Image Quality Assessment of Diabetic Retinopathy Based on ADD-Net

Peiming Zhang, Tong Hu, Xiaoxiang Han, Xiao Liu, Jia Ni, Jialin Shen, Qiaohong LiuIEEE Access 2023 · DOI

This paper presents a significant step forward in making automated diabetic retinopathy (DR) screening more practical for real-world, resource-constrained environments. By introducing a lightweight architecture called ADD-Net and a refined set of image quality standards, the researchers demonstrate that high-performance medical AI doesn't always require massive computational overhead.

TL;DR

The authors developed ADD-Net, a lightweight deep learning model that achieves 97.03% accuracy in assessing retinal image quality while using significantly fewer parameters than mainstream models. By re-labeling existing datasets with a new, clinically-focused quality standard and utilizing dilated convolutions, the team created a tool optimized for deployment on portable fundus cameras and mobile devices.

What Problem Does This Paper Address?

Diabetic Retinopathy is a leading cause of blindness, and early screening via fundus (retinal) images is vital. However, a major bottleneck exists: roughly 25% of retinal images are of poor quality due to operator error, equipment limitations, or patient conditions.

Existing Retinal Image Quality Assessment (RIQA) models face two main hurdles:

  1. Computational Heavyweights: Most high-performing models (like VGG16 or ResNet50) are too large and slow for the portable or smartphone-based cameras increasingly used in rural or underserved areas.
  2. Dataset Deficiencies: Publicly available datasets often have unclear classification criteria or are heavily unbalanced, leading to models that are biased or fail to recognize subtle features necessary for a "gradable" diagnosis.

Methods

The researchers built ADD-Net on a DenseNet backbone, known for its efficient feature reuse. They introduced two primary technical innovations to optimize it for RIQA:

  • Dilated Convolutions in DenseLayers: Standard convolutions have a limited "view" (receptive field). Usually, researchers use pooling to see more of the image, but this loses spatial resolution. ADD-Net uses Hybrid Dilated Convolutions (HDC), which expand the receptive field without losing data or creating "holes" in pixel utilization (the "gridding effect").
  • CBAM (Convolutional Block Attention Module): The team integrated seven CBAM modules. These act as a "spotlight," helping the network focus on both where important features are (spatial attention) and which features matter most (channel attention), such as the optic disc or subtle lesions.

Additionally, the team moved away from simple "good vs. bad" labels. They developed a four-tier criteria (Excellent, Minor problems, More problems, Unqualified) based on the visibility of specific retinal structures (macula, vessels) and DR lesions.

Datasets

The study utilized three public datasets: DR1, DRIMDB, and the Kaggle DR image quality dataset.

  • The Kaggle dataset was the primary focus; the authors re-labeled and reclassified its images according to their new four-tier criteria.
  • To ensure a robust model, they balanced the final training set into a 1:1 ratio of Gradable (Excellent/Minor) vs. Ungradable (More problems/Unqualified) images.
  • Data augmentation (rotations and flips) was used to expand the training set from 3,776 to over 30,000 images.

Key Insights and Ideas

  • Quality is Task-Specific: The authors argue that RIQA shouldn't just look at "blur." An image is only "good" if a doctor can actually see the lesions needed to diagnose DR. Their new criteria bake clinical utility directly into the ground truth.
  • Efficiency via Receptive Fields: By stacking dilated convolutions with specific rates (1, 2, and 5), they achieved a large receptive field similar to much deeper networks but with a fraction of the parameters.
  • The "Masterpiece" Balance: The results suggest a "sweet spot" in architecture. ADD-Net matches the accuracy of ResNet50 but with roughly 65% fewer parameters and lower computational complexity (GMacs).

Key Results

ADD-Net outperformed or matched mainstream classical models across nearly every metric:

  • Accuracy: 97.03% (matching ResNet50, the best of the "heavy" models).
  • Sensitivity: 98.31% (crucial for medical screening to ensure poor images aren't missed).
  • Lightweight Profile: Only 8.92M parameters. For comparison, VGG16 has 138.36M parameters.
  • Computational Speed: Complexity was measured at 2.76 GMacs, making it significantly faster and more suitable for mobile deployment than ResNet50 (4.12 GMacs) or VGG16 (15.50 GMacs).

Limitations and Future Work

While the paper demonstrates excellent performance on the re-labeled Kaggle set, the authors acknowledge that the model's ultimate goal is deployment.

  • Real-world Hardware: The paper does not show the model running on an actual mobile device or portable fundus camera yet; this is the logical next step.
  • Generalization: While they used multiple datasets, the criteria rely on expert manual re-labeling. Future work could explore if these standards hold up across different global populations and diverse camera hardware.
  • Integration: The authors suggest this RIQA step should be a "warning system" that triggers an immediate retake if an image is poor, but the integration of this "loop" into a clinical workflow remains for future study.

Summary updated

BibTeX
@article{zhang2023image,
  author = {Peiming Zhang and Tong Hu and Xiaoxiang Han and Xiao Liu and Jia Ni and Jialin Shen and Qiaohong Liu},
  title = {{Image Quality Assessment of Diabetic Retinopathy Based on ADD-Net}},
  journal = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3318876},
}
Figure