Method for processing an image obtained using a first imaging modality with a view to generating a second image belonging to a second imaging modality

Abstract

The invention relates to a method for processing a first image (1) of dimension N obtained using a first imaging modality, comprising a step of generating a synthetic second image (3) of dimension M belonging to a second imaging modality, this synthetic second image being obtained from the first image (1) of dimension N using an artificial intelligence model, this artificial intelligence model comprising: - at least one extractor (9) capable of extracting latent variables from the first image (1) of dimension N, and - at least one generator (11) for generating the synthetic second image of dimension M from the latent variables.

Description

Description Title of the invention: Method for processing an image obtained according to a first imaging modality to generate a second image belonging to a second imaging modality Technical field

[0001] The present invention relates to an image processing method for the detection of discriminating characteristics, particularly for diagnosis.

[0002] The present invention applies in particular in the medical field for the detection of pathologies such as knee osteoarthritis. But the invention has a broader scope since it can be applied outside the medical field, particularly when it involves super resolution of images. State of the prior art

[0003] Knee osteoarthritis, or KOA, is a degenerative, potentially disabling disease caused by the progressive wear and tear and loss of joint cartilage. It is most commonly seen in the elderly, with clinical symptoms and progression rates varying between individuals. Symptoms typically become more severe and frequent over time, resulting in pain, stiffness, swelling, and discomfort in the knee after prolonged sitting or rest.

[0004] Unfortunately, the exact cause of knee OA is still unknown, and there is no cure. Therefore, it is crucial to predict the early onset of knee OA and implement timely interventions, such as weight loss, to delay disease progression. Medical imaging technology has advanced significantly, and magnetic resonance imaging (MRI) has become a popular diagnostic tool for knee OA. MRI uses strong magnetic fields, magnetic field gradients, and radio waves to produce detailed images of organs and structures in the body. In addition to the always-necessary routine clinical examination of the symptomatic joint, the standard diagnosis of knee OA relies on radiographic imaging (plain X-rays), which are safe, cost-effective, and widely available.Despite these advantages, plain radiographs are insensitive when it comes to detecting early changes in knee OA, making early diagnosis of knee OA difficult in clinical practice.

[0005] However, MRI is very rarely available and is much more expensive than an X-ray and takes longer to perform. It is used as a last resort to confirm a diagnosis of knee osteoarthritis, for example. Furthermore, it cannot be performed on people with cardiac devices and ferromagnetic metal implants due to the strong magnetic field.

[0006] We know the document Z. Wang, A. Chetouani, D. Hans, E. Lespessailles, and R. Jennane, “Siamese-gap network for early detection of knee osteoarthritis” 2022, IEEE 19th International Symposium on Biomedical Imaging (ISBI), pp. 1-4, 2022, describing an approach, called Siamese-GAP network (“Global Average Pooling”), for the early detection of knee osteoarthritis using a KL grade classification.

[0007] We know the document A. Tiulpin, J. Thevenot, E. Rahtu, P. Lehenkari, and S. Saarakkala, “Automatic knee osteoarthritis diagnosis from plain radio-graphs: A deep learning-based approach” Scientific Reports, 2018, describing a transparent computer-assisted diagnosis method based on the “Deep Siamese Convolutional Neural Network” to automatically assess the severity of knee osteoarthritis according to the Kellgren-Lawrence (KL) grading scale.

[0008] We know the paper Y. Nasser, R. Jennane, A. Chetouani, E. Lespessailles, and ME Hassouni, “Discriminative regularized auto-encoder for early detection of knee osteoarthritis: Data from the Osteoarthritis Initiative” IEEE Transactions on Medical Imaging, vol. 39, no. 9, pp. 2976-2984, 2020, describing a discriminative regularized auto-encoder (DRAE) that allows learning both relevant and discriminative properties to improve classification performance.

[0009] Also known is the paper J. Masci, U. Meier, D. Ciresan, and J. Schmidhuber, “Stacked convolutional auto-encoders for hierarchical feature extraction” in International conference on artificial neural networks, pp. 52-59, Springer, 2011, describing the use of a convolutional auto-encoder.

[0010] We also know the document Sanuwani Dayarathna et al., “Deep learning based synthesis of MRI, CT and PET: Review and analysis” in Medical Image Analysis 92 (1er December 2023) 103046, describing different image synthesis techniques.

[0011] The present invention aims to detect pathologies such as gonarthrosis for example, at low cost.

[0012] The invention also aims to accelerate the detection of these pathologies. Statement of the invention

[0013] At least one of the objectives is achieved with a method of processing a first image of dimension N obtained according to a first imaging modality. This method comprises a step of generating a second synthetic image of dimension M belonging to a second imaging modality, this second synthetic image being obtained from the first image of dimension N using an artificial intelligence model. This artificial intelligence model comprises:

[0014] - at least one extractor capable of extracting latent variables from the first image of dimension N, and

[0015] - at least one generator for generating the second synthetic image of dimension M from the latent variables.

[0016]

[0017] With the method according to the invention, it was realized that starting from an image according to a first modality, for example a standard X-ray image, it is possible to generate a synthetic image of another modality, for example a pseudo 3D MRI image (Magnetic Resonance Imaging). The 3D image is made up of several two-dimensional images / two-dimensional planes. This method of generating a 3D MRI synthetic image from a 2D image is not intuitive for those skilled in the art. Indeed, the invention makes it possible in particular to generate several two-dimensional images at different depths from a single two-dimensional image. This is made possible by using an extractor whose purpose is to generate latent variables, these variables representing significant attributes of the input image.

[0018] With the invention, these latent variables feed a generator intended to generate images of a modality different or not from that of the input image.

[0019] In other words, the resulting latent variables from the extractor constitute the input to the generator, which uses the extracted representational features to generate images, such as synthetic MRI images.

[0020] The first image is advantageously a real X-ray image, but it can also be a synthetic image that is used as input to obtain a second synthetic image of a second modality.

[0021]

[0022] Combining an extractor with a generator according to the invention makes it possible to obtain output images that appear enriched compared to the input image. This process seems counterintuitive to those skilled in the art.

[0023] For example, the extractor can be configured to reduce the size of the first image.

[0024] With the invention, when the second synthetic image of dimension M is a 3D image of the MRI type, synthetic MRI data are thus obtained at low cost and quickly compared to a real MRI.

[0025] According to an advantageous characteristic of the invention, the generator can be a deconvolutional neural network, dCNN for “deConvolutional Neural Network” in English.

[0026] Other generator techniques can be used, such as a 3D convolutional autoencoder (the decoder acting as the generator), techniques based on UNet (fully convolutional network), VGG (Visual Geometry Group), ResNet (Residual Neural Network), etc., configured to perform deconvolution.

[0027] According to the invention, the extractor can advantageously be an encoder. Other extractors can be used, such as UNet ("fully convolutional network"), VGG for "Visual Geometry Group", ResNet ("Residual neural Network"), etc., which can also generate a latent vector to feed the generator.

[0028] According to an advantageous characteristic of the invention, the artificial intelligence model may further comprise a decoder which, associated with the encoder, forms an auto-encoder capable of generating a third synthetic image belonging to the first imaging modality.

[0029] In other words, we use a classic autoencoder comprising an encoder and a decoder, but we modify it by integrating a generator which captures the characteristics extracted by the encoder.

[0030] The complete autoencoder function, i.e., the reconstruction of a synthetic image corresponding to the input image, can be preserved. The resulting synthetic image, the third image, can also be used for classification as mentioned later.

[0031] With the invention, from a first image of a first modality, it is possible to generate both a second image which is a synthetic image of another modality, and a third image which is a synthetic image of the same modality as the first image.

[0032] Preferably, the autoencoder may be of the convolutional type.

[0033] Such an autoencoder provides excellent performance in X-ray reconstruction for example and the features extracted from the encoder show remarkable suitability for generating MRI images.

[0034] The technique according to the invention allows the use of numerous CNN blocks (Convolutional Neuronal Network) in the encoder, generator and decoder. This allows, in particular, the spatial information of the input images to be preserved as much as possible.

[0035] Other types of encoders can be used, such as the FCN (Fully Convolutional Network), the variational autoencoder (VAE), the contractive autoencoder or stacked Auto-encoder » in English, etc.

[0036] According to one embodiment of the invention, during a learning phase of the artificial intelligence model, for each given first image N, third synthetic images and second synthetic images can be generated iteratively, and at each iteration, a mean square error can be minimized according to the following formula:

[0038] With Xi and Æ2 two hyperparameters,

[0039] X a first image of dimension N obtained according to the first imaging modality,

[0040] X a third synthetic image belonging to the first imaging modality,

[0041] X2a real image of dimension M obtained according to the second imaging modality,

[0042] X2a second synthetic image of dimension M belonging to the second imaging modality.

[0043] Thus, during the training phase, we minimize both the difference between the third synthetic image and the first image, and the difference between the second synthetic image and real images corresponding to the first image. Said real images here are for example the MRI images of the same object.

[0044] It is thus understood that only the first images are used as training data feeding the model. The real images corresponding to the synthetic images are simply used to minimize the mean square error.

[0045] According to the invention, in the context of MRI images for example, different MRI sequences (IW, Tl, MPR, etc.) and different views (sagittal, coronal or axial) can be used.

[0046]

[0047] According to an advantageous characteristic of the invention, the hyperparameters can be defined as follows:

[0048] Xi + X2 = 1, with Xi and X2 included in [le-1,1],

[0049] Preferably, Xi = 0.6 and / ..2 = 0.4. Such values ​​allow to obtain good results.

[0050]

[0051] According to an advantageous embodiment of the invention, the method may further comprise a classification step during which a classification model is fed with the first image of dimension N and the second synthetic image of dimension M so as to extract characteristics discriminant ticks in each image and classify the first N-dimensional image according to predetermined groups.

[0052] In this embodiment, an additional operation of classification of the first image is carried out. Advantageously, the classification process is enriched by injecting into the classification model the second synthetic image which was generated from this same first image.

[0053] The classification model therefore receives as input, in parallel, on the one hand the first image of dimension N, and on the other hand the second synthetic image of dimension M.

[0054] According to another advantageous embodiment of the invention, the method may further comprise a classification step during which a classification model is fed with the third synthetic image and the second synthetic image of dimension M so as to extract discriminating characteristics in each image and classify the first image of dimension N according to predetermined groups.

[0055] In this embodiment, the first image is not used, but its synthetic image, which is the third image. However, it is possible to provide in a suitable classification model the parallel use of the first image, the second image, and the third image. Other inputs may be provided, such as data relating to the imaged object, such as a patient, in the case of medical imaging.

[0056] For example, the use of synthetic MRI images significantly improves the performance of the classification model compared to the use of standard X-ray images alone. The method according to the invention makes it possible to envisage a high-precision tool for diagnosis.

[0057] According to one embodiment of the invention, the classification model may comprise:

[0058] - a first branch of the model which is a neural network to independently generate a first vector of discriminating characteristics from the first image of dimension N,

[0059] - a second branch of the model which is a neural network to independently generate a second vector of discriminating characteristics from the second synthetic image of dimension M,

[0060] - a function to concatenate the first vector and the second vector,

[0061] - a probability distribution function allowing the classification of the first image of dimension N according to predetermined groups.

[0062] For the first branch of the model, other techniques can be used such as DenseNet, ResNet, VGG, etc.

[0063] For the second branch of the model, other techniques can also be used such as 3D Inception V3, 3D DenseNet, ResNet 3D, 3D VGG, etc.

[0064] Preferably, the first vector and the second vector may be of the same size. Depending on the merging technique used, the first vector and the second vector may be of different sizes.

[0065] According to an advantageous embodiment of the invention:

[0066] - the first image can be a two-dimensional image,

[0067] - the second computer-generated image can be a three-dimensional computer-generated image,

[0068] - the extractor can be an encoder of a two-dimensional convolutional neural network, and

[0069] - the generator can be a three-dimensional deconvolutional neural network.

[0070] Such an arrangement fits perfectly with a first modality being an X-ray and a second modality being an MRI imaging. The second synthetic image obtained is a 3D image made up of several two-dimensional images commonly called planes, sections or "slices" in English, that is to say a sequence of images (planes or sections) according to a given view.

[0071] Advantageously, the classification model may be a fusion classification model in which:

[0072] - the first branch of the model is a two-dimensional convolutional neural network,

[0073] - the second branch of the model is a three-dimensional convolutional neural network.

[0074] Advantageously, the first modality and the second modality can be constituted by one of the following pairs:

[0075] - Standard X-ray / Tomography (CT)

[0076] - Dual Energy X-ray Absorptiometry (DEXA) / Standard X-ray,

[0077] - Standard X-ray / Dual Energy X-ray Absorptiometry (DEXA)

[0078] - Tomography (CT) / MRI

[0079] - Standard X-ray / EOS X-ray; EOS being a system allowing images of the entire body to be taken in a functional position.

[0080] Many combinations are possible. In a preferred example, the dimension M is greater than or equal to the dimension N. We can also consider the opposite case where N is greater than M.

[0081] According to another aspect of the invention, there is provided a computer program product comprising instructions which, when the program is executed by a computer, lead it to implement the process as described above.

[0082] Also provided is a computer-readable recording medium comprising instructions which, when executed by a computer, cause the computer to carry out the method as described above. Description of figures and embodiments.

[0083] Other advantages and particularities of the invention will appear on reading the detailed description of implementations and embodiments which are in no way limiting, and the following appended drawings:

[0084] [Fig. 1] Figure 1 is a schematic view of the overall method according to the invention,

[0085] [Fig. 2] Figure 2 is a schematic view of a conventional autoencoder according to the prior art,

[0086] [Fig. 3] Figure 3 is a schematic view of a convolutional autoencoder according to the invention integrating a generator,

[0087] [Fig. 4] Figure 4 is a schematic view of a classification model according to the invention.

[0088] The embodiments which will be described below are in no way limiting; it will be possible in particular to implement variants of the invention comprising only a selection of characteristics described below isolated from the other characteristics described, if this selection of characteristics is sufficient to confer a technical advantage or to differentiate the invention compared to the state of the prior art. This selection comprises at least one preferably functional characteristic without structural details, or with only a part of the structural details if this part alone is sufficient to confer a technical advantage or to differentiate the invention compared to the state of the prior art.

[0089] In the figures, elements common to several figures retain the same reference.

[0090] Although the invention is not limited thereto, a method according to the invention will now be described for the classification of a 2D standard X-ray image by generating a 3D MRI synthesis image for the early detection of knee osteoarthritis (KOA).

[0091] Figure 1 is a schematic view of the overall method according to the invention.

[0092] A first 2D image 1 of an X-ray of a knee is distinguished. This first image 1 feeds a generation model 2 according to the invention to generate a second 3D synthetic image 3 of the MRI type.

[0093] The first image 1 and the second image 3 then feed a fusion classification model 4 to predict a radiological KL score. The KL score is a composite index created by Kellgren and Lawrence and allows a classification into 5 classes: no osteoarthritis, doubtful osteoarthritis, minimal, certain, advanced.

[0094] The generation 2 model, with or without the fusion classification model 4, constitutes the artificial intelligence model according to the invention.

[0095] The method according to the invention relates to a new approach based on deep learning for, for example, the early detection of knee osteoarthritis. In the example described, two models are distinguished: the MRI generation model and the fusion classification model. The MRI generated from an X-ray is combined with this X-ray to obtain additional information for predicting the KL score.

[0096] The MRI generation model uses an autoencoder as seen in Figure 3.

[0097] A classic autoencoder (AE) is depicted in Figure 2. It is an unsupervised learning model composed of an encoder 5 and a decoder 6. It is based on a backpropagation algorithm and optimization methods (such as the gradient descent method), and uses the input data X as supervisory data to make the neural network learn a mapping relationship, thus obtaining a reconstructed output X. Encoder 5 aims to encode the high-dimensional input X into a low-dimensional hidden variable h so that the neural network learns the most informative features. Decoder 6 then processes h to reconstruct the input data. The differences between the reconstructed data and the initial data are used to measure the error made by the autoencoder.Training consists of modifying the autoencoder parameters in order to reduce the reconstruction error measured on different training data, i.e. minimize(dist(X,X)).

[0098] The process of encoding the original data X from the input layer to the hidden layer is as follows:

[0099] h = gei (X) = o(WiX + bi)

[0100] where gei is the encoder function, G is an activation function, Wi and bi are the encoder weight and bias matrix respectively.

[0101] The process of decoding the hidden variable h from the hidden layer to the output layer is as follows:

[0102] X = g 02 (A) = o(W2à + b 2)

[0103] where ge2 is the decoder function, W2 and b2 are the decoder weight and bias matrix respectively.

[0104] Usually, the AE model minimizes the reconstruction error using the Mean Square Error or MSE cost function, which is defined as:

[0106] where N is the number of samples in each training batch, and T represents the training set.

[0107] Figure 3 is a schematic view of the generation 2 model according to the invention, in this case it is MRI type synthetic images which are generated. The autoencoder chosen for the invention is a convolutional autoencoder (CAE) 7 capable of constructing a synthetic image 8 from the first 2D radiography image 1.

[0108] According to the invention, the latent variables h extracted between the encoder 9 and the decoder 10 are used as input data of the generator 11, which is based in this example on a 3D deconvolutional neural network. A deconvolutional neural network architecture is notably described in the document by Feilong Cao et al., “Deconvolutional neural network for image super-resolution”, Neural Networks Volume 132, December 2020, Pages 394-404.

[0109] In the example according to the invention, the output of the generator 11 are 3D synthetic images equivalent to a 3D MRI sequence. These are 3D synthetic images resembling MRI images that would have been obtained by performing an MRI of the same imaged object to obtain the first 2D dimension image 1.

[0110] Within the framework of the invention, in the operating phase, we can envisage a generation 2 model consisting solely of the encoder 9 and the generator 11.

[0111] The autoencoder's encoder is cleverly used to extract informative features and feed the generator.

[0112] Encoder 9 in Figure 3 consists of six 2D CNN (Convolutional Neural Network) blocks. These blocks consist of a set of 2D convolutional layers of the same 3x3 kernel size with different depths (i.e., 32, 64, 128, 256, 512, and 1024) - a Batch Normalization (BN) layer - a LeakyReLU (Leaky Rectified Linear Unit) layer with a predefined slope of 0.2.

[0113] Decoder 10 in Figure 3 has an almost mirror structure of encoder 9. It consists of seven 2D deconvolutional neural network blocks. These blocks consist of a set of 2D deconvolutional layers of the same 4x4 kernel size with different depths (i.e., 1024, 512, 256, 128, 64, 32, and 1) - a BN layer - a LeakyReLU layer with a predefined slope of 0.2.

[0114] Generator 11 is based on a 3D deconvolutional neural network (dCNN) architecture. This architecture consists of seven 3D dCNN blocks and two 2D CNN blocks. Each 3D dCNN block consists of a set of 3D deconvolutional layers of the same kernel size of 4x4x4 with different depths (i.e., 1024, 512, 256, 128, 64, 32, and 16) - a 3D BN layer - a ReLU layer. Then, for the next first 2D CNN block, it consists of a set of 2D convolutional layers of kernel size 1x1 with depths of 8 - a BN layer - a ReLU layer. Finally, for the last 2D CNN block, it consists of a 2D convolutional layer of kernel size 1x1 with depths of 128 with an output channel of 20.The resulting latent variables between encoder 9 and decoder 10 constitute the input to generator 11, which uses the extracted representation features to generate the corresponding MRI images 3.

[0115] During the training phase, outputs 8 and 3 are used to minimize on the one hand the deviation of output 8 with the input image 1 and on the other hand the deviation of output 3 with real MRI images 12. However, these real images 12 do not feed the generation model.

[0116] This hybrid minimization of the MSE error is expressed according to a hybrid loss function J

[0117] Hybrid =

[0118] with Xi and X2 being hyperparameters G [le— 1, 1] requiring XI + X2 = 1. Preferably, XI = 0.6 and XI = 0.4.

[0119] In Figure 4 is represented a fusion classification model 4. It consists of two branches (i.e. a 2D CNN and a 3D CNN), its input is composed of the first 2D radiographic image 1 and a synthetic image 3 which is of the 3D MRI sequence type. Instead of the first image 1, or in parallel if we add a new branch to the network, we can use the synthetic image 8.

[0120] Input images 1 and 3 are combined as an input pair of the fusion classification model to assess the severity of knee osteoarthritis (KOA) in the described example. This combination of X-ray and MRI images significantly improves the performance of the classification model compared to using the X-ray image alone or the actual MRI images alone.

[0121] This makes it possible to simplify and reduce the cost of systems, for example, for the detection of knee osteoarthritis.

[0122] The fusion classification model according to the invention makes it possible to combine the 2D X-ray images and 3D MRI-like computer-generated images of a patient and extract discriminatory features via two independent CNN-based models, respectively.

[0123] The first branch of the fusion classification model 4 uses a 2D convolutional neural network 14 that generates a vector 15 of size 2048 from the 2D X-ray image 1.

[0124] The second branch of the fusion classification model 4 uses a 3D convolutional neural network 16 which generates a vector 17 of size 2048 from the 3D MRI-like synthetic image 3.

[0125] The 3D convolutional neural network 16 has thirteen layers, namely nine CNN layers and four pooling layers in four blocks (i.e., 2, 3, 4, and 4). Each of these four blocks is composed of a combination of 3D convolution with filters of different sizes (i.e., 32, 64, 128, and 256) for layers having a same kernel size (3x3x3), rectified linear unit (ReLU), and a 3D max-pooling layer of size (2x2x2). Then, two fully connected layers with node lengths of 4,096 and 2,048 are used to obtain the feature vector 17 of length 2,048.

[0126] Vectors 15 and 17 are of the same size so as to take into account equally the characteristics of both inputs (X-ray images and pseudo MRI). Other combination techniques can be used with the same size or different sizes between the two vectors.

[0127] These resulting vectors are then concatenated to obtain a final 18 feature vector of length 4096. Finally, the probability distribution of KL-0 and KL-2 grades corresponding to early detection of osteoarthritis is obtained via a Softmax layer 13.

[0128] As a classification loss function, we use the cross-entropy (CE) loss defined by:

[0132] with y s a ground-truth label converted into vectors of zeros and ones of sample s; Ÿ and T are the predicted and actual labels, respectively; k represents the KL score of sample s, Py |T is the conditional probability distribution; K is a set of KL scores (KL-0 and KL-2).

[0133] The generation model can be implemented in the following way. Initialization of the model weights can be done using the Kaiming initialization method and training of the model can be done using the Adam optimizer for 1000 epochs. A batch size can be set to 32 and learning rates of 0.001 and 0.0001 are used for the encoder and decoder on the one hand, and the generator on the other hand, respectively.

[0134] The classification model can be implemented as follows. The initialization of the model weights can be performed using the Kaiming initialization method, and the model training can be performed using the Adam optimizer for 500 epochs. The batch size can be set to 8 and a learning rate of 0.001. To handle the imbalanced dataset, the bootstrap-based oversampling method is used. Then, the balanced dataset was randomly divided into training, validation, and test sets in a ratio of 7:1:2 for each KL score, respectively. Data augmentations, including random rotation, brightness, contrast, jitter, and gamma correction, are performed randomly. In addition, to avoid overfitting, weight decay with a coefficient of 0.0003 was applied.

[0135] The present invention is a novel approach to improve the detection performance of early grades of knee osteoarthritis (KO A), i.e., KL-0 and KL-2. The overall model includes a pseudo MRI generation model and a fusion classification model. The generation model is based on a CAE and a 3D dCNN to generate pseudo MRI sequence images of the knee from an X-ray. These pseudo MRIs are combined with the X-ray to assist in the diagnosis of KOA for patients who only have X-rays.

[0136] The convolutional autoencoder (CAE)-based generation model is capable of holistically extracting the features embedded in 2D X-ray images. These features are advantageously used to generate MRI sequence images via a generator.

[0137] Knee osteoarthritis (KOA) is a common musculoskeletal disorder, and radiographs are commonly used for its diagnosis due to their cost-effectiveness. On the other hand, magnetic resonance imaging (MRI) is a more advanced imaging technique that provides detailed soft tissue information and has become an option in supplementary diagnosis for the assessment of KOA. Unfortunately, the high cost of MRI means that most patients with KOA only have radiographic images available. The present invention thus addresses this problem by proposing a novel approach to improve the early diagnosis of KOA, including specifically distinguishing between KL-0 and KL-2 stages.

[0138] Of course, the invention is not limited to the examples just described. Many modifications can be made to these examples without departing from the scope of the present invention as described.

Claims

Claims

1. 1. Method for processing a first image (1) of dimension N obtained according to a first imaging modality, comprising a step of generating a second synthetic image (3) of dimension M belonging to a second imaging modality, this second synthetic image being obtained from the first image (1) of dimension N using an artificial intelligence model, this artificial intelligence model comprising: - at least one extractor (9) capable of extracting latent variables from the first image (1) of dimension N, these variables representing significant attributes of the first image, and - at least one generator (11) for generating the second synthetic image of dimension M from the latent variables, the first modality and the second modality being a specific pair of imaging technologies.

2. 2. Method according to claim 1, characterized in that the generator (11) is a deconvolutional neural network (dCNN).

3. 3. Method according to claim 1 or 2, characterized in that the extractor is an encoder (9).

4. 4. Method according to claim 3, characterized in that the artificial intelligence model further comprises a decoder (10) which, associated with the encoder, forms an auto-encoder (7) capable of generating a third synthetic image (8) belonging to the first imaging modality.

5. 5. Method according to claim 4, characterized in that the autoencoder is of the convolutional type.

6. 6. Method according to claim 5, characterized in that during a learning phase of the artificial intelligence model, for each given first image N, third synthetic images and second synthetic images are iteratively generated, and at each iteration, a mean square error is minimized according to the following formula: With Xi and Æ2 two hyperparameters, X a first image of dimension N obtained according to the first imaging modality, X a third synthetic image belonging to the first imaging modality, X2 a real image of dimension M obtained according to the second imaging modality, X2a second synthetic image of dimension M belonging to the second imaging modality.

7. 7. Method according to claim 6, characterized in that Xi + X2 = 1, with Xi and ta included in [le-1,1],

8. 8. Method according to claim 7, characterized in that Xi = 0.6 and ta = 0.

4.

9. 9. Method according to any one of the preceding claims, characterized in that it further comprises a classification step during which a classification model (4) is fed with the first image (1) of dimension N and the second synthetic image (3) of dimension M so as to extract discriminating characteristics in each image and classify the first image of dimension N according to predetermined groups.

10. 10. Method according to any one of claims 4 to 8, characterized in that it further comprises a classification step during which a classification model (4) is fed with the third synthetic image (8) and the second synthetic image (3) of dimension M so as to extract discriminating characteristics in each image and classify the first image of dimension N according to predetermined groups.

11. 11. Method according to claim 9 or 10, characterized in that the classification model comprises: - a first branch of the model which is a neural network for independently generating a first vector (15) of discriminating characteristics from the first image (1) of dimension N, - a second branch of the model which is a neural network for independently generating a second vector (17) of discriminating characteristics from the second synthetic image (3) of dimension M, - a function to concatenate the first vector and the second vector, - a probability distribution function (13) making it possible to classify the first image of dimension N according to predetermined groups.

12. 12. Method according to claim 11, characterized in that the first vector and the second vector are of the same size.

13. 13. Method according to any one of the preceding claims, characterized in that: - the first image is a two-dimensional image, - the second computer-generated image is a three-dimensional computer-generated image, - the extractor is an encoder of a two-dimensional convolutional neural network, and - the generator is a three-dimensional deconvolutional neural network.

14. 14. Method according to any one of claims 11 to 13, characterized in that the classification model is a fusion classification model in which: - the first branch of the model is a two-dimensional convolutional neural network, - the second branch of the model is a three-dimensional convolutional neural network.

15. 15. Method according to any one of the preceding claims, characterized in that the first modality and the second modality are constituted by one of the following pairs: - Standard X-ray / Tomography (CT) - Dual Energy X-ray Absorptiometry (DEXA) / Standard X-ray, - Radiography / Dual Energy X-ray Absorptiometry (DEXA) - Tomography (CT) / MRI - Standard X-ray / EOS X-ray.

16. 16. Computer program product comprising instructions which, when the program is executed by a computer, cause the latter to implement the method according to any one of the preceding claims.

17. 17. A computer-readable recording medium comprising instructions which, when executed by a computer, cause the computer to carry out the method according to any one of claims 1 to 15.

Images