Method for characterizing a patient's organs in a medical image - Patent Application 20070122997

Abstract

The present invention relates to a method implemented by computer means for characterizing at least one organ of a patient in at least one medical image, said method comprising the following iterative steps: segmenting (6) said organ in at least one image (7) representing said healthy organ to obtain at least one segmentation mask (8) of said healthy organ; generating additional segmentation masks of said healthy organ using at least one data augmentation method (9) based on the segmentation mask obtained in step (a), said data augmentation method (9) being a shape-preserving method; and training an autoencoder (10) using the segmentation mask obtained in step (a) and the additional segmentation masks obtained in step (b) to encode the shape of said organ in a latent space (11).

Description

[Technical Field]

[0001] The present invention relates to a computer implemented method for characterizing at least one organ of a patient in at least one medical image. [Background technology]

[0002] Anatomical changes in organs such as the brain or pancreas can provide information about functional disorders. For example, hippocampal atrophy and pancreatic duct dilation are well-known markers of Alzheimer's disease and pancreatic ductal adenocarcinoma. Therefore, in these cases, quantifying anatomical differences holds great potential for determining a patient's clinical condition, predicting their future progression or regression, and assisting in treatment planning.

[0003] The computational anatomical literature has proposed several statistical shape modeling (SSM) methods that embed geometric shapes into a metric space where concepts of distance and difference can be defined and quantified. Utilizing these representations, statistical shape models have subsequently been proposed to perform group analyses of shape collections.

[0004] In particular, atlas models learn geometric distributions in terms of "mean" representational shapes and associated variability, generalizing Euclidean mean-variance analysis. In medical imaging, learning atlases from healthy examples allows for the definition of normative models of anatomical structures or organs, such as subcortical regions segmented from brain MRI or neuroimaging data, and thus provides a natural framework for detecting abnormal anatomical structures.

[0005] In practice, leveraging the atlas model to compute the likelihood of a given shape belonging to the underlying distribution requires identifying landmarks or solving the registration problem, which implies a computational cost for the shape embedding operation.

[0006] Alternative approaches have been proposed to drop topological assumptions by relying on autoencoder variations or variational autoencoders to learn normative models that are subsequently used to perform anomaly detection (AD). These methods compress and reconstruct images of healthy subjects to capture normative models of organs. These methods are typically applied to raw imaging data and therefore run the risk of extracting features related to the intensity distribution of the dataset that are not necessarily specific to the organ's anatomy.

[0007] This document makes it possible to address these shortcomings in whole or in part. Summary of the Invention [Means for solving the problem]

[0008] To that end, the present document describes a method implemented by computer means for characterizing at least one organ of a patient in at least one medical image, said method comprising the following iterative steps: (a) segmenting a healthy organ in at least one image representing said organ to obtain at least one segmentation mask of said healthy organ; (b) generating an additional segmentation mask of the healthy organ using at least one data augmentation method based on the segmentation mask obtained in step (a), wherein the data augmentation method is a shape-preserving method; (c) training an autoencoder using the segmentation mask obtained in step (a) and the additional segmentation mask obtained in step (b) to encode the shape of the organ in a latent space; We propose a method including:

[0009] In digital image processing and computer vision, image segmentation is a method that aims to divide a digital image into at least one segment or set of pixels. More precisely, image segmentation is the process of assigning a label to every pixel or voxel in an image, such that pixels with the same label share certain characteristics.

[0010] In the context of this document, the result of step (a) is a set of pixels or voxels, also called a segmentation mask, that belong to said organ.

[0011] Furthermore, data augmentation is a set of techniques used to increase the amount of data by adding slightly modified copies of existing data or newly created synthetic data from existing data.

[0012] Furthermore, an autoencoder is a type of artificial neural network used to learn efficient coding of unlabeled data (unsupervised learning). An autoencoder consists of two parts: an encoder and a decoder. The encoder compresses the data from a high-dimensional space into a low-dimensional space (also called the latent space), while the decoder does the reverse, transforming the latent space back into a higher-dimensional space. The decoder is used to ensure that the latent space is able to capture most of the information from the dataset space by forcing the encoder to output what was fed as input.

[0013] The method includes the following further iterative steps: (d) segmenting the images representing the healthy organ and the images representing the pathological organ to obtain segmentation masks of the healthy organ and the pathological organ, respectively; (e) encoding the segmentation masks using the autoencoder trained in step (c) to obtain representations of healthy and pathological organs in a latent space; may include:

[0014] The method comprises the following further steps: (f) training a classifier capable of classifying the representations of healthy and pathological organs in the latent space by using at least some of the representations obtained in step (e) as a training set for the classifier.

[0015] In the context of this document, classification is the problem of identifying to which of a set of categories or classes an observation (here a coded representation of an organ) belongs, said classes may include a first class or healthy class and a second class or pathological class.

[0016] In such applications, after such training, the method comprises the following successive steps: - segmenting an image representative of an organ to be characterized to obtain a segmentation mask of said organ to be characterized; - encoding the segmentation mask using the autoencoder trained in step (c) to obtain a representation of the organ for characterization in a latent space; - classifying said expression with the classifier trained in step (f) to obtain a class of said expression; outputting said class; The method may include an inference phase including:

[0017] The application may be referred to as a few-shot learning method.

[0018] In another application, the method comprises the further step of: (f1) segmenting an image representing an organ to be characterized to obtain a segmentation mask of said organ to be characterized; (f2) encoding the segmentation mask using the autoencoder trained in step (c) to obtain a representation of the organ for characterization in a latent space; (f3) calculating the distance in the latent space between the representation of the organ to be characterized and the points representing healthy organs; may include:

[0019] Such distance provides information regarding whether the pancreas is healthy or pathological.

[0020] Said application may be called a zero-shot learning method.

[0021] The point can be the mean of the representation of the healthy organ in the latent space.

[0022] The distance can be an L2 or Euclidean distance.

[0023] In another application, the method comprises the further step of: (g1) selecting at least one representation in the latent space; (g2) using the autoencoder trained in step (c) to generate an organ segmentation mask by decoding the selected representation; (g3) outputting the generated segmentation mask; may include:

[0024] In such applications, the method may be used to reconstruct or generate segmentation masks of organs that are healthy, diseased, or "in between." Such reconstructions can be used by physicians to visualize organs in different states.

[0025] The selected representation may be a representation selected between the average representation of a healthy organ and the average representation of a pathological organ.

[0026] In particular, consider a set of healthy and pathological objects or organs and their segmentation masks, each represented in a latent space as z healthy and

number

[0027] an expression with t∈R (1-t).z healthy +tz pathological Consider the line t. Moving along this line for increasing values ​​of t, we progress from a healthy mean latent shape representation to a pathological representation, and can reconstruct the corresponding segmentation mask using the decoder part of a trained autoencoder.

[0028] The image may be a 3D image.

[0029] The image may be a CT scan image, for example a portal CT scan image.

[0030] Images may also be provided by other radiological modalities such as MRI, PET / CT or PET / MR.

[0031] The organ may be a pancreas.

[0032] The autoencoder may be a variational autoencoder.

[0033] A variational autoencoder may be based on UNet, where skip connections are removed to conform to the variational autoencoder framework.

[0034] nnUNet (a known framework that re-implements UNet) may be used to determine the topology of the UNet (number of layers, number of filters, filter size, etc...).

[0035] More specifically, the nnUNet framework automatically designs pipelines based on the UNet architecture by relying on heuristics applied to the data that allow for estimating important parameters. Dataset characteristics are estimated to automatically perform preprocessing steps. This is followed by the automatic definition of model design choices (number of layers, convolution kernel size, convolution blocks, etc.). Training procedures (planned learning rate, etc.) are also implemented.

[0036] An implementation of the nnUnet framework can be found at the following link as of the date of writing of this application: https: / / github.com / MIC-DKFZ / nnUNet.

[0037] The classifier may be a support vector machine classifier.

[0038] The shape-preserving data augmentation method may perform at least one random translation, rotation, and / or scaling of the segmentation mask.

[0039] This document also proposes computer software comprising instructions for carrying out at least part of the method according to any one of the preceding claims when the software is executed by a processor.

[0040] This document also: an input interface for receiving medical images; a memory for storing at least the instructions of a computer program according to the preceding claims; a processor that reads the instructions and then accesses the memory to perform the method according to the present document; The present invention proposes a computer device comprising:

[0041] The computing device may also include an output interface for providing output of the method.

[0042] This document also proposes a computer-readable non-transitory recording medium on which computer software is registered so as to perform the methods according to this document when the computer software is executed by a processor.

[0043] Other features, details and advantages are set forth in the following detailed description and drawings. [Brief explanation of the drawings]

[0044] [Figure 1] FIG. 1 is a diagram illustrating an example of a computer device according to the present writing. [Figure 2] FIG. 1 illustrates a pipeline for the application of the method according to the present document. [Figure 3] FIG. 10 illustrates a pipeline for another application of the method according to the present document. [Figure 4] FIG. 1 illustrates the architecture of an autoencoder model. [Figure 5] 1A and 1B are diagrams showing examples of normal and abnormal pancreatic shapes. [Figure 6] FIG. 1 shows a 2D reduction of a subject's latent representation. DETAILED DESCRIPTION OF THE INVENTION

[0045] The accompanying drawings contain significant color. This application is intended to be published in black and white, but color versions of the accompanying drawings have been submitted to the European Patent Office.

[0046] Figure 1 shows a schematic diagram of an example of a computer device 1 according to the invention. an input interface 2; a memory 3 for storing at least the instructions of a computer program; a processor 4 accessing a memory 3 to read said instructions and to perform the method according to the present document; -Output interface 5 and Equipped with.

[0047] General Pipeline The present document is illustrated in FIG. 2 and describes a method implemented by a computer means for characterizing at least one organ (e.g., the pancreas) of a patient in at least one medical image, said method comprising the following iterative steps: (a) segmenting 6 a healthy organ in at least one image 7 representing said organ to obtain at least one segmentation mask 8 of said healthy organ; (b) generating an additional segmentation mask of the healthy organ using at least one data augmentation method 9 based on the segmentation mask obtained in step (a), wherein the data augmentation method is a shape-preserving method; (c) training an autoencoder 10 using the segmentation mask obtained in step (a) and the additional segmentation mask obtained in step (b) to encode the shape of the organ in a latent space 11; We propose a method including:

[0048] The autoencoder 10 comprises an encoder 12 and a decoder 13 .

[0049] Such a pipeline is described in more detail below.

[0050] method Organ shape modeling Consider an image acquired via standard imaging techniques. For a given organ in the image, its anatomy is i ∈{0,1} and d is the number of voxels in the image. i , i=1...d}. The method aims to study the shape of this organ, which is assumed to be characterized by a set of underlying features that can be extracted from the segmentation mask. We therefore assume the following generation process for the segmentation mask:

number

[0051] In one aspect, we are interested in inferring the parameters θ of the generative process as well as approximating the posterior distribution of the latent variables z given the segmentation mask X. The method relies on a variational autoencoder (VAE) framework (Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes. In: 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada (2014)) to estimate the model parameters. Thus, p(z) is defined as a multivariate Gaussian distribution with mean and identity covariance 0. The approximate posterior distribution q parameterized by φ is φ (z|X) is also introduced to optimize the lower bound L of the marginal log-likelihood, which is the segmentation mask X of object p. p can be written as follows for

number

number

[0052] To capture shape features, this method relies on a convolutional network and employs the U-Net encoder-decoder architecture without skip connections between the encoder and decoder. The U-Net architecture is disclosed in the following paper: Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer (2015), pp. 234-241.

[0053] In practice, the number of convolutional layers and convolutional blocks are automatically inferred from the nnUNet self-configuration procedure. Such an nnUnet framework is disclosed in the following paper: Isensee, F. et al.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021).

[0054] Due to this encoder-decoder architecture, the segmentation mask is progressively downsampled to obtain low-resolution feature maps that are mapped through a linear transformation to the latent variable z. The latent codes are then decoded by a symmetric path to reconstruct the original mask.

[0055] Anomaly Detection (AD) This method proposes to learn a normative model of organ shape by applying the previously presented VAE framework to segmentation masks of a large cohort of N healthy patients, allowing the model to capture the low-dimensional embedding characteristics of normal organ anatomy in latent space. Furthermore, as mentioned above, data augmentation is applied with such a procedure consisting of random translations, rotations, and scalings to enable the network to extract shape features that are invariant to these transformations. Based on this trained model, the method proposes two techniques for performing anomaly detection (AD) by leveraging the latent representation of normal organ shape.

[0056] Zero-shot learning method After training, the recognition model q φ (z|X) to obtain a segmentation mask X for a cohort of healthy subjects. p , we can project the normal shape onto the latent space 11 to obtain the empirical distribution of normal shapes (reference number 14 in Figure 2). The method relies on this low-dimensional distribution of normality to detect abnormal shapes.

[0057] To do so, the average of the projections 15 of healthy subjects is calculated and the abnormality of a representation 16 in the latent space is defined via the L2 distance 17 to this average latent representation.

[0058] Future learning method Another approach is to classify normal or healthy and abnormal or pathological shapes based on their low-dimensional representations. In practice, segmentation masks from a set of healthy and pathological objects in the latent space are used to model the recognition model q. φ (z|X). Thus, a set of low-dimensional organ shape features is obtained for all these subjects, and such a set is used to train any type of classifier (e.g., a linear SVM).

[0059] Examination of differences in organ shape Such a framework can also be used to study organ differences between groups, as shown in Figure 3. In one example, a set of healthy and pathological subjects and their segmentation masks 18 are considered. φ Based on (z|X), z healthy and z pathological For each group 19, 20, denoted by

number

[0060] experiment In this section, the method is applied to the case of the pancreas. A normative model of pancreatic shape was trained on a large cohort of healthy subjects and then utilized for anomaly detection (AD) on an independent test cohort. Several configurations were proposed to evaluate model performance, including the impact of the number of training subjects and latent space dimensionality on AD performance. Detection was performed using a support vector machine (SVM) with a few-shot learning method (i.e., making predictions based on a limited number of samples). Finally, it is shown how the proposed framework can be used to visualize differences between healthy and pathological pancreases.

[0061] training data Training dataset D 訓練(Reference number 23 in Figure 2) was created from a private cohort containing 2606 abdominal portal CT scans of patients with potential liver cancer. Several exclusion criteria were applied to ensure pancreatic health and shape (see section below). Finally, 1200 portal vein CT scans were retained. To examine the effect of the number of samples seen during training, subset D with an increasing number of subjects was created. N (See Table 1 below.) N For , 80% and 20% of the samples were used for training and validation, respectively. The partitioning was done so that the pancreas volume distribution was balanced across partitions.

[0062] Inclusion criteria Training data were selected from a dataset consisting of 2,606 portal venous phase abdominal CT scans acquired from patients with underlying liver disease. Each scan was accompanied by a radiological report and segmentation of the liver and liver lesions, performed manually by a radiologist. To exclude patients whose condition may affect the pancreas, different exclusion criteria were applied. First, all patients with metastatic cancer or lesions reported as "unevaluable" (or unable to be assessed) were excluded. Second, to avoid excessively large liver lesions, patients with lesion volumes above the 99th percentile were also excluded. Third, cases with lesions protruding from the liver were also excluded from the study. Finally, pancreas segmentation masks were automatically generated for the remaining patients, and only cases in which the mask consisted of one single connected component were retained. This last criterion served as a quality control for the segmentation masks. Finally, 1,200 cases were retained to build a healthy pancreas database for training.

[0063] [Table 1]

[0064] Test Data Test Database D 試験 (reference number 24 in Figure 2) is a set of two data sets, namely: i) A dataset of individuals including 144 cases diagnosed with pancreatic cancer (Ref. 24a) and assessed by expert radiologists as having abnormal pancreatic morphology.

number

number

[0065] More specifically, Figure 5 shows 試験 Examples of normal and abnormal pancreatic shapes from the pancreas are shown. Green and red numbers represent the

number

[0066] Pretreatment The first step was to obtain a pancreas segmentation mask.

number

number

[0067] Autoencoder Architecture As shown in Figure 4, the proposed autoencoder is a variational autoencoder (VAE) that follows a 3D encoder-decoder architecture. The network topology (number of convolutions per block, filter size) was selected based on the nnUNet self-configuration procedure. Thanks to stacked convolution and pooling operations, the network first encodes the 3D input into a 1D vector, also called the latent representation. Then, a transposed convolution decodes the latent representation to produce an output of the same size as the input.

[0068] Implementation details The model was trained for 800 epochs. A stochastic gradient descent optimizer was used with an initial learning rate of 10-4 followed by polynomial decay. A batch size of 8 was used with a gradient accumulation step of 5 to avoid memory issues and mimic larger batch sizes that are more suitable for VAEs.

[0069] result Zero-shot Anomaly Detection (AD) The model runs on different datasets

number

[0070] By increasing the dimension L of the latent space, each dataset D N The classification performance of improved. Furthermore, for each dataset size, L is set to the maximum value L 1024 The best results were consistently obtained when L was set to . The effect of the latent space dimension on performance appears to decay as the dataset size increases, and 16 From L 1024 When you move to D300 , D 600 , D 900 , D 1200 It is observed that the average AUC for D is improved by 10.7, 5.9, 4.0, and 3.1 points, respectively. Regarding the effect of database size, it is observed that increasing the training set size appears to improve the AUC scores overall. For example, D 300 From D 600 Moving on to the next section, all the experiments were successful, especially the L 16 The classification performance was improved.

[0071] This beneficial effect of both a larger training set and latent dimension was also observed in the Dice scores between the original and reconstructed segmentation masks. Therefore, in the following experiments, we chose the latent dimension L, as it gave the best results in terms of AUC and Dice scores. 1024 D 1200 We selected a model trained with

[0072] [Table 2]

[0073] For each experiment corresponding to a particular training size D and latent space dimension L, the mean and standard deviation of the AUC scores are reported in % obtained by bootstrapping with 10,000 iterations. The best results per row are underlined in bold per column.

[0074] To visualize the separation between normal and abnormal shapes, 試験 Each object from the recognition model q φ (z|X). Based on the latent representation of the object, three-dimensional reduction techniques were applied: principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and isomap. The results are displayed in Figure 6, where each point represents the latent projection of the reduced test object on a 2D plane. Independently of the projection technique, we observed a tendency for normal and abnormal shapes to separate into two distinct clusters.

[0075] More specifically, Figure 6 shows a 2D reduction of a subject's latent representation.

number

number

[0076] Fuseshot AD A linear support vector machine (SVM) classifier was tested using layered k-fold cross-validation. 試験 The classifier was trained on a latent representation of . The number of folds, k, was varied to test the performance of the classifier depending on the training / test sample ratio. Experiments ranged from a 0.05 training / test ratio to leave-one-out cross-validation and are presented in Table 3. Using only eight healthy and 15 abnormal training samples, performance increased to 78.9%. Furthermore, the AUC score and equilibrium accuracy were observed to increase with the number of training samples, reaching a maximum of 91.1% and 83.2%, respectively, for the leave-one-out configuration.

[0077] [Table 3]

[0078] For each experiment, the number of training samples and the number of outlier samples (in parentheses) are shown. The mean and standard deviation of AUC (area under the curve (%)) and balanced accuracy (%) were obtained by bootstrapping with 10,000 iterations.

[0079] Study of differences in pancreatic shape between groups To model the differences in pancreatic shape between healthy and pathological groups, we used the procedure presented earlier.

number

[0080] Such a framework may therefore be used to study and visualize morphological differences between organ shapes in different clinical groups based on latent space exploration. Anatomical changes may be consistent with clinical evidence, as shrinkage suggests partial parenchymal atrophy.

[0081] Reconstruction results As an additional indicator of the good reconstruction ability of the proposed model, in Table 4, D 試験 The Dice scores calculated between the input and output of the VAE for 244 subjects are also reported. The Dice score is a common metric used to measure the similarity of two samples. Here, it evaluates the overlap between the input segmentation mask and its reconstruction. The Dice score ranges from 0 (no overlap) to 100% (perfect match). With respect to the AUC scores obtained with zero-shot AD, the Dice scores increased with larger latent space dimensions and larger training sets. In particular, the best reconstruction performance was achieved with the D with latent dimension L = 1024. 1200 This was consistent with the best detection performance obtained with a model trained on

[0082] [Table 4]

[0083] For each experiment, we report the mean and standard deviation of the Dice scores obtained by bootstrapping with 10,000 iterations, corresponding to the specific training size and latent space dimension. The best results by row are underlined, and the best results by column are in bold.

[0084] Comparison with other methods Comparison with the baseline method The method presented here was compared to a baseline method that classifies shapes based on their volume. This method was then compared to a baseline method that classifies shapes based on their volume using bootstrap sampling. 試験 and obtained a mean AUC of 51% with a 95% confidence interval of [49.9; 51.7], below the previously reported maximum AUC scores of 65.4% and 91.1% for zero-shot and few-shot, respectively.

[0085] Comparison with other SSM methods The method presented here has also been compared with two state-of-the-art methods, namely, Active Shape Modeling (ASM) using Deformetrica software and Large Deformation Differential Metric Mapping (LDDMM).

[0086] The details and results reported in the following sections show that the method outperforms these two state-of-the-art methods in both zero-shot and few-shot configurations.

[0087] For ASM, a signed distance map of the pancreas 3D contour was calculated for each subject. For LDDMM, a Bayesian atlas parameterized by 576 control points was estimated. In both cases, PCA was performed on the shape coding parameters to obtain a latent vector of dimension 1024 for each subject. We also used the latent dimension L = 1024 to compare them with the method described in this paper. All methods were compared using D 1200 The results reported in Table 5 show that the method outperforms ASM and LDMM when the number of training samples is small (train / test ratio ≦0.05).

[0088] [Table 5]

[0089] For each experiment, the mean and standard deviation of the AUC (%) obtained by bootstrapping with 10,000 replicates is reported. The best results per column are in bold. [Explanation of symbols]

[0090] 1. Computer equipment 2 Input Interface 3. Memory 4 processors 5 Output Interface 6 Segmentation 7 images 8. Segmentation Mask 9 Data Augmentation Methods 10 Autoencoder 11 Latent space 12 Encoders 13 Decoder 14 Empirical distribution of normal shapes 15 Expressions 16 Expressions 17 distance 18 Segmentation Mask 19 groups 20 groups 21 Line 22 Segmentation Mask 23 Training Dataset D 訓練 24a Images showing pathological organs 24b Image showing healthy organs

Claims

1. 1. A computer-implemented method for characterizing at least one organ of a patient in at least one medical image, said method comprising the repeated steps of: (a) segmenting (6) an organ in at least one image (7) representing said healthy organ to obtain at least one segmentation mask (8) of said healthy organ; (b) generating an additional segmentation mask of the healthy organ using at least one data augmentation method (9) based on the segmentation mask obtained in step (a), wherein the data augmentation method (9) is a shape-preserving method; (c) training an autoencoder (10) using the segmentation mask obtained in step (a) and the additional segmentation mask obtained in step (b) to encode the shape of the organ in a latent space (11); A method comprising:

2. Further iterative steps of: (d) segmenting the image (24b) representing the healthy organ and the image (24a) representing the pathological organ to obtain segmentation masks (18) of the healthy organ and the pathological organ, respectively; (e) encoding the segmentation masks using the autoencoder (10) trained in step (c) to obtain representations (15, 16) of healthy and pathological organs in the latent space (11); The method of claim 1 , comprising:

3. Further steps below, (f) training a classifier capable of classifying representations of healthy and pathological organs in the latent space by using at least some of the representations (15, 16) obtained in step (e) as a training set for the classifier; The method of claim 2 , comprising:

4. Further steps below, (f1) segmenting an image representative of an organ to be characterized to obtain a segmentation mask of said organ to be characterized; (f2) encoding the segmentation mask using the autoencoder trained in step (c) to obtain a representation of the organ for characterization in the latent space; (f3) calculating the distance (17) in the latent space between the representation of the organ to be characterized and a point representing a healthy organ; The method of claim 2 , comprising:

5. The method of claim 4 , wherein the point is the mean of the representation of healthy organs in the latent space.

6. The method according to claim 4 or 5, wherein the distance is an L2 or Euclidean distance.

7. Further steps below, (g1) selecting at least one representation in the latent space; (g2) using the autoencoder trained in step (c) to generate the segmentation mask (22) of an organ by decoding the selected representation; (g3) outputting the generated segmentation mask; The method of claim 2 , comprising:

8. 8. The method of claim 7, wherein the selected representation is a representation selected between the average of the representations (19) of healthy organs and the average of the representations (20) of pathological organs.

9. The method according to any one of claims 1 to 8, wherein the image is a 3D image.

10. The method of any one of claims 1 to 9, wherein the image is a CT scan image.

11. The method according to any one of claims 1 to 10, wherein the organ is the pancreas.

12. The method according to any one of claims 1 to 11, wherein the autoencoder (10) is a variational autoencoder.

13. The method of claim 12 , wherein the variational autoencoder network is based on a UNet in which skip connections are removed to conform to the variational autoencoder framework.

14. The method of any one of claims 1 to 13, wherein the classifier is a support vector machine classifier.

15. The method according to any one of claims 1 to 14, wherein the shape-preserving data augmentation method (9) performs at least one random translation, rotation and / or scaling of the segmentation mask.

Images