Context encoder self-supervised approaches for eye fundus analysis

Abstract

The broad availability of medical images in current clinical practice provides a source of large image datasets. In order to use these datasets for training deep neural networks in detection and segmentation tools, it is necessary to provide pixel-wise annotations associated to each image. However, the image annotation is a tedious, time consuming and error prone process that requires the participation of experienced specialists. In this work, we propose different complementary context encoder self-supervised approaches to learn relevant characteristics for the restricted medical imaging domain of retinographies. In particular, we propose a patch-wise approach, inspired in the previous proposal of broad domain context encoders, and complementary fully convolutional approaches. These approaches take advantage of the restricted application domain to learn the relevant features of the eye fundus, situation that can be extrapolated to many medical imaging issues. Different representative experiments were conducted in order to evaluate the performance of the trained models, demonstrating the suitability of the proposed approaches in the understanding of the eye fundus characteristics. The proposed self-supervised models can serve as reference to support other domain-related issues through transfer or multi-task learning paradigms, like the detection and evaluation of the retinal structures or anomaly detections in the context of pathological analysis.