CoT-XNet: contextual transformer with Xception network for diabetic retinopathy grading

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) grading relies on fundus image analysis, but subtle lesions pose challenges for current deep learning models like CNNs.
• Vision transformers offer alternative visual representations, potentially improving upon CNN performance in medical image analysis.

Purpose of the Study:

• To develop and evaluate a novel deep learning model, CoT-XNet, for enhanced accuracy in diabetic retinopathy grading.
• To leverage the complementary strengths of vision transformers and Xception networks for improved DR detection.

Main Methods:

• Proposed a two-path CoT-XNet model integrating contextual transformer (CoT) and Xception network representations.
• Implemented dedicated pre-processing, data resampling, and test-time augmentation strategies.
• Evaluated performance on over 50,000 images across three public datasets (DDR, APTOS2019, EyePACS) and compared with SOTA models.

Main Results:

• CoT-XNet outperformed existing state-of-the-art models in DR grading accuracy and Kappa scores across all tested datasets.
• Achieved accuracies of 83.10%, 84.18%, and 84.10% with Kappa values of 0.8496, 0.9000, and 0.7684, respectively.
• Class activation maps indicated that CoT and Xception networks provide different and complementary visual insights.

Conclusions:

• The CoT-XNet model demonstrates superior performance and generalizability in grading diabetic retinopathy from fundus images.
• Integrating diverse visual representations from CoT and Xception networks enhances diagnostic accuracy.
• CoT-XNet shows promise for widespread application in AI-assisted DR screening programs.