A Benchmark for Studying Diabetic Retinopathy: Segmentation, Grading, and Transferability.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients.
• High blood glucose damages retinal blood vessels, leading to DR.
• Deep learning shows promise for DR diagnosis, but current systems lack interpretability due to limited annotated data.

Purpose of the Study:

• To address the need for high-quality annotated data in diabetic retinopathy (DR) research.
• To introduce a large-scale, fine-grained annotated DR dataset (FGADR) for improved model training and evaluation.
• To establish benchmark tasks for DR lesion segmentation, grading, and multi-disease identification using transfer learning.

Main Methods:

• Construction of the FGADR dataset with 2,842 images, including 1,842 with pixel-level lesion annotations.
• Image-level grading by six ophthalmologists ensuring intra-rater consistency.
• Development of an inductive transfer learning method for ocular multi-disease identification.

Main Results:

• The FGADR dataset provides comprehensive annotations for advancing DR diagnosis research.
• Benchmark tasks facilitate evaluation of segmentation, joint classification/segmentation, and transfer learning for DR.
• Experimental results on FGADR establish baseline performance for state-of-the-art methods.

Conclusions:

• The FGADR dataset is a valuable resource for developing more accurate and interpretable AI-driven DR diagnostic tools.
• The established benchmarks and transfer learning methods pave the way for future advancements in automated ocular disease detection.
• The dataset will be publicly released to foster collaborative research in the field.