Anomaly-guided image segmentation on retinal OCT images

Abstract

Medical image segmentation can profit from the integration of contextual information provided by anomaly detection methods. This work investigates different strategies for using anomaly detection results as guidance in supervised segmentation of pathological retinal scans. Three state-of-the-art anomaly detection algorithms, differing in architecture, training methods, and original use cases, are evaluated in five guidance strategies, including an Attention MobileNet architecture. The best-performing approach achieves an improvement of 6.7% in the detection rate score compared to the baseline. This framework offers an enhanced tool for automated analysis of retinal scans, supporting earlier diagnosis and improved treatment planning.