operation is sufficient, or whether it is more appropriate to continue the treatment, for example by performing additional ablations.

[0102] It should be emphasized that the electronic device 181 can advantageously be integrated into the helmet 150 .

[0103] The post-treatment evaluation method 200 includes a first step 210 of acquiring a post-operative medical image of the target anatomical structure 130 .

[0104] It should be emphasized that the preoperative and postoperative medical images are preferably obtained using computed tomography. As an alternative, they can be obtained using a magnetic resonance imaging device.

[0105] The techniques used to acquire the pre-operative medical images and the post-operative medical images are preferably similar, or even identical.

[0106] In other words, advantageously, the technique used to acquire the pre-operative and post-operative medical images is the same technique used to acquire the medical images in the training database of the machine learning method.

[0107] However, the technology used to acquire postoperative medical images may be different from the technology used to acquire preoperative medical images.

[0108] In this case, advantageously, the technique used to acquire the pre-operative and post-operative medical images is the same as the technique used to acquire the medical images in the training database for the machine learning method.

[0109] In a non-limiting example of the present invention, advantageously, the preoperative and postoperative medical images are images acquired in three dimensions. In practice, each medical image acquired in three dimensions typically corresponds to a set of two-dimensional medical images, each image corresponding to a cross section of the target anatomical structure 130 taken at fixed time intervals along a predetermined axis. A three-dimensional representation of the target anatomical structure can be reconstructed from this set of two-dimensional medical images. Therefore, the term three-dimensional image will be understood to refer to the set of medical images and the three-dimensional representation. The term voxel will be understood as a basic unit related to the resolution of a three-dimensional image.

[0110] As an option, both preoperative and postoperative medical images are acquired in two dimensions, and the basic unit associated with the resolution of a two-dimensional image is usually referred to as a pixel.

[0111] Preoperative and postoperative medical images are images that contain the entire target anatomy or are cropped around the ablation zone within a predefined frame. In a three-dimensional image, the frame around the ablation zone corresponds to a cube. On the other hand, in a two-dimensional image, the frame corresponds to a square.

[0112] A frame around the ablation zone, also called a "bounding box", can be automatically generated around the ablation zone following an operation of the operator. For example, such an operation can correspond to the fact that the operator indicates a point belonging to the ablation zone in a post-operative medical image, around which the frame is subsequently generated. By way of example, in the case of minimally invasive ablation therapy of small lesions, that is to say, for example, for lesions of the order of 5 cm + / - 10% of the maximum diameter, or even preferably 3 cm + / - 10% of the maximum diameter, each edge of the cube or each side of the square measures 5-10 cm.

[0113] Figure 3 3 is an illustration of a three-dimensional medical image 300 in which an ablation zone 310 is surrounded by a frame 320. The frame 320 is cubic and corresponds to a square 330 in cross-sectional views 340, 350, and 360 along the sagittal view, along the axial view, and along the coronal view, respectively.

[0114] exist Figure 2 In the second step 220 of the illustrated method 200, the postoperative medical image is realigned with the preoperative medical image 170. The realignment to find a match between anatomical points of the two medical images is performed using methods known to those skilled in the art. The realignment can be performed rigidly, that is, all points in the image are transformed in the same way, or non-rigidly, that is, each point in the image can have a specific transformation.

[0115] The two realigned pre-operative and post-operative images form a pair of medical images of the target anatomical structure 130 of the individual 110 .

[0116] In a third step 230, a pair of medical images of the target anatomical structure 130 is analyzed by a neural network, a machine learning method, in order to automatically assess the risk of recurrence associated with the ablation therapy.

[0117] To this end, the neural network has been previously trained on a database of paired medical images of the same target anatomical structure (thus, in this case, the liver) of a group of patients during a preliminary training phase 290. Each pair of medical images includes a preoperative image and a postoperative image having the same function as the target anatomical structure 130.

[0118] Advantageously, the pair of medical images of the target anatomical structure 130 of the individual 110 is acquired using the same technique as the paired medical images in the training database of the neural network. In other words, the preoperative medical image is acquired using a first medical imaging technique and the postoperative medical image is acquired using a second medical imaging technique, which may be the same as or different from the first medical imaging technique. Using the same technique (possibly with different parameters) for the same type of medical images, i.e., the preoperative medical image and / or the postoperative medical image, may improve the results obtained by the neural network by reducing the bias associated with the different medical imaging techniques.

[0119] When the medical images of the target anatomy 130 of the individual 110 are cropped, advantageously the dimensions of the cube or square of these medical images are the same as the dimensions of the cube or square used to train the neural network.

[0120] To train the neural network, each pair of medical images in the database is associated with the recurrence status of the liver lesion of that person, which indicates whether a recurrence has occurred and may be associated with the date of recurrence. It should be emphasized that each of these pairs of medical images is usually from a different person.

[0121] Typically, the stage 290 of training a neural network is performed in several steps:

[0122] - training step 291;

[0123] - verification step 292;

[0124] -Test step 293.

[0125] Therefore, the paired medical image database is divided into three databases, each containing a different pair of medical images. The three databases are referred to as a training database, a validation database, and a test database, respectively. In a non-limiting example of the present invention, 60-98% of the paired medical images in the medical image database are grouped into the training database, 1-20% are grouped into the validation database, and 1-20% are grouped into the test database. These percentages are generally related to the number of images in the medical image database and are given here by way of illustration.

[0126] The first two steps 291 and 292 of the phase 290 of training the neural network are the main steps and can be repeated several times. The third test step is optional in its part.

[0127] In the first step 291 of the training phase 290, the weights W and biases b of each neuron of the neural network are determined from pairs of medical images in the training database.

[0128] It should be emphasized that advantageously, the training database may contain image pairs without ablation zones.

[0129] Furthermore, the image set in the database preferably contains as many voxels belonging to the ablation zone as to the non-ablation zone. This ratio is calculated based on the classification of the voxels determined manually by the operator.

[0130] In other words, for all cropped medical image pairs in the database, the individual body parts contained in the images are divided into a plurality of basic units of a single size, and the number of basic units is divided into two almost equal parts between the human body part delineated by the ablation zone and the rest of the individual body parts contained in the images.

[0131] Almost equal parts are to be understood as when both groups of basic units consist of the same number of basic units, or when the difference between the number of basic units in both groups is, for example, lower than 5% of the number of basic units in both groups.

[0132] Furthermore, in some of the rescaled medical image pairs, the location of the ablation zone can be advantageously randomly shifted relative to the center of the image in a random manner to avoid introducing bias into the neural network, which would recognize that the ablation zone is primarily an area in the center of the frame. Specifically, there are some medical images around the ablation zone that are incorrectly framed by the operator. Therefore, advantageously, a bounded random variable is added to the location of the frame to limit this bias in locating the lesion in the center of the frame.

[0133] The second step 292 of the training phase 290 makes it possible to verify the weights W and biases b previously determined for each neuron of the neural network from pairs of medical images in the validation database, and thus to verify the results of the neural network, in particular the prediction error, that is, by comparing each pair of medical images in the validation database (i.e. the risk of recurrence obtained using the neural network) with the recurrence state associated with this pair of medical images.

[0134] If the prediction error is too large at the end of the second step, the training 291 and validation 292 steps are implemented again in order to retrain the neural network by repeatedly using the same medical image pairs to improve the values ​​of the weights W and bias b of each neuron.

[0135] As an option, in the process of retraining the neural network, the first step 291 resamples the medical images by considering the medical image pairs in the training database and some of the medical image pairs in the validation database for training. Then, the remaining medical image pairs in the validation database are used to validate the weights W and biases b obtained at the end of the first retraining step.

[0136] It should be emphasized that the neural network can be retrained as many times as needed until the prediction error is acceptable, i.e., smaller than a predetermined value.

[0137] When the two steps 291 and 292 of the training phase 290 have been implemented at least once, the final performance of the neural network can be tested with pairs of medical images from the test database in a possible third test step 293. These pairs of medical images, different from those in the training database and the validation database, make it possible to validate the neural network (such as the parameters W and b configured for each neuron), making it possible to accurately predict the risk of recurrence in all the situations that the neural network is likely to face. However, unlike the validation step 292, this possible third test step 293 does not lead to a new training cycle of the neural network.

[0138] It should be emphasized that the images used in the testing step 293 are typically carefully selected to cover a variety of locations and sizes of ablation zones in the target anatomy in order to optimally test the predictive ability of the neural network.

[0139] The prediction of the risk of recurrence associated with surgical treatment was obtained directly using a neural network that had been previously trained on a database of medical image pairs, each of which was associated with a recurrence status.

[0140] For example, when the lesion does not recur within a certain period after ablation, the associated recurrence status is equal to 0 (negative status); or when the lesion recurs within a certain period after ablation, the associated recurrence status is equal to 1 (positive status). The period associated with the evaluation of the recurrence status is preferably 6 months. However, when the recurrence status is positive, the recurrence status on different dates after ablation can also be evaluated by associating the recurrence date. For example, the neural network can also be trained to predict the recurrence status associated with a pair of medical images every 6 months after ablation or 1 month, 3 months, 6 months, 1 year, 2 years, 5 years or 10 years after ablation.

[0141] Typically, the risk value for lesion recurrence takes the form of a probability between 0 and 1.

[0142] If the value of the risk of recurrence is greater than a predetermined threshold, that is, a recurrence is likely to occur within a given period, additional treatment may be proposed by the post-treatment evaluation method 200 by evaluating an additional ablation mask evaluated in the fourth step 240 of the post-treatment evaluation method 200. Such additional, non-binding treatment corresponds, for example, to additional ablations in areas where the ablation margin is less than a given value (e.g., 5 mm).

[0143] The step 240 of generating additional ablation masks generally includes five sub-steps. Figure 2 The numbers are 241 to 245.

[0144] In sub-step 241, lesions 165 are detected in a pre-operative medical image of target anatomical structure 130 of individual 110. This detection may be performed automatically or manually by an operator.

[0145] In sub-step 242, an ablation mask is generated by a second neural network around at least one lesion of the target anatomical structure based on learning from medical images of a second database, wherein most medical images in this second database contain an ablation zone that was previously detected and segmented on a patient exhibiting a liver lesion.

[0146] Figure 4 Four medical images 400 from the second database are shown. Each medical image 400 is manually annotated by an operator who delineated an ablation zone 410. An ablation mask 420 is generated by the neural network.

[0147] It should be emphasized that, for each medical image in the second database, the ablation zone is advantageously manually determined by at least two operators in order to improve the relevance of the training and thus the relevance of the analysis results obtained by the second neural network. In particular, it may be difficult for operators to delineate the ablation zone, especially when the contrast of the image is poor, e.g. Figure 4 400 in the image. Therefore, using several operators to annotate the medical image, it is possible to improve the identification of ablation zones. Therefore, in a non-limiting example of the present invention, the ablation zone associated with a pair of readjusted pre-operative and post-operative medical images is equivalent to the union of the ablation zones suggested by the operators. As another example, the ablation zone associated with the medical image can correspond to the intersection, consensus, or ruling of the ablation zones suggested by the operators. In addition, the neural network is also trained to classify voxels of the medical image as ablation zones or non-ablation zones.

[0148] As an option, an expert annotator can be used to delineate ablation zones in medical images for training. The operator’s experience is important for the second neural network to be able to derive well-defined ablation zones.

[0149] It should also be emphasized that, in order to limit the training bias of the second neural network, the second database can have the same advantages as the first database, specifically, it can contain images without ablation zones, the locations of the ablation zones are not systematically located in the center of the images, and the set of images in the database contains as many voxels of ablation zones as non-ablation zones of the anatomical parts visible in the images.

[0150] From the post-operative medical images of the individual target anatomical structure 130, the second neural network classifies each voxel as an ablation zone or a non-ablation zone. This prediction can take the form of an ablation mask superimposed on a pair of medical images of the target anatomical structure 130. The ablation mask is typically rescaled to the voxels belonging to the ablation zone predicted by the second neural network. It should be emphasized that in the context of a two-dimensional image, the ablation mask is generally delineated by a region or a contour.

[0151] In sub-step 243, lesion 165 is automatically segmented on the pre-operative medical image of target anatomy 130 of individual 110. Alternatively, the segmentation is performed manually by an operator.

[0152] The lesions are automatically segmented using methods known to those skilled in the art, for example, using image histogram-based methods, such as the Otsu method, or using deep learning methods.

[0153] The segmentation sub-step 243 is Figure 5 , which shows a preoperative medical image 500, in which automatic segmentation based on a deep learning method is performed to determine the three-dimensional position of a lesion 510 and an ablation zone 520. Using a third neural network different from the two previously used neural networks, an equivalent result can be obtained.

[0154] Then, in sub-step 244, an ablation margin is determined between the previously established lesion segmentation and the ablation mask. The ablation margin corresponds to the smallest margin, i.e., the smallest distance indicated between the segmentation of the lesion and the ablation mask. In other words, the ablation margin is equivalent to the smallest distance calculated between a point of the lesion and a point of the ablation zone, and is calculated for all points of the lesion.

[0155] Determining the ablation margins can ensure that the ablation zone truly covers the lesion.

[0156] Figure 6 An example of performing additional ablations after ablation of a lesion 600 at risk of recurrence is shown. The post-treatment evaluation method 200 has determined areas where the ablation margin 605 between the lesion 600 and the ablation zone 610 is insufficient, and an additional ablation mask 620 has been generated. It should be emphasized that the additional ablation mask 620 is generated in an attempt to limit the ablation zone as much as possible. The target point 630 to be reached by the ablation needle can then be further defined in the mask 620.

[0157] In sub-step 245, the prediction of the recurrence location can be further evaluated by comparing the calculated ablation margin with an ablation margin reference value associated with the recurrence state and stored in a database. For example, when the ablation margin is greater than or equal to 5 mm, the risk of recurrence is considered to be low or even zero.

[0158] Additionally or alternatively, recurrence risk predictors other than ablation margins may be used. The location of the recurrence may be assessed by weighting all or some of these different predictors.

[0159] For example, predictors of recurrence risk can be:

[0160] - the distance between the lesion surface or part of the lesion surface and the ablation zone;

[0161] - The distance between the centroid of the lesion and the centroid of the ablation zone;

[0162] - the distance between the lesion surface or part of the lesion surface and the ablation zone, as well as the distance between the lesion centroid and the ablation zone centroid, taking into account the proximity of the capsule to the anatomical structures involved, in particular in the case of subcapsular lesions;

[0163] - Regularity of the edges of the ablation zone relative to the surrounding healthy tissue;

[0164] - The clarity of the edges of the ablation zone relative to the surrounding healthy tissue;

[0165] - Ratio of the volume of the lesion to the volume of the ablation zone;

[0166] -Location of the lesion within the target anatomy.

[0167] The reference value of the center of mass depends on the ablation margin. If the ablation margin is equal to 10 mm and the reference value of the ablation margin is 5 mm, the distance between the center of mass of the lesion and the center of mass of the ablation zone should be less than or equal to 5 mm.

[0168] The method 200 may further include a step 250 of planning a trajectory to be followed by the medical instrument 120 associated with the ablation mask or the additional ablation mask, so as to guide an operator to operate the medical instrument 120 along the planned trajectory in a step 260 of guiding the medical instrument 120 .

[0169] An example of a planning method is described in French patent application No. 1914780 entitled “Method for automatically planning a trajectory for a medical intervention”.

[0170] It should be emphasized that, in a non-limiting example of the present invention, the guidance is visual, involving displaying the planned trajectory and / or guidance instructions on the screen 155 of the helmet 150 .

[0171] As an option, the medical instrument 120 may be guided by a navigation system that provides position and orientation information for the medical instrument 120. This may involve mechanical guidance using a robotic device coupled to such a navigation system.

[0172] It should be emphasized that steps 230 to 260 may be repeated until the risk of recurrence is zero or almost zero, or until the ablation margin is sufficient.

Claims

1. A method (200) for post-treatment evaluation of ablation of a portion of a target anatomical structure (130) of an individual (110), the target anatomical structure comprising at least one lesion (165, 510, 600), the ablated portion of the target anatomical structure being referred to as an ablation zone (160), It is characterized in that The post-treatment evaluation method comprises the following steps: - acquiring (210) a post-operative medical image of the target anatomical structure of the individual, including all or part of the ablation zone; - realigning (220) the postoperative image and the medical image (170, 500) of the individual target anatomical structure obtained before the surgical treatment, referred to as the preoperative medical image, the realigned preoperative medical image and the realigned postoperative medical image constituting a pair of medical images of the individual target anatomical structure; - Assessing the risk of recurrence of a target anatomical structure lesion of an individual using a neural network-based machine learning method, analyzing a pair of medical images of the target anatomical structure of the individual, wherein the machine learning method has been previously trained in a so-called training phase (290) on a database of multiple pairs of medical images (300, 400) of the same target anatomical structure of a group of patients, wherein each pair of medical images in the database is associated with a recurrence status of the lesion of the target anatomical structure of the patient.

2. The post-treatment evaluation method of claim 1, wherein the risk of recurrence is evaluated at a predetermined date after treatment, and when the recurrence status is positive, each pair of medical images in the database is also associated with the recurrence date.

3. A post-treatment evaluation method as described in claim 1, wherein after realignment, all or some of the image pairs in the database are cropped around the ablation zones contained in the post-operative images of all or some of the training image pairs, and the images are cropped within a common frame of a predetermined size, and the center set of the ablation zones in the cropped image pairs forms a constellation of distinct points within the common frame.

4. A post-treatment evaluation method as described in claim 3, wherein for all pre-cropped medical image pairs in the database, the individual body parts contained in the image are divided into a plurality of basic units of a single size, and the number of basic units is divided into two almost equal parts between the human body part delineated by the ablation zone and the rest of the individual body parts contained in the image.

5. The post-treatment evaluation method of claim 1, wherein the database of paired medical images includes at least one pair of images without an ablation zone.

6. The post-treatment evaluation method of claim 1, further comprising the step of determining an additional ablation mask when the risk of recurrence is greater than a predetermined threshold.

7. A post-treatment evaluation method as claimed in claim 6, wherein the step of determining an additional ablation mask includes the sub-step of detecting the ablation zone in a post-operative image of the target anatomical structure of the individual.

8. The post-treatment evaluation method of claim 6, wherein the step of determining an additional ablation mask comprises the sub-step of segmenting the ablation zone in the post-operative image of the individual's target anatomical structure.

9. The post-treatment evaluation method of claim 6, wherein the step of determining an additional ablation mask includes the substep of detecting lesions in pre-operative images of the individual's target anatomical structure.

10. The post-treatment evaluation method of claim 6, wherein the step of determining an additional ablation mask includes a sub-step of segmenting lesions in a pre-operative image of the individual's target anatomical structure.

11. The post-treatment evaluation method of claim 6, wherein the step of determining an additional ablation mask comprises a sub-step of determining the location of the recurrence as a function of at least one risk prediction factor selected from the following list: - ablation margins between ablation zones and lesions, - the distance between the centroid of the lesion and the centroid of the ablation zone, - Regularity and definition of the edges of the ablation zone relative to the surrounding healthy tissue, - the ratio of the volume of the lesion to the volume of the ablation zone, -Location of the lesion relative to the center of the target anatomy.

12. The post-treatment evaluation method of claim 1, wherein the medical image is a three-dimensional image.

13. The post-treatment evaluation method of claim 1, wherein each pre-operative image is obtained using a first image acquisition technique, and each post-operative image is obtained using a second image acquisition technique, and the first image acquisition technique and the second image acquisition technique are the same or different.

14. An electronic device (150) comprising a processor and a computer memory storing instructions for the post-treatment evaluation method of any one of the preceding claims.

15. The electronic device as claimed in claim 14, which can be a control device, a navigation system, a robotic device or an augmented reality device.