Two-Stage Topological Refinement Network for Retinal Artery/Vein Classification

Abstract

Automated retinal artery/vein (A/V) classification could significantly speed up computer-aided diagnosis of various cardiovascular and systemic diseases. Despite the successful application of deep learning methods to A/V segmentation and classification, exploiting topological information in deep learning methods remains a challenging task. We propose a novel two-stage cascaded deep learning framework to spread the workload across a U-Net with dual decoders and a topological refinement GAN, with a focus on the pixel-level features and topological features respectively. The proposed framework accomplishes state-of-the-art performance in A/V classification on the public AV-DRIVE, INSPIRE-AVR and LES-AV datasets and effectively improves the topological connectedness of the classification results.