Système d'apprentissage multitâche dédié à la segmentation des lésions sombres et claires de la rétine dans les images de fond d'oeil

This work focuses on automatic diagnosis on fundus images, which are a bidimensional representation of the inner structure of the eye. The aim of this master's thesis is to discuss a solution for an automatic segmentation of the lesions that can be observed in the retina. The proposed methodology regroups those lesions in two categories: red and bright. Obtaining a simultaneous double segmentation is a novel approach; most of the previous works focus on the detection of a single type of lesions. However, due to time constraints and the tedeous nature of this work, clinicians usually can not test all the existing methods. Moreover, from a screening perspective, the clinician has no clue a priori on the nature of the pathology he deals with and thus on which algorithm to start with. Therefore, the proposed algorithm requires to be versatile, fast and easily deployable. Conforted by the recent progresses obtained with machine learning methods (and especially deep learning), we decide to develop a novel convolutional neural network able to segment both types of lesions on fundus images. To reach this goal, our methodology relies on a new multitask architecture, trained on a hybrid method combining weak and normal supervised training. The architecture relies on hard parameter sharing: two decoders (one per type of lesion) share a single encoder. Therefore, the encoder is trained on deriving an abstrast representation of the input image. Those extracted features permit a discrimination between both bright and red lesions. In other words, the encoder is trained on detecting pathological tissues from normal ones. The training is done in two steps. During the first one, the whole architecture is trained with patches, with a groundtruth at a pixel level, which is the typical way of training a segmentation network. The second step consists in weak supervision. Only the encoder is trained with full images and its task is to predict the status of the given image (pathological or healthy), without specifying anything concerning the potential lesions in it (neither location nor type). In this case, the groundtruth is a simple boolean number. This second step allows the network to see a larger number of images: indeed, this type of groundtruth is considerably easier to acquire and already available in large public databases. This step relies on the hypothesis that it is possible to use an annotation at an image level (globally) to enhance the performance at a pixel level (locally). This is an intuitive idea, as the pathological status is directly correlated with the presence of lesions.