Abstract
SEA-Net is a novel deep learning architecture designed to improve the multi-class grading of Diabetic Retinopathy (DR). By integrating Squeeze-and-Excitation (SE) blocks within a spatial attention framework and utilizing a hybrid loss function, the model achieves superior performance in distinguishing between subtle disease severity levels in retinal fundus images.
Introduction
Diabetic Retinopathy (DR) remains a primary cause of vision loss worldwide. While early intervention is vital, manual grading by clinicians is labor-intensive and prone to inter-observer variability. Automated grading systems face two significant challenges: the extremely subtle visual differences between adjacent severity levels (e.g., Mild vs. Moderate) and the inherent class imbalance in medical datasets, where healthy cases often dominate. The proposed SEA-Net addresses these by focusing on pathological features like microaneurysms while suppressing irrelevant background noise.
Methodology
The SEA-Net architecture leverages three core components to enhance feature representation:
- Backbone: A pre-trained ResNet-50 serves as the primary feature extractor.
- Dual Attention Mechanism: The model combines spatial attention (via $1 \times 1$ convolutions) to localize lesions and channel attention (via SE blocks) to recalibrate feature map importance. The authors found that interleaving SE blocks within the spatial attention network (Configuration D in Figure 1) yielded the best results.
- Hybrid Loss Function: To improve class separation, the model employs a combination of Weighted Cross-Entropy (to handle imbalance) and Center Loss (to minimize intra-class variance and maximize inter-class distance).
Experimental Setup
The study utilized the EyePACS dataset, comprising 33,566 retinal images across five severity grades (0: No DR to 4: Proliferative DR). Preprocessing included cropping, resizing to $610 \times 610$, and histogram equalization for contrast enhancement. Testing was conducted on a balanced set of 1,560 images to ensure fair comparison with prior work.
Results and Discussion
SEA-Net demonstrated significant improvements over baseline models, including VGG-based approaches and BiRA-Net. As shown in the table below, the optimized SEA-Net ($\lambda=0.1$) achieved an Average Classification Accuracy (ACA) of 0.5994 and an AUC of 0.8760.
The synergy between spatial and channel attention allows the model to adaptively focus on both the location and the nature of retinal lesions. Furthermore, the use of Center Loss proved crucial for fine-grained classification. Visual evidence from t-SNE plots in the study confirmed that the hybrid loss successfully clustered the five classes more distinctly than standard loss functions, particularly for the often-confused early stages of the disease.
Conclusion
While SEA-Net shows promise, future work should address its increased computational complexity compared to standard architectures and validate its performance on diverse clinical datasets. The dual-attention approach holds potential for other fine-grained medical imaging tasks, such as skin lesion or histopathology grading.
Summary updated