Comparative Analysis of Vision Transformers and Conventional Convolutional Neural Networks in Detecting Referable...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) detection from retinal images is crucial for preventing vision loss.
• Convolutional neural networks (CNNs) have dominated image classification tasks, including DR detection.
• Vision transformers (ViTs) show promise but their efficacy in referable DR detection remains underexplored.

Purpose of the Study:

• To compare the performance of Vision transformers (ViTs) and Convolutional Neural Networks (CNNs) for referable diabetic retinopathy (DR) detection.
• To evaluate the diagnostic accuracy of different ViT and CNN models using retinal photographs.

Main Methods:

• A retrospective study involving 48,269 retinal images from Kaggle, Messidor-1, and SEED datasets.
• Development and comparison of 5 CNN models and 4 ViT models for referable DR detection.
• Performance evaluation using Area Under the Operating Characteristics Curve (AUC), specificity, and sensitivity on internal and external test sets.

Main Results:

• The SWIN transformer (a ViT model) achieved the highest AUC (95.7% internal, 97.3% SEED, 96.3% Messidor-1), significantly outperforming all CNN models.
• At 80% specificity, the SWIN transformer demonstrated superior sensitivity (94.4%) compared to CNNs (76.3%-83.8%).
• These performance advantages were consistent across internal and external validation datasets.

Conclusions:

• Vision transformers (ViTs) offer superior performance compared to CNNs for detecting referable diabetic retinopathy from retinal images.
• ViT models, particularly the SWIN transformer, hold significant potential for enhancing deep learning-based DR detection systems.
• These findings suggest ViTs can optimize automated analysis of retinal photographs for improved DR screening and management.