Abstract
The heterogeneous acquisition, variability of orientation, and subtle lesions continue to challenge the screening of diabetic retinopathy through color fundus photographs. We formulate DR grading as a binary triage task (No-DR vs DR) and propose the All-ViT Hybrid framework, integrating complementary pretrained transformer backbones within a stability-oriented training schedule (head-only warm-up, partial unfreezing), optimized using AdamW with OneCycle scheduling, class-weighted cross-entropy, and mixed precision. Preprocessing includes luminance-space CLAHE and retinal field-of-view masking, and data splitting is performed using stratified sampling to preserve class balance. We perform post-hoc threshold tuning and test-time augmentation (optional), which is available for operating-point control and robustness. Over the competitive baselines (ConvNeXt, MobileNetV3-Large, EfficientNet-B0, DenseNet201), All-ViT Hybrid has an Accuracy of 0.9754, F1 of 0.9761, Precision of 0.9658, Recall of 0.9866, and measures of agreement κ of 0.9509, MCC of 0.9511, and Jaccard of 0.9532. Compared to the best baseline, F1 achieves a gain of +1.11 percentage points, an improvement in accuracy of +1.13, and a recall improvement of +0.81 percentage points without compromising controllable precision and these experiments were conducted on the APTOS 2019 Blindness Detection dataset. These visualizations provide qualitative interpretability and are not quantitatively validated against lesion-level annotations due to dataset limitations. These findings suggest that integrating complementary transformer representations within a unified fusion pipeline provides strong and threshold-adjustable performance for DR triage under real-world variability. The framework remains modular and extensible, with potential applicability to larger input resolutions, multi-class grading, and multi-modal clinical settings. The future research program will test external generalization and probability calibration across devices and centers.