System and method for semi-automatic image segmentation by user interaction loop learning and training method thereof

Abstract

The invention relates to a semi-automatic image segmentation system comprising a neural network (12) configured, for each iteration, to receive input data comprising at least one image (14) to be segmented and to provide output data comprising at least one mask (16) for segmenting each image, characterised in that it further comprises a user interaction loop (18) comprising a user interaction module (20) configured to provide at least one user mask representative of an interaction with a real user and / or a simulated user on each segmentation mask, the user interaction loop (18) being configured to provide, as input data for the neural network (12), each segmentation mask provided by the neural network from a past iteration of the neural network (12), referred to as the past segmentation mask, and the user mask associated with each past segmentation mask.

Description

[0001] DESCRIPTION

[0002] TITLE OF THE INVENTION: SYSTEM AND METHOD FOR SEMI-AUTOMATIC IMAGE SEGMENTATION BY USER INTERACTION LOOP LEARNING AND ASSOCIATED TRAINING METHOD

[0003] Technical field of the invention

[0004] The invention relates to a 2D or 3D image segmentation system comprising a neural network, an image segmentation method using this system and a method for training the segmentation system, in particular its neural network. The invention is intended in particular for the segmentation of medical images, for example originating from magnetic resonance imaging (MRI) or computed tomography (CT) for the identification of organs on each of the section images obtained by FIRM or CT in order to recreate a 3D model of the organ from these segmentations.

[0005] Technological background

[0006] Image segmentation is a technique for associating each pixel or voxel of a 2D or 3D image with a label representing the object, texture or state, etc. corresponding to said pixel or voxel.

[0007] Image segmentation is used in particular in a medical context to determine the organs, tissues, fluids, etc. visible on a 2D or 3D image. Such an image is obtained, for example, by MRI and makes it possible to form a 3D model of the targeted organ and, for example, of its environment and the tumor to be operated on. This creation of a 3D model requires the segmentation of the images obtained by FIRM.

[0008] Current segmentation techniques fall into several categories: manual, semi-automatic (also called interactive) or fully automatic segmentation.

[0009] The manual approach is time-consuming to implement, as each pixel or voxel must be categorized one by one by a user and can thus require several hours for a single slice. The automatic approach allows segmentation without user intervention, but the lack of control can negatively affect the result and therefore the practical application. In addition, the result is never perfect, and to obtain a better result, the human operator must be able to maintain control for validation reasons, and therefore be able to correct errors.

[0010] The semi-automatic method allows interactivity in which the segmentation is both mostly automatic but with a degree of control by a user, who can notably act on the initialization and on a modification during certain stages.

[0011] The various segmentation processes can implement methods based on deep learning, particularly using artificial neural networks. Deep learning is particularly used for fully automatic processes, but semi-automatic segmentation solutions implementing deep learning have received limited attention compared to the number of automatic solutions. Neural networks have greatly improved the performance of automatic segmentation solutions, but semi-automatic segmentations have not benefited from as many improvements due to difficulties in successfully integrating neural networks.

[0012] In particular, the use of neural networks in semi-automatic segmentation poses problems in taking into account user interactions during the training of the neural network. This consideration of user interactions creates problems in training the system, because the training data that takes into account user interactions must adapt to the output of the neural network to which the interactions are applied, the neural network not being trained at the beginning of the process. This leads to a problematic dependency, the input data necessary for training being themselves dependent on the output data to which the user interactions have been applied.

[0013] The proposed solutions to address these issues do not provide the desired performance, particularly because user interactions are generally not taken into account dynamically and have little impact on the training quality, as they focus on the input data. The proposed solutions generally approach training a neural network for semi-automatic segmentation tasks in the same way as for automatic segmentation. They generate user interaction masks from the annotated data, either statically before training or dynamically during training. Static generation methods form input structures to the neural network closest to automatic segmentation, and do not take into account the dynamic nature of user interactions.Existing dynamic generation methods allow for diversifying training data and increasing performance but do not directly address the problematic dependency between outputs and inputs of the system taking into account interactions. In particular, user interactions used in dynamic processing do not reproduce a typical sequence of user interaction because they do not take into account the order of these interactions.

[0014] The inventors thus sought to improve semi-automatic segmentation solutions to enable better consideration of user interactions, in particular to take into account during training the impact of user interactions during the deployment of the solution.

[0015] Objectives of the invention

[0016] The invention aims to provide a system and a method for semi-automatic segmentation of images, in particular for medical type images, for example sections obtained by MRI.

[0017] The invention also aims to provide a method for training such a segmentation system.

[0018] The invention aims to provide, in at least one embodiment, a semi-automatic segmentation system that can be easily trained by simulating user interactions.

[0019] The invention aims to provide, in at least one embodiment, a segmentation system whose operation and training allow previous segmentation results to be taken into account.

[0020] The invention aims to provide, in at least one embodiment, a segmentation system whose operation and training allow previous user interactions to be taken into account.

[0021] Statement of the invention

[0022] To this end, the invention relates to a semi-automatic image segmentation system comprising a neural network configured to, at each iteration, receive input data comprising at least one image to be segmented and to provide output data comprising at least one segmentation mask of each image, characterized in that it further comprises a user interaction loop comprising a user interaction module configured to provide at least one user mask representative of an interaction of a real user and / or a simulated user on each segmentation mask, said user interaction loop being configured to provide, as input data of the neural network: each segmentation mask provided by said neural network coming from a previous iteration of the neural network, called past segmentation mask, and the user mask associated with each past segmentation mask.

[0023] An image segmentation system according to the invention therefore makes it possible to take into account user interactions both during training and during use for image segmentation, and to associate these user interactions with a segmentation mask obtained at the output of the neural network as a new input to the neural network.

[0024] In particular, the training data are not obtained solely by retrieving a large dataset but are generated by the segmentation system itself through the user interaction loop. The user mask adapts according to the network output data and therefore cannot be prepared in advance. Without this loop, as in automatic segmentation systems, the system would only be trained from the images to be segmented and therefore could not, in use, take into account user interactions. Using only an input dataset without the loop would therefore not allow for a satisfactory result, regardless of the volume of this input dataset.In particular, using only an input dataset without the loop would allow obtaining a system that does not allow modification of the obtained result, such modification being necessary in many fields, in particular in the medical field where a user must validate or edit the result before its use in a therapeutic act.

[0025] Furthermore, using the output segmentation mask as input allows for sequentiality in data processing by taking past segmentation masks into account when defining better future segmentation. Using the output segmentation mask and user interaction as input data allows for a sequential internal memory to be built that keeps track of past segmentations and user interactions to maintain the sequentiality of the system. In particular, the sequential internal memory keeps all previous user interactions and segmentation masks obtained as output from the neural network, to provide them as new input to the neural network. Segmentation is also accelerated because sequential user interactions have more impact on the result.

[0026] The user interaction loop is used both when training the neural network, when testing the segmentation system from test data, and when using the segmentation system for new data.

[0027] Real user interaction is defined as an interaction by a human user who interacts directly with the segmentation mask proposed as output. Simulated user interaction is defined as the simulation of an interaction as a human user might do, defined by a user interaction simulation algorithm. User interaction is, for example, the assignment of a particular label to a specific point in the segmented image. User interaction makes it possible, in particular, to correct the segmentation mask obtained at the output of the neural network, for example by indicating areas of the 2D or 3D image that have not been correctly labeled. Taking the user into account from the system training stage improves the impact of user corrections when using the system.Advantageously, and thanks to its training method, the image segmentation system can perform the initial segmentation itself when trained and take into account user corrections for improved segmentation. The initial segmentation can be trivial, in particular little or no informative, or calculated by another segmentation method, typically a fully automatic method.

[0028] The system can advantageously be used for the segmentation of MRI images but can be applied to other types of 2D or 3D images, medical or not. The system can be integrated into a larger system using the segmentation results for a particular application, for example for creating a 3D model from the segmented data and displaying the 3D model in augmented reality on a video.

[0029] The neural network can be of different types, the segmentation system not depending on a particular neural network for its use. Indeed, the image segmentation system is particularly characterized by the user interaction loop which is external to the neural network. The neural network can integrate modules of the recurrent neural network type or RNN for recurrent neural network in English.

[0030] A module may for example consist of a computing device such as a computer, a set of computing devices, an electronic component or a set of electronic components, or for example a computer program, a set of computer programs, a library of a computer program or a function of a computer program executed by a computing device such as a computer, a set of computing devices, an electronic component or a set of electronic components. Advantageously and according to the invention, each user mask representative of a simulated user comprises a simulation of at least one interaction carried out by the simulated operator on the segmentation mask associated with the user mask, the position of said interaction on the image being determined as a function of the difference between said output segmentation mask and a ground truth.

[0031] According to this aspect of the invention, the position of the simulated interaction is determined to approximate what a real user would do, i.e., position their interaction in an area that has been incorrectly segmented or labeled, or in an area that does not have a label.

[0032] Advantageously and according to the invention, the position of the interaction is determined as a function of the center of a region chosen as being to be corrected by the user interaction.

[0033] Advantageously and according to the invention, the system comprises a graph cutting module configured to apply a graph cut to the output data of the neural network to obtain each segmentation mask.

[0034] According to this aspect of the invention, this step allows the smoothing of the output data by using a graph cut algorithm (graph cut in English) to improve the relevance of the labels integrated into the segmentation mask. The graph cut algorithm is applicable for binary segmentation and derived algorithms are used for multi-class segmentations.

[0035] Advantageously and according to the invention, the user interaction loop comprises a sequential internal memory in which each segmentation mask passed and each user mask obtained at each iteration is sequentially stored.

[0036] Advantageously and according to this last aspect of the invention, the internal sequential memory comprises a “first in-first out” or FIFO (for First In-First Out in English) type queue, the queue storing at each iteration the segmentation mask and the user mask corresponding to the iteration.

[0037] The invention also relates to a method for training an image segmentation system according to the invention, via at least one data set comprising images to be segmented, characterized in that it comprises, for each image of the data set: at least two iterations in which the neural network processes the input data to provide output data, a first iteration in which the input data comprise at least one image to be segmented and a predefined initialization mask, and subsequent iterations in which the input data comprise each image to be segmented, each past segmentation mask obtained in the previous iteration, and a user mask representative of an interaction of a real user and / or a simulated user, a step of updating the weights of the neural network from the output data after each iteration or when a predetermined number of iterations have been performed by the neural network.

[0038] A training method according to the invention makes it possible to simulate user interactions to maximize the effectiveness of the training and the relevance of the neural network at the end of the training. Unlike conventional models of the prior art in which an update of the weights is carried out after each iteration, the training method according to the invention advantageously uses the user interaction loop of the system to allow the use of a user interaction simulation and output data, i.e. the segmentation mask of a previous iteration, as input data of the system. Thus, the image segmentation system makes it possible to improve the user interaction simulation based on quality checks of the segmentation obtained, and to take into account past segmentation masks and user interactions.Training thus allows the reduction of the number of false positives and false negatives by simulating what a user would do when checking the quality of the segmentation. The training sequence formed by the iteration sequence makes it possible to improve the accuracy of the corrections and thus increase the quality of the segmentation at each iteration.

[0039] The weight update is for example implemented by a loss function minimization algorithm, whose gradient is for example calculated by backpropagation. The weight update can be implemented at each iteration, in particular each time output data is produced by the neural network, or only at the end of all iterations so that the weight update is only carried out from the best segmentation obtained.

[0040] The initialization mask is for example a trivial mask, an empty mask, a mask obtained by automatic pre-segmentation or a mask derived from the ground truth, for example by simulating user interactions from this ground truth.

[0041] In subsequent iterations, the initialization mask is added to the user mask which indicates the errors between the segmentation mask obtained at the output of the neural network and the result desired by the user.

[0042] Each dataset can include one or more images, different or not. Even if two datasets include the same images, the training process will treat two identical images differently thanks to simulated user interactions that allow for changes in training.

[0043] Simulated user interactions are obtained from ground truth. Ground truth is formed by predefined labeled data from which the error between the segmentation mask obtained at the output of the neural network and the desired result is determined. Training is thus supervised using previously segmented data, for which the image regions have labels assigned to them. Ground truth is more commonly called "ground truth" in English.

[0044] In the segmentation methods of the prior art, training is generally carried out only from these labeled data which are not sufficient since the user and the segmentation masks of the previous iteration are not taken into account in the training. Furthermore, the training method according to the invention makes it possible to provide different simulated user interactions for the same image, as would be the case with a real user who will never provide the same interactions for the same image.

[0045] Advantageously and according to the invention, the predetermined number of iterations is determined for each image to be segmented and for each data set independently of the predetermined number of iterations determined for the other images to be segmented and the other data sets.

[0046] According to this aspect of the invention, the number of iterations is predetermined when processing a dataset or an image to be segmented but can be modified over time in order to reproduce the behavior of a real user who will not systematically make the same number of corrections. A maximum number of iterations can be set to avoid training being too long. Setting the maximum number of iterations also makes it possible to avoid making the network dependent on interactions during the test phase. Since the interactions are based on the ground truth, it is preferable to target that a minimum of interactions are provided during the training phase while maximizing performance, so that a minimum of interactions are necessary during the test phase thereafter.

[0047] The invention also relates to a semi-automatic image segmentation method comprising at each iteration a step of processing input data comprising each image to be segmented by a neural network to provide output data comprising at least one segmentation mask of each image, characterized in that each iteration further comprises: a step of receiving a user mask representative of an interaction of a real user on each segmentation mask, a step of transmitting, as input data of the neural network to the following iteration, each segmentation mask provided by said neural network, and the user mask. According to this aspect of the invention, the image segmentation method allows the user's interactions to be taken into account as many times as necessary until the user considers that the quality of the segmentation of the image is in line with his expectations.If the quality is not sufficient, the user interactions and the segmentation mask are passed as input data to provide a new segmentation mask.

[0048] Advantageously, the image segmentation system according to the invention is configured to implement the image segmentation method according to the invention.

[0049] Advantageously, the image segmentation method according to the invention is implemented by an image segmentation system according to the invention.

[0050] The invention also relates to a computer program product for training an image segmentation system according to the invention, said computer program product comprising program code instructions for executing, when said computer program product is executed on a computer, the steps of the training method according to the invention.

[0051] The invention also relates to a semi-automatic image segmentation computer program product, said computer program product comprising program code instructions for executing, when said computer program product is executed on a computer, the steps of the image segmentation method according to the invention.

[0052] The invention also relates to an automatic segmentation system, an automatic segmentation method, an automatic segmentation computer program product, a training method and a training computer program product characterized in combination by all or some of the features mentioned above or below.

[0053] List of figures

[0054] Other aims, characteristics and advantages of the invention will appear on reading the following description given solely for non-limiting purposes and which refers to the appended figures in which:

[0055] [Fig. 1] is a schematic view of a semi-automatic image segmentation system according to one embodiment of the invention.

[0056] [Fig. 2] is a schematic view of a semi-automatic image segmentation method according to one embodiment of the invention.

[0057] [Fig. 3] is a schematic representation of the steps of a method for training a semi-automatic image segmentation system according to one embodiment of the invention.

[0058] Detailed description of an embodiment of the invention

[0059] In the figures, scales and proportions are not strictly respected, for the purposes of illustration and clarity.

[0060] Furthermore, identical, similar or analogous elements are designated by the same references in all figures.

[0061] Figure 1 schematically represents a semi-automatic image segmentation system 10 according to one embodiment of the invention.

[0062] The segmentation system 10 comprises a neural network 12 configured to, at each iteration, receive input data comprising at least one image 14 to be segmented and to provide output data comprising at least one segmentation mask 16 of each image. The segmentation mask defines for each part of the image a label making it possible to identify regions of the image, corresponding to a particular class. The segmentation can be binary but the segmentation can also be a multi-class segmentation and makes it possible to identify more than two classes of regions. At the output of the neural network, a graph cut can make it possible to smooth the output data to obtain the segmentation mask.

[0063] The segmentation system 10 integrates around an existing machine learning model. For example, the system uses an existing encoder-decoder neural network architecture comprising recurrent neural network (RNN) modules. One embodiment, for example, uses a ResNet34 encoder described in the publication “He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770-778.” and a decoder equipped with a pair of standard convolutional layers and a corresponding long short-term memory (LSTM) recurrent layer at each step of the upsampling path.

[0064] To counter the dataset imbalance, the system uses focal loss as described in "Lin, TY, Goyal, P., Girshick, RB, He, K., Dollàr, P., 2017. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), 2999-3007." and pre-computed dataset-wide class weights.

[0065] For example, the neural network is pre-trained on the ImageNet dataset and then refined on a dataset dedicated to the desired task. A dataset can, for example, be a dataset comprising medical images such as MRIs of the female pelvis, manually segmented using manual segmentation software such as 3DSlicer and MITK.

[0066] A female pelvic MRI dataset created by the applicant comprises ninety-seven MRI series representing three thousand sixty-six slices in total. This female pelvic MRI dataset is used for training the neural network, at a rate of: seventy-seven MRI series representing two thousand four hundred forty-nine slices for the training set, ten MRI series representing three hundred eight slices for the validation set, and ten MRI series representing three hundred nine slices for the test set.

[0067] The slices of the female pelvic MRI dataset can be preprocessed via normalization, standardization and bias correction, and random data augmentation can be provided, e.g., vertical and horizontal flipping, intensity shift for brightness, gamma correction for contrast, as well as blur and unsharp mask for sharpness adjustment, etc.

[0068] In the context of MRI-type medical images for gynecological application, examples of labels can be the uterus, the bladder, a possible tumor, a cavity, the background, etc., forming segmentation classes.

[0069] The image segmentation system 10 differs from the prior art in that it further comprises a user interaction loop 18 comprising a user interaction module 20 configured to provide at least one user mask representative of an interaction of a real user and / or a simulated user on each segmentation mask.

[0070] The user interaction loop 18 is configured to provide, as input data to the neural network: each segmentation mask 16 provided by said neural network originating from a previous iteration of the neural network 12, called past segmentation mask, as represented by the arrow 22, the user mask associated with each past segmentation mask 16 obtained by the user interaction module 20, as represented by the arrow 24.

[0071] The segmentation mask 16 and the user mask are combined via a state module 26 to form a state 28 formed by the combination of the segmentation mask and the user mask to form new input data. During the first iteration, the state 28 is fed by an equivalent empty structure 30, a mask obtained by a pre-segmentation obtained automatically or a mask derived from the ground truth when the system is trained. The state 28 thus combines all the past segmentation masks and user masks, thus forming an internal sequential memory for the segmentation system.

[0072] User interaction loop 18 is external to the neural network and does not depend on a specific neural network architecture, although examples of neural networks that can be used are described previously.

[0073] A conventional prior art memoryless interaction loop takes into account user interactions by providing the neural network as input the image to be segmented and the user interactions. The user interactions are represented using N binary masks, with N the number of segmentation classes. The neural network input is thus concatenated into a tensor of size HxWxC, with H the image height, W the image width and Cbase the number of channels, which is equal to 1+N in a memoryless interaction loop, corresponding to one channel for the image and N channels for the N segmentation classes.

[0074] Prior art solutions have proposed adding an internal memory, characterized by a set of N binary masks for each segmentation class in which user interactions are accumulated. The input tensor therefore comprises COM = 1 + N channels since all user interactions are merged. This type of method can be called "cumulative interaction memory" in view of its operating principle and is implemented according to different variants, notably in the following publications:

[0075] Amrehn, M., Gaube, S., Unberath, M., Schebesch, F., Horz, T., Strumia, M., Steidl, S., Kowarschik, M., Maier, A., 2017. ULNet: Interactive Artificial Neural Networks for Iterative Image Segmentation Based on a User Model, in: Eurographics Workshop on Visual Computing for Biology and Medicine, The Eurographics Association;

[0076] Zhou, B., Chen, L., Wang, Z., 2019. Interactive deep editing framework for medical image segmentation, in: Shen, D., Liu, T., Peters, T.M., Staib, L.H., Essert, C., Zhou, S., Yap, P.T., Khan, A. (Eds.), Medical Image Computing and Computer Assisted Intervention - MICCAI 2019, Springer International Publishing, Cham. pp. 329-337 ;

[0077] Liao, X., Li, W., Xu, Q., Wang, X., Jin, B., Zhang, X., Zhang, Y., Wang, Y., 2020. Iteratively -refined interactive 3d medical image segmentation with multi-agent reinforcement learning. 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9391-9399.

[0078] The image segmentation system 10 according to the invention therefore comprises, thanks to the preservation of the state of the previous interactions, a sequentiality stored in an input tensor comprising CSIM = 1 + 2DN channels, with D the number of stored states, that is to say the number of interactions implemented. The factor 2 corresponds to the presence of N interaction masks and N probability maps corresponding to the intermediate segmentation results.

[0079] The internal sequential memory is thus formed by a FIFO type stack where the successive states corresponding to the interactions are recorded. This memory is external to the neural network and has no interaction with the internal memory of the neural network, in particular with the weights assigned to each node of the neural network which are specific to this neural network and to the type of data.

[0080] Figure 2 schematically represents a method 40 for semi-automatic image segmentation according to one embodiment of the invention.

[0081] The segmentation method comprises a step 42 of receiving the images to be segmented to provide them to the neural network as described in the segmentation system.

[0082] The method comprises at each iteration a step 44 of processing input data comprising each image to be segmented by the neural network to provide output data comprising at least one segmentation mask of each image.

[0083] If the quality of the segmentation is considered correct at a quality control step 46, the segmentation process is completed and a final step 52 provides the segmentation mask associated with the last segmentation.

[0084] If the quality of the segmentation is not correct, each following iteration further comprises a step 48 of receiving a user mask representative of an interaction of a real user on each segmentation mask, and a step 50 of transmitting, as input data of the neural network to the following iteration, each segmentation mask provided by said neural network, and the user mask.

[0085] Figure 3 schematically represents steps of a method for training a semi-automatic image segmentation system according to an embodiment of the invention,

[0086] The training method 100 comprises for each image of a data set, a predetermined number of iterations, each iteration being noted from Ii to Each iteration comprises a step 112 of processing input data 114 by the neural network to provide output data 116. In the first iteration Ii, the input data comprises at least one image to be segmented and a predefined initialization mask 110. The predefined initialization mask 110 is here for example an empty mask, a trivial mask, a mask obtained by automatic pre-segmentation or a mask derived from the ground truth.

[0087] In the following iterations, from L to L, the input data comprises each image to be segmented, each past segmentation mask obtained in the previous iteration, and a user mask representative of an interaction of a real user and / or a simulated user, as symbolized by the dotted arrows which link the output data 116 of one iteration to the input data 114 of the following iteration.

[0088] The user mask is for example obtained by comparing the segmentation mask with the ground truth, i.e. previously labeled images, allowing supervised learning. The user mask thus simulates the interaction of a user who knows the desired result of the segmentation.

[0089] The user mask is for example a binary mask per segmentation class, symbolizing a click of the user in the foreground. The position of the click is chosen randomly in the largest region of the class, following a probability map in which the maximum is in the center of the region, decreasing away from the center and very reduced outside the region.

[0090] Using the internal sequential memory as described above, training therefore allows filling the memory of D states with realistic values ​​produced by a virtual user. The D states are trained by D iterations to fill the internal sequential memory before updating the weights. The states are preferably trained by inference via the neural network, i.e. without updating the neural network weights.

[0091] The method also comprises, after each iteration or when the predetermined number of iterations have been performed by the neural network, a step 120 of updating the weights of the neural network from the output data. In this case, in the embodiment shown in FIG. 3, the step 120 of updating the weights of the neural network is carried out at the last iteration, on the basis of the last output data 116.

[0092] The predetermined number of interactions is determined independently for each class and for each image set. The maximum number of interactions is typically three interactions to simulate a typical user, the minimum being zero.

[0093] The probability of adding an additional interaction decreases linearly after each interaction to reach zero when the maximum number of interactions is reached.

[0094] In one embodiment of the invention, a class is randomly chosen at each epoch of the training process with a dataset, for which no user interaction is generated.

[0095] In one embodiment of the invention, a percentage of the generated interactions is retained, typically 20%.

Claims

CLAIMS Semi-automatic image segmentation system comprising a neural network (12) configured to, at each iteration, receive input data comprising at least one image (14) to be segmented and to provide output data comprising at least one segmentation mask (16) of each image, characterized in that it further comprises a user interaction loop (18) comprising a user interaction module (20) configured to provide at least one user mask representative of an interaction of a real user and / or a simulated user on each segmentation mask, said user interaction loop (18) being configured to provide, as input data to the neural network: each segmentation mask provided by said neural network (12) from a previous iteration of the neural network (12), said past segmentation mask, and the user mask associated with each past segmentation mask.A segmentation system according to claim 1, characterized in that each user mask representing a simulated user comprises a simulation of at least one interaction performed by the simulated operator on the segmentation mask (16) associated with the user mask, the position of said interaction on the image being determined based on the difference between said output segmentation mask (16) and a ground truth value. A segmentation system according to claim 1 or 2, characterized in that it comprises a graph cutting module configured to apply a graph cut to the output data of the neural network (12) to obtain each segmentation mask (16). A segmentation system according to any one of claims 1 to 3, characterized in that... that the user interaction loop (18) includes a sequential internal memory in which each passed segmentation mask and each user mask obtained at each iteration are sequentially stored. Method for training an image segmentation system according to any one of claims 1 to 4 via at least one dataset comprising images to be segmented, characterized in that it comprises, for each image in the dataset: at least two iterations (Ii, L, L) in which the neural network (12) processes the input data to provide output data, a first iteration (Ii) in which the input data includes at least one image to be segmented and a predefined initialization mask (110), and subsequent iterations in which the input data includes each image to be segmented, each passed segmentation mask obtained at the previous iteration,and a user mask representative of an interaction of a real user and / or a simulated user, a step (120) of updating the neural network weights from the output data after each iteration or when a predetermined number of iterations have been performed by the neural network. Training method according to claim 5, characterized in that the predetermined number of iterations is determined for each image to be segmented and for each dataset independently of the predetermined number of iterations determined for the other images to be segmented and the other datasets. Semi-automatic image segmentation method comprising, at each iteration, a step of processing input data comprising each image to be segmented by a neural network to provide output data comprising at least one segmentation mask of each image, characterized in that each iteration further comprises: a step of receiving a user mask representative of an interaction of a real user on each segmentation mask; a step of transmitting, as input data from the neural network to the next iteration, each segmentation mask provided by said neural network, and the user mask. Product: computer program for training an image segmentation system according to any one of claims 1 to 4, said product: computer program comprising program code instructions for executing, when said product: computer program is run on a computer, the steps of the training method according to any one of claims 5 or 6.semi-automatic image segmentation computer program product said computer program product comprising program code instructions for the execution, when said computer program product is run on a computer, of the steps of the image segmentation process according to claim 7.

Images