Task optimized vision transformer for diabetic retinopathy detection and classification in resource constrained early...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic Retinopathy (DR) is a major cause of preventable blindness globally.
• Early DR detection and classification are crucial but challenging, especially in resource-limited areas.
• Existing deep learning models struggle with long-range dependencies and computational demands for DR analysis.

Purpose of the Study:

• To develop an efficient AI model for Diabetic Retinopathy detection and severity classification.
• To optimize a Vision Transformer model for enhanced feature extraction and computational efficiency.
• To enable real-time DR screening on low-cost hardware like Raspberry Pi-4.

Main Methods:

• Introduced a Task-Optimized Vision Transformer (TOViT) model with optimized learning rates, attention heads, and embedding dimensions.
• Employed structured pruning and 8-bit quantization for model compression.
• Evaluated performance on three large-scale public datasets and implemented on Raspberry Pi-4 hardware.

Main Results:

• Achieved 99% classification accuracy and F1-scores over 93% across all DR stages.
• Demonstrated real-time performance on Raspberry Pi-4, processing at 8 frames per second with 120 ms latency.
• Confirmed the model's suitability for portable, point-of-care screening devices.

Conclusions:

• The TOViT model offers a scalable and clinically relevant solution for automated DR diagnosis.
• Optimized AI models can overcome computational limitations of traditional deep learning for medical imaging.
• This approach has significant potential to expand access to early retinal screening in global healthcare systems.