Reproduce “Fundus-Enhanced Disease-Aware Distillation Model for Retinal Disease Classification from OCT Images” on OpenResearch

About this paper

Lehan Wang, Weihang Dai, Mei Jin, Chubin Ou, Xiaomeng Li

The Fundus-enhanced Disease-aware Distillation Model (FDDM) improves retinal disease classification from OCT images by distilling knowledge from abundant, unpaired fundus images during training, without requiring fundus images during inference. This method achieved 69.06% Mean Average Precision (MAP) and 69.17% F1 score on a new clinical dataset, outperforming single-modal OCT models and existing multi-modal approaches that rely on paired data.

Key insights

• Distilling generalized, disease-specific class prototype vectors from a fundus teacher to an OCT student (CPM) effectively reduces feature noise and improves classification accuracy, especially for majority classes.
• Aligning class similarity matrices between modalities (CSA) successfully transfers crucial inter-disease relational knowledge and label co-occurrence patterns, proving particularly beneficial for minority class classification.
• It is possible to significantly enhance an OCT diagnostic model by leveraging knowledge from readily available, unpaired fundus images, thereby eliminating the need for strict multi-modal data pairing and enabling single-modality inference for clinical practicality.

Summary by alphaXiv.