A Multi-Model Image Enhancement and Tailored U-Net Architecture for Robust Diabetic Retinopathy Grading

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a primary cause of preventable vision loss in diabetic patients.
• Early DR detection is crucial but challenging due to subtle progression and manual screening limitations.
• This study addresses early diagnosis and fine-grained lesion identification in DR.

Purpose of the Study:

• To introduce an AI-based framework for robust multiclass diabetic retinopathy classification.
• To enhance early diagnosis and lesion discrimination in retinal fundus images.
• To develop a scalable and interpretable solution for DR screening.

Main Methods:

• A hybrid AI framework incorporating preprocessing, a Hybrid Local-Global Retina Super-Resolution (HLG-RetinaSR) module, and hierarchical classification.
• The HLG-RetinaSR module combines deformable convolutional networks and vision transformers.
• Classification utilizes a hierarchical approach with CNN, DenseNet-121, and a RefineNet-U architecture.

Main Results:

• The HLG-RetinaSR and RefineNet-U approach achieved precision, recall, F1-score, and accuracy exceeding 99% across all DR severity levels.
• The system effectively highlights vascular abnormalities and reduces background noise.
• Performance surpassed existing state-of-the-art methods in accuracy and robustness.

Conclusions:

• The proposed hybrid AI pipeline offers a scalable, interpretable, and clinically relevant solution for DR screening.
• The system improves diagnostic reliability, supporting early intervention for diabetic retinopathy.
• This AI tool has the potential to significantly assist ophthalmologists in reducing preventable vision loss.