Weakly Supervised Anomaly Localization and Segmentation of Biomarkers in OCT Images

Abstract

Identifying biomarkers from optical coherence tomography images is critical in diagnosing and treating ophthalmic diseases. Most existing biomarker segmentation methods require pixel-level annotations for training, which is time-consuming and labor-intensive. This paper proposed a novel weakly supervised biomarker localization and segmentation method. The framework includes a classification network and a teacher-student network to exploit category annotated data through contrastive learning and anomaly localization strategies based on knowledge distillation. The classification network combines cross-entropy loss and self- supervised contrastive loss to ensure that the model focuses on the characteristics of the biomarker of interest. We introduce a knowledge distillation-based anomaly localization method to localize biomarker-related pathological regions accurately. The trained classification network acts as a teacher model to guide the training of the student network to learn the features of normal OCT images. The biomarker regions can be accurately localized by the differences between the feature maps generated by the two networks. Experiment results on the public dataset demonstrate the effectiveness of the proposed method.