Dassault Systèmes

The disclosure notably relates to a computer-implemented method for forming a dataset of medical images. The dataset is configured to be used for training a model. The trained model is thereby configured to output a prediction from an input medical image. The dataset-forming method comprises obtaining a first set of medical images. The first set consists of annotated medical images. The first set is unbalanced with respect to a meta-property relative to each medical image. The dataset-forming method further comprises obtaining a second set of medical images. The second set comprises unannotated medical images. The dataset-forming method further comprises outputting a third set of medical images. The third set comprises the first set and unannotated medical images of the second set. The third set is balanced with respect to the meta-property. The dataset-forming method is an improved solution for forming a dataset of medical images.

The disclosure notably relates to a computer-implemented method for processing a medical image of a body of a patient on a horizontal support. The method comprises obtaining the medical image. The medical image comprises voxels. Each voxel is associated with a value. The value represents tissue density at the position represented by the voxel. The method further comprises discarding each voxel of the medical image associated with a value smaller than a predefined threshold. The discarding splits the image into connected components separated by discarded voxels. The method further comprises identifying, among the connected components, a connected component having a largest number of voxels. The method further comprises computing a curve separating the identified connected component from the remaining part of the medical image. The voxels bounded by the curve correspond to the body of the patient.

The disclosure notably relates to a computer-implemented method for generating, from one or more medical images each of an organ of a patient, a medical report compliant with a predefined set of medical report rules. The method comprises, automatically, determining positions and diameters of one or more lesions in the one or more medical images. The method further comprising, automatically, determining, based on determined positions and diameters of the lesions, a list of target lesions in accordance with the predefined set of medical report rules. The method further comprises, automatically, generating the medical report based on the determined target lesions and their positions and diameters.

The disclosure concerns a method for machine-learning a function configured constraints prediction for a sketch representing at least a portion of a mechanical part. The method comprises providing a training dataset of training examples, each comprising a graph representing a sketch. The graph comprises object nodes each representing a geometric object, geometry nodes each representing a geometric property, topological edges each representing a topological link, geometrical edges each connecting a geometry node to one object node, and constraint edges each representing a constraint and labeled with a label of a predefined set of labels each representing a constraint of a predefined constraints set. The method comprises training the function to take as input a graph representing a sketch and to output constraints each of a constraint edge of the graph and each belonging to the predefined constraints set.

The disclosure notably relates to a computer-implemented method for generating a template image of an organ. The method comprises obtaining a dataset comprising a set of medical images, and an autoencoder pre-trained on a set of medical images. The method comprises determining the template image, by minimizing in the latent space a loss which is a function of the latent vector and, for each medical image of the dataset, parameters of a diffeomorphic deformation applied to the template image to obtain the medical image. The loss includes a first term penalizing a discrepancy between a posterior probability of the parameters of the diffeomorphic deformation, given the medical image, and a Gaussian prior probability of the parameters of the diffeomorphic deformation. The loss includes a second term penalizing a negative log-likelihood of the set of medical images of the dataset, given the latent vector and the parameters of the diffeomorphic deformation.

The disclosure notably relates to a computer-implemented method for computing a score representing a percentage of occurrence of a clinical event. The method comprises obtaining a database including a list of P characteristics within a vector X, associated with the clinical event. The method comprises obtaining, for a given patient, the modalities corresponding to a subset of the characteristics in the list (p ∈ X*, wherein X* ⊂ X). The method comprises extracting, from the database, the modality weight associated with each modality obtained for the patient. The method comprises computing the score (score_patient) for the given patient by summing the extracted modalities weights, using the formula score_patient = Σ_{pj ∈}_X* α(p_j ). The method comprises computing a dispersion range associated with the computed score. Such a method forms an improved solution for computing a score representing a percentage of occurrence of a clinical event for a patient.

The disclosure notably relates to a computer-implemented method for machine-learning a neural network comprising a Spike Neural Network for predicting a time occurrence of a clinical event. The method comprises providing a training dataset of training examples. Each training example comprises one or more time series of values of respective physiological measures of a patient over a past time period and a corresponding ground truth of a time occurrence of the clinical event. The method further comprises training the neural network based on the training dataset. The neural network is trained for taking as input one or more time series of values of the respective physiological measures of a patient over a past time period and for outputting a prediction of probabilities over a time period of the occurrence of the clinical event. The method is an improved solution for predicting a time occurrence of a clinical event.

The disclosure notably relates to a computer-implemented method for encoding a time series of values of a physiological measure of a patient. The method comprises obtaining the time series of values of the physiological measure of the patient. The method further comprises, for each time point of the time series of values, obtaining a slope coefficient of the time series of values at the time point. The obtaining a slope coefficient is based on the time point, a predetermined number of previous time points, and a predetermined number of next time points. The method further comprises, for each time point of the time series of values, comparing the slope coefficient to a reference slope coefficient. The method is an improved solution for encoding a time series of values of a physiological measure of a patient.

A method for generation of a sequence of synthetic ultrasound images of an organ. The method comprises providing video diffusion model trained to take as input a sequence of semantically labelled 2D representations of the organ and a corresponding sequence of noise images, to generate a corresponding denoised sequence of ultrasound images respecting the labels. The method comprises providing several consecutive sequences of labelled 2D representations of the organ and corresponding consecutive sequences of noise images, iteratively applying the video diffusion model to each couple formed by one said consecutive sequence and one corresponding noise images sequence, including, at each iteration, using, for each denoising step of the backward process of the video diffusion model, a part of the denoised synthetic ultrasound images generated at the previous iteration for that denoising step in replacement of a part of images of the given couple for that denoising step.

The disclosure notably relates to a computer-implemented method for annotating a set of medical images of at least one anatomical target of a same patient. The method comprises: obtaining at least one query from the set of medical images, each query being an unannotated 2D slice of a medical image; obtaining a support from the set of medical images, the support comprising a set of pairs, each pair comprising a 2D slice of the medical image and a respective 2D annotation; obtaining a pre-trained few-shot model being configured: for taking as input at least one query and a support; and for outputting, based on the inputted support, a 2D annotation associated with the inputted at least one query. The method also comprises predicting, by using the pre-trained few-shot model, a 2D annotation associated with the obtained at least one query.