Abstract

Retinal anomaly detection plays a pivotal role in screening ocular and systemic diseases. However, progress in this field has been hindered by the lack of a comprehensive and publicly available benchmark for fair evaluation and methodological advancement. Existing retinal anomaly detection studies often rely on limited anomaly types, nearly saturated test sets, and insufficient generalization evaluation. Meanwhile, medical anomaly detection benchmarks mainly focus on one-class supervised settings, overlooking the labeled abnormal data and unlabeled data commonly available in clinical practice. To bridge these gaps, we introduce BenchReAD, a systematic benchmark for retinal anomaly detection that covers both data and algorithmic perspectives. By categorizing and benchmarking previous methods, we find that a fully supervised approach based on disentangled representations of abnormalities achieves strong performance but degrades on certain unseen anomalies. Inspired by memory bank mechanisms in one-class supervised learning, we further propose NFM-DRA, which integrates a Normal Feature Memory into DRA to reduce this degradation and establish a new state of the art.