An interpretable dual attention network for diabetic retinopathy grading: IDANet

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a primary cause of adult vision impairment globally.
• DR often lacks early symptoms, leading to delayed treatment and irreversible vision loss.
• Accurate DR grading is difficult due to subtle and varied lesion patterns, especially small ones.

Purpose of the Study:

• To develop an advanced deep learning model for precise diabetic retinopathy grading.
• To enhance the detection of small, critical lesions indicative of DR.
• To improve classification performance in fine-grained DR grading.

Main Methods:

• A novel bi-directional spatial and channel-wise parallel attention network (IDANet) was proposed.
• The attention block was integrated into a backbone network to extract discriminative features.
• Model interpretability was assessed using LIME-generated activation maps for lesion visualization.

Main Results:

• The proposed IDANet demonstrated superior performance in DR grading across four benchmark datasets.
• The network effectively improved the detection of small-sized DR lesions.
• IDANet outperformed existing state-of-the-art methods in both classification and lesion detection tasks.

Conclusions:

• The IDANet offers a significant advancement in automated diabetic retinopathy grading.
• The attention mechanism is crucial for identifying subtle lesions and improving diagnostic accuracy.
• This approach holds promise for earlier and more effective management of diabetic retinopathy.