A Comprehensive Federated Learning Framework for Diabetic Retinopathy Grading and Lesion Segmentation

Diabetic retinopathy (DR) is a debilitating ocular complication demanding timely intervention and treatment. The rapid evolution of deep learning (DL) has notably enhanced the efficiency of conventional manual diagnosis. However, the scarcity of existing DR datasets hinders the progress of data-driven DL models, especially for pixel-level lesion annotation datasets, which severely impedes the advancement of DR lesion segmentation tasks required for precise interpretations of DR grading. Furthermore, the escalating concerns surrounding medical data security and privacy induce data collection challenges for traditional centralized learning, exacerbating the issue of data silos. Federated learning (FL) emerges as a privacy-preserving distributed learning paradigm. Nevertheless, the existing literature lacks a comprehensive FL framework for DR diagnosis and fails to exploit multiple diverse DR datasets simultaneously. To address the challenges of data scarcity and privacy, we construct a high-quality pixel-level DR lesion annotation dataset (TJDR) and propose a novel FL-based DR diagnosis framework including both DR grading and multi-lesion segmentation. Moreover, to tackle the scarcity of pixel-level DR lesion datasets, we propose <inline-formula><tex-math notation="LaTeX">$\bm {\alpha }$</tex-math></inline-formula>-Fed and adaptive-<inline-formula><tex-math notation="LaTeX">$\bm {\alpha }$</tex-math></inline-formula>-Fed, two efficient cross-dataset FL algorithms. Extensive experiments demonstrate the effectiveness of our proposed framework and the two cross-dataset FL algorithms.