Anomaly-guided weakly supervised lesion segmentation on retinal OCT images

Jiaqi Yang, Nitish Mehta, Gözde Merve Demirci, Xiaoling Hu, Meera S. Ramakrishnan, Mina Naguib, Chao Chen, Chia-Ling Tsai

Abstract: Highlights•Presents a weakly-supervised learning with image-level labels for medical lesion segmentation.•Constructs anomaly-discriminative representations to incorporate anomalies in the training.•Utilizes global anomaly information to improve pseudo label precision for segmentation.•Demonstrates adaptability in handling various types of diseases or lesions.•Validates the effectiveness across three OCT datasets.