CANet

CANet performs joint grading of diabetic retinopathy (DR) and diabetic macular edema (DME) by leveraging cross-disease attention to exploit their interrelationship using image-level supervision.

Key Features:

• Disease-Specific Attention Module: Selectively learns features pertinent to each disease individually (DR and DME) via attention mechanisms.
• Disease-Dependent Attention Module: Captures the internal relationship between DR and DME to model disease interactions relevant for joint grading.
• Integrated Deep Network Architecture: Integrates both attention modules into a deep learning framework to extract disease-specific and cross-disease features for grading.
• Image-level Supervision: Trains using only image-level labels, eliminating the need for location-specific annotations such as macula and soft/hard exudate markings.
• Performance Evaluation: Validated on the ISBI 2018 IDRiD challenge dataset and the Messidor dataset, reporting superior results on ISBI 2018 IDRiD and improved performance on Messidor compared with other methods.

Scientific Applications:

• Joint DR and DME grading: Automates simultaneous grading of diabetic retinopathy and diabetic macular edema from image-level supervision.
• Clinical decision support: Provides graded outputs to inform ophthalmologists' treatment planning and patient management.
• Benchmarking on public datasets: Serves as a validated method for comparison on ISBI 2018 IDRiD and Messidor datasets.

Methodology:

CANet integrates disease-specific and disease-dependent attention modules within a deep neural network trained with image-level supervision and evaluated on the ISBI 2018 IDRiD and Messidor datasets.