Modélisation interprétable du diagnostic de pathologies rétiniennes par apprentissage profond

This thesis deals with the automatic recognition of retinal pathologies in two modalities: Optical Coherence Tomography (OCT) and fundus imaging. The work relies on various types of neural networks for disease recognition while imposing an additional condition necessary for their clinical acceptability: beyond the performance of an algorithm, it is essential to optimize its ability to produce predictions interpretable by a human operator. Several approaches have been experimented with. The first approach consists of using a segmentation model to decompose an image into a set of semantically meaningful structures for diagnosis. A convolutional neural network is trained to segment four classes of lesions in fundus imaging. Several databases are used for this purpose, and their detailed characterization reveals heterogeneity, which poses additional constraints on the model’s training due to divergent annotation styles. This raises questions about the model’s generalization capability, which is studied through the prism of inter-database compatibility. The result is the creation of a general model capable of modularly adopting various segmentation styles. The adaptation is achieved through the concept of adversarial attacks on an image origin identification probe placed at the heart of the network’s encoder. The outcome is a model with multiple segmentation styles, configurable by the user in a predictable manner, without retraining or modification. The obtained segmentation is converted into a graph representation, preserving the lesions’ locality in a lightweight image format. This graph is used for the automatic grading of diabetic retinopathy, comparing an existing architecture with an original proposal, both based on Graph Neural Networks (GNNs). By limiting the graph structure to segmented lesions, the causal link between these lesions and the diagnosis is intrinsically induced in the model, distinguishing it from previous published models and promoting transparency in decision-making. The second approach focuses on a recent image classification model, the Transformer. A new mechanism is introduced based on resampling, first by simple interpolation and then stochastically conditioned on the self-generated attention map of the model. This algorithm, called Focused Attention, produces a local attribution map indicating the structures read in the image by the network with precision, thus contributing to another form of interpretability in automated diagnosis. To test the developed algorithms and models, two clinical applications are studied: the first dedicated to monitoring the longitudinal evolution of age-related macular degeneration within a patient cohort, and the second to the development of a diagnostic support platform for tele-screening purposes. This work has resulted in several original technical contributions. Regarding generalization, the adversarial attack approach allows for multi-domain adaptation of a single conventional architecture without modifying or retraining it, setting it apart from existing works in the literature. It paves the way for training a single model with heterogeneous data where there may not necessarily be distinct source and target domains. In terms of interpretability, the work on Transformer architectures leads to a technique for generating attribution maps approved by clinicians. Transformers do not handle high-dimensional images, and Focused Attention opens the path to new architectures that are free from these memory constraints.