Attention dual transformer with adaptive temporal convolutional for diabetic retinopathy detection

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a leading cause of vision loss, necessitating accurate and early detection.
• Current DR detection methods often struggle with complex spatial and temporal features in retinal images.

Purpose of the Study:

• To propose and evaluate an Attention Dual Transformer with Adaptive Temporal Convolutional (ADT-ATC) model for improved DR detection.
• To enhance the analysis of multi-scale spatial features and temporal dependencies in retinal fundus images.

Main Methods:

• The ADT-ATC model utilizes a dual spatial transformer network to process multi-scale spatial features.
• An adaptive temporal convolutional unit captures temporal dependencies, crucial for DR identification.
• A hierarchical cross-attention module fuses spatial and temporal features for comprehensive analysis.

Main Results:

• The ADT-ATC model achieved 98.2% accuracy on the DRIVE dataset.
• The model demonstrated 97.7% accuracy on the Diabetic Retinopathy dataset.
• Performance surpassed conventional deep learning models in DR detection.

Conclusions:

• The proposed ADT-ATC model offers a significant advancement in automated DR detection.
• The model's ability to integrate spatial and temporal features leads to superior diagnostic performance.
• ADT-ATC shows promise for clinical application in early DR screening.