Leveraging Complementary Attention Maps in Vision Transformers for OCT Image Analysis

[Download Paper]

Optical coherence tomography produces large volumes of retinal scans, so biomarker detection needs an automatic first pass. Local biomarkers and global ones call for different inductive biases: MaxViT is stronger on local structure, while EVA-02 is stronger on global context. Ensembling the two placed first in the IEEE VIP Cup 2023 OCT biomarker detection challenge, with a patient-wise F1 of 0.8527 in the final phase. Knowledge distillation then compresses the ensemble into a single MaxViT that is cheaper to run and still stronger.