Explainable AI for diabetic retinopathy detection using vision transformers.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) screening via color fundus photographs faces challenges due to image variability and subtle lesions.
• Accurate DR detection is crucial for preventing vision loss in diabetic patients.

Purpose of the Study:

• To develop and evaluate the All-ViT Hybrid framework for binary DR grading (No-DR vs. DR).
• To enhance the accuracy and robustness of automated DR screening systems.

Main Methods:

• Formulated DR grading as a binary triage task.
• Proposed the All-ViT Hybrid framework integrating complementary pretrained transformer backbones.
• Employed a stability-oriented training schedule with AdamW optimization and OneCycle scheduling.
• Utilized preprocessing techniques including luminance-space CLAHE and retinal field-of-view masking.
• Performed data splitting via stratified sampling and incorporated post-hoc threshold tuning and optional test-time augmentation.

Main Results:

• Achieved high performance metrics: Accuracy 0.9754, F1 0.9761, Precision 0.9658, Recall 0.9866.
• Demonstrated superior performance over competitive baselines, with notable gains in F1, accuracy, and recall.
• Results were validated on the APTOS 2019 Blindness Detection dataset.

Conclusions:

• The All-ViT Hybrid framework effectively integrates complementary transformer representations for robust and threshold-adjustable DR triage.
• The modular and extensible design holds potential for broader applications in multi-class grading and multi-modal settings.
• Future work will focus on external generalization and probability calibration across different devices and centers.