Annonce

Deep learning has demonstrated remarkable performance in the medical field, with accuracy exceeding that of medical experts. However, it has a major problem in that these models are "black box" structures, meaning they are not transparent and difficult for humans to understand, making them unreliable. The lack of interpretability of these models is a barrier to their use in clinical practice. To overcome this problem, several interpretability studies have been proposed within the field of Explainable Artificial Intelligence (XAI). One of the main objectives of this proposal is to explore and survey existing XAI methods, starting with the Gradient-Weighted Class Activation Mapping (Grad-CAM) method, which provides explainability by visualising the area the model focuses on when making predictions in the form of a heat map, intuitively showing the basis used by the model to make decisions. These methods will then be integrated within the application of medical image segmentation applied to various medical datasets available at http://medicaldecathlon.com/. The second objective is to integerate uncertainty based measures within the neural network to estimate and produce an uncertainty map, this part of the work is already in progress by the current intern. The final objective is to use both the XAI heat map and the uncertainty map to better interpret as well as improve the performance of the network.