Angular contrastive distillation driven self-supervised scanner independent screening and grading of retinopathy

Abstract

Retinopathy is a group of retinal diseases that causes severe retinal damage, resulting in partial visual impairment or complete vision loss. Due to the capability of optical coherence tomography in revealing early retinal abnormalities, many researchers have utilized it to develop autonomous retinal screening systems. However, to the best of our knowledge, the majority of these systems are vulnerable against diversified scanner specifications, hence, requiring extensive training supervision (on large-scale datasets) to accurately learn retinal screening tasks. In this paper, we present a novel self-supervised segmentation-driven classification pipeline that employs a proposed angular contrastive distillation scheme to extract retinal lesions (from the multi-vendor data) in order to give lesion-aware scanner independent screening and grading of retinopathy. The diagnostic capacity of the proposed framework is further enhanced by the integration of a novel co-attention mechanism, which enables the underlying network to focus its attention on retinal abnormalities to effectively grade retinal diseases without incurring supervision from the ground truth labels. The proposed framework is rigorously validated on seven public datasets, acquired with four different scanners, where it outperforms its competitors by achieving 9.22% improvement in terms of mean intersection-over-union for extracting retinal lesions and 10.71% improvement in terms of F1 score for grading retinopathy.