Image & Video

ORViT-DR hybrid model hits 57% accuracy for low-res diabetic retinopathy grading

Hybrid CNN-Transformer tackles 28x28 retinal scans with ordinal-aware grading

Deep Dive

Diabetic retinopathy remains a leading cause of vision loss, and automated grading systems are critical for accurate screening. However, retinal images are often captured at low resolution in real-world clinics. ORViT-DR, presented at ECCT 2026, directly addresses this by combining convolutional local feature extraction with transformer-based global context modeling. Its backbone integrates BiT-ResNetv2 with a Vision Transformer in a hybrid architecture, specifically designed to respect the ordinal structure of DR severity — where neighboring classes like 'mild' and 'moderate' share overlapping visual patterns.

On the official RetinaMNIST test set (part of MedMNISTv2), ORViT-DR achieved 57.00% classification accuracy, a quadratic weighted kappa of 0.5963, and a macro-F1 of 0.4293. To stabilize training on these challenging 28x28 images, the authors applied progressive layer unfreezing, layer-wise learning rate decay, exponential moving average parameter updates, and ensemble-based inference. While accuracy remains modest, the strong kappa score indicates the model respects ordinal severity ordering — a key requirement for clinical grading. The results suggest hybrid CNN-Transformer models are a viable direction for low-resource DR screening.

Key Points
  • ORViT-DR fuses BiT-ResNetv2 and a Vision Transformer backbone to capture both local features and global context
  • Achieves 57.00% accuracy and 0.5963 quadratic weighted kappa on 28x28 RetinaMNIST images
  • Training uses progressive unfreezing, layer-wise LR decay, EMA updates, and ensemble prediction
  • Accepted at ECCT 2026 and available on arXiv (2608.16958)

Why It Matters

Reliable low-resolution DR grading could enable affordable, automated diabetic retinopathy screening in under-resourced healthcare settings.

📬 Get the top 10 AI stories daily