Sea-Net: Squeeze-And-Excitation Attention Net For Diabetic Retinopathy Grading

Ziyuan Zhao, Kartik Chopra, Zeng Zeng, Xiaoli Li

TL;DR

SEA-Net is a novel deep learning architecture that improves the multi-class grading of Diabetic Retinopathy (DR) by combining spatial and channel-wise attention mechanisms. By interleaving Squeeze-and-Excitation (SE) blocks with spatial attention layers and employing a hybrid loss function, the model achieves state-of-the-art performance in distinguishing subtle severity levels in retinal images.

What Problem Does This Paper Address?

Diabetic Retinopathy (DR) is a leading cause of blindness globally. While early detection is critical, manual grading by experts is a slow, expensive, and subjective process. Automated systems using computer vision have shown promise, but they face two major hurdles:

  1. Subtle Inter-class Differences: The visual differences between severity levels (e.g., "Mild" vs. "Moderate") are often extremely subtle, requiring "fine-grained" classification rather than simple binary detection.
  2. Data Imbalance: In most datasets, "Normal" cases vastly outnumber "Severe" cases, making it difficult for standard models to learn the features of advanced disease stages.

The authors propose SEA-Net to address these issues by focusing the model's "attention" on the most relevant pathological features (like microaneurysms or hemorrhages) while ignoring background noise.

Methods

The SEA-Net architecture is built on three main pillars:

  • The Backbone: A pre-trained ResNet-50 is used for initial feature extraction from retinal fundus images.
  • Dual Attention (The "SEA" in SEA-Net):
    • Spatial Attention: A series of $1 \times 1$ convolution layers refine features in the spatial dimension, helping the model "look" at the right parts of the retina.
  • Channel Attention (SE Blocks): Squeeze-and-Excitation blocks recalibrate the importance of different feature channels. This allows the network to emphasize meaningful feature maps while suppressing less useful ones.
  • Integration: The authors experimented with different placements for these blocks and found that placing SE blocks within the spatial attention network (alternating them) provided the best performance.
  • Hybrid Loss Function: To tackle class imbalance and improve feature separation, the authors combined:
  • Weighted Cross-Entropy Loss: Gives higher importance to minority classes.
  • Center Loss: Minimizes the distance between features of the same class and increases the distance between different classes, creating tighter, more distinct clusters in the feature space.

Datasets

The study utilized the EyePACS dataset from a popular Kaggle competition.

  • Scale: 33,566 retinal images.
  • Classes: 5-point scale (0: No DR, 1: Mild, 2: Moderate, 3: Severe, 4: Proliferative DR).
  • Evaluation: To ensure fair comparison with prior work, they used a balanced testing set of 1,560 images.
  • Preprocessing: Images were cropped, resized to $610 \times 610$, and underwent histogram equalization to improve contrast.

Key Insights and Ideas

  • Synergy of Attention: The "aha moment" here is that spatial and channel attention are not just additive; they are synergistic. By interleaving them, the model can adaptively recalibrate its focus based on both where a lesion is and what kind of feature it represents.
  • Beyond Accuracy: The use of Center Loss is a clever way to handle the "fine-grained" nature of medical grading. By forcing the model to learn distinct "centers" for each disease stage, it becomes much better at distinguishing between adjacent severity levels that look very similar to the naked eye.
  • Architecture Matters: The authors' ablation study showed that simply tacking SE blocks onto the end of a network is less effective than integrating them into the attention mechanism itself.

Key Results

SEA-Net outperformed several established baselines across multiple metrics:

  • Average Classification Accuracy (ACA): 0.5994 (compared to 0.5051 for a standard VGG-based approach and 0.5431 for BiRA-Net).
  • Macro-F1 Score: 0.6047, indicating robust performance across all five classes despite the inherent difficulty of the task.
  • AUC (Area Under Curve): 0.8760, demonstrating strong diagnostic potential.
  • Visual Evidence: t-SNE plots of the learned features showed 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.

Limitations and Future Work

The authors do not explicitly detail limitations, but a few are inherent to this type of study:

  • Computational Complexity: While more efficient than some bilinear pooling methods, the dual-attention mechanism adds parameters and compute time compared to a vanilla ResNet.
  • Generalization: The model was tested on a balanced subset of one specific dataset; its performance on "wild" clinical data with different camera types remains to be seen.
  • Future Directions: The paper suggests that this architecture could be extended to other medical imaging tasks where fine-grained classification is required, such as skin lesion grading or histopathology.

Summary updated

BibTeX
@inproceedings{zhao2020sea,
  author = {Ziyuan Zhao and Kartik Chopra and Zeng Zeng and Xiaoli Li},
  title = {{Sea-Net: Squeeze-And-Excitation Attention Net For Diabetic Retinopathy Grading}},
  booktitle = {ICIP},
  year = {2020},
  url = {https://openreview.net/forum?id=wB8SvbtkMyP},
}

Loading related content...

Figure