Saliency Mask-Guided Local-Context Consistency for Retinal Anomaly Classification

Automated optical coherence tomography (OCT) diagnosis is clinically important for retinal disease assessment, but image-level classification can be limited by class imbalance and localized pathological patterns. Standard image-level classifiers may compress small pathological regions together with large areas of normal retinal tissue, reducing the contribution of subtle structural biomarkers during global pooling. To address this limitation, we propose Saliency Mask-Guided Local-context Consistency (SMGLC), a training framework that uses saliency maps from a frozen teacher proxy to extract lesion-focused local crops and align their feature representations with the corresponding whole-scan representations. This consistency objective encourages a lightweight student backbone to preserve local pathological cues while retaining a standalone inference pathway. We evaluate SMGLC on OCTDL and OCT2017 against four representative baseline architectures and same-student ablations under a fixed 8:1:1 evaluation protocol. On the OCTDL test split, SMGLC achieves an accuracy of 95.88%, an F1 score of 88.92%, and an AUC of 99.09%. On the OCT2017 test split, it reaches an accuracy of 95.58%, an F1 score of 92.93%, and an AUC of 99.37%. These results show that train-time local-context supervision remains competitive under the revised protocol without requiring an additional saliency branch at inference time.