HIRD-Net: An Explainable CNN-Based Framework with Attention Mechanism for Diabetic Retinopathy Diagnosis Using...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a major cause of vision loss globally.
• Computer-Aided Diagnosis (CAD) systems using fundus images face challenges like variable lesion scales, image blur, class imbalance, and computational demands.

Purpose of the Study:

• To develop an efficient, end-to-end deep learning framework for accurate Diabetic Retinopathy diagnosis.
• To address limitations in existing CAD systems for DR detection.

Main Methods:

• An enhanced preprocessing pipeline combining Contrast Limited Adaptive Histogram Equalization (CLAHE) and Dilated Difference of Gaussian (D-DoG) filtering.
• A novel deep learning architecture, Hierarchical-Inception-Residual-Dense Network (HIRD-Net), featuring hierarchical feature fusion, multiscale inception-residual-dense blocks, Squeeze-and-Excitation Channel Attention (SECA), and Focal Loss.
• Incorporation of Gradient-weighted Class Activation Mapping (Grad-CAM) for model interpretability.

Main Results:

• The proposed HIRD-Net framework achieved high classification accuracies: 93.46% on IDRiD-APTOS2019, 82.45% on DDR, and 79.94% on EyePACS datasets.
• The model demonstrated strong generalization capability and computational efficiency with only 4.8 million parameters.
• Explainable AI (XAI) using Grad-CAM provided visualization of the model's decision-making process.

Conclusions:

• The proposed HIRD-Net framework offers an effective and efficient solution for automated Diabetic Retinopathy diagnosis.
• The integration of advanced preprocessing and a novel deep learning architecture improves diagnostic accuracy and interpretability.
• This framework holds promise for early DR detection, potentially preventing vision impairment.