Deep Learning Network with Illuminant Augmentation for Diabetic Retinopathy Segmentation Using Comprehensive...

Area of Science:

• Medical image analysis
• Computer vision
• Ophthalmology

Background:

• Diabetic retinopathy (DR) segmentation is challenged by domain shift and false positives from varied retinal backgrounds.
• Existing methods often fail to integrate crucial anatomical context, such as blood vessels and DR lesions, into training datasets.

Purpose of the Study:

• To enhance diabetic retinopathy segmentation by integrating comprehensive anatomical context into a DeepLabV3+ framework.
• To develop a novel training dataset that systematically includes DR lesions and complete retinal anatomical structures.

Main Methods:

• Utilized a DeepLabV3+ framework augmented with anatomical context.
• Created the first training dataset integrating DR lesions with optic disc, fovea, blood vessels, and retinal boundaries.
• Implemented illumination-based data augmentation and a two-stage training strategy (cross-entropy and Tversky loss).

Main Results:

• Achieved competitive performance (AUC-PR: 0.7715, IoU: 0.6651, F1: 0.7930) on IDRiD, DDR, and TJDR datasets.
• Demonstrated improved generalization on unseen datasets and reduced false positives through anatomical context awareness.
• Outperformed or matched state-of-the-art methods, including transformer approaches.

Conclusions:

• Comprehensive anatomical context integration is more critical than architectural complexity for DR segmentation.
• Systematic annotation and data augmentation can improve conventional network performance, efficiency, and interpretability.
• This approach establishes a new paradigm for medical image segmentation in DR.