Fine-grained attention & knowledge-based collaborative network for diabetic retinopathy grading.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate diabetic retinopathy (DR) grading is vital for preventing vision loss.
• Current deep learning (DL) systems struggle with subtle DR lesions and lack medical knowledge integration.
• This limits the practical application and interpretability of automated DR grading systems.

Purpose of the Study:

• To develop a novel deep learning model, FA+KC-Net, for improved diabetic retinopathy (DR) grading.
• To address limitations in existing DL models by incorporating fine-grained attention and medical knowledge.
• To enhance the accuracy and interpretability of DR grading for clinical use.

Main Methods:

• Proposed a novel FA+KC-Net combining a fine-grained attention network and a knowledge-based collaborative network.
• The fine-grained attention network captures subtle, small image features from fundus images.
• The knowledge-based network extracts a priori medical knowledge features of DR lesions (MA, SE, EX, HE).
• Decision rules fuse results from both networks for final DR grading.

Main Results:

• FA+KC-Net demonstrated high accuracy and stability across four datasets (DDR, Messidor, APTOS, EyePACS).
• Achieved state-of-the-art performance on the DDR, Messidor, and APTOS datasets.
• The model effectively captures subtle lesions and integrates medical knowledge for improved grading.

Conclusions:

• The proposed FA+KC-Net significantly improves diabetic retinopathy grading accuracy.
• Integrating fine-grained attention with medical knowledge offers a promising direction for AI in ophthalmology.
• The model's performance suggests potential for practical application in clinical settings.