Abstract
Early screening for ophthalmic diseases is a public health imperative for preventing irreversible vision loss. However, real-world fundus datasets often exhibit multi-label comorbidities and long-tailed distributions, inducing diagnostic bias in conventional deep learning models. This letter proposes the Contrastive Dual-Prototypical Learning (CDPL) framework to address these challenges. CDPL utilizes a two-stage curriculum: Self-Supervised Learning (SSL) pre-training to extract invariant features, followed by orthogonality-regularized dual-prototypical fine-tuning that enforces angular separation to decouple overlapping pathological signals. Evaluated on the ODIR dataset using a ResNet-18 backbone, CDPL achieves a 9.7% Macro-F1 improvement over standard benchmarks. Qualitative t-SNE and Grad-CAM analyses confirm the structural refinement of the metric space and superior sensitivity to rare tail-class pathologies. The framework is designed for extensibility to multi-label medical imaging domains characterised by class imbalance, with theoretical generalizability supported by the dataset-agnostic construction of $\mathcal {L}_{\rm{ortho}}$.
Funding Information
  • Anusandhan National Research Foundation
  • Department of Science & Technology (DST), Govt. of India

This publication has 29 references indexed in Scilit: