INTELLIGENT SCREENING AND UNCERTAINTY QUANTIFICATION OF FUNDUS OCT LESIONS BASED ON AMSF-CPNET

Abstract

In response to the prominent challenges in primary-level fundus screening—including high missed diagnosis rates, opaque decision-making of deep learning models, and uncalibrated probabilistic outputs—this study develops an adaptive multi-scale statistical fusion network integrated with conformal prediction (AMSF-CPNet). The framework leverages multi-branch convolutional neural networks to capture fine-grained local features and Swin Transformer to model long-range global dependencies, enhanced by channel-spatial adaptive attention for optimal feature integration. Conformal prediction is introduced to provide statistically rigorous uncertainty quantification with guaranteed coverage rates, while a concept-guided interpretability module establishes links between model decisions and clinical semantics. Evaluated on the public OCT2017 dataset and external OCTDL dataset, AMSF-CPNet achieves an accuracy of 96.8% and an AUC of 99.1%. At α = 0.05, the conformal prediction coverage reaches 94.7% with an average prediction set size of 1.23. Cross-domain validation demonstrates strong generalization under device shifts, with fine-tuned accuracy reaching 92.4%. This work offers a reliable, interpretable, and robust AI solution for grassroots ophthalmic screening, supporting the advancement of tiered healthcare systems.