Related Experiment Video
Updated: Aug 29, 2025
05:54
Quantification of Diabetes-induced Adherent Leukocytes in Retinal Vasculature
Published on: January 24, 2025
401
Resampling-based cost loss attention network for explainable imbalanced diabetic retinopathy grading.
Haiyan Li1, Xiaofang Dong1, Wei Shen2
1School of Information, Yunnan University, Kunming, 650504, PR China.
Computers in Biology and Medicine
|September 4, 2022
Summary
A novel deep learning network addresses challenges in diabetic retinopathy (DR) grading by balancing data, enhancing lesion detection with attention, and improving accuracy for small sample classes. This explainable AI approach achieves state-of-the-art results in DR grading.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic retinopathy (DR) is a leading cause of blindness globally.
- Current DR grading methods face challenges including imbalanced datasets, difficulty detecting small lesions, low accuracy for underrepresented classes, and limited explainability.
- Automated DR grading systems require robust solutions to overcome these limitations for clinical application.
Purpose of the Study:
- To propose a novel, explainable, and attention-based deep learning network for improved diabetic retinopathy grading.
- To address data imbalance, enhance detection of subtle lesions, and increase grading accuracy, particularly for small sample classes.
- To provide visual interpretability of the model's decision-making process for clinical trust and validation.
Main Methods:
- A progressively-balanced resampling strategy combining instance-based and class-based sampling to create balanced training data.
- A novel neuron and normalized channel-spatial attention module (Neu-NCSAM) with 3-D weights and sparsity penalty for detailed feature learning, including small lesions.
- A weighted 'cost loss' function integrating Cost-Sensitive regularization and Gaussian label smoothing to penalize misclassifications and improve small class accuracy.
- Gradient-weighted Class Activation Mapping (Grad-CAM) for visual localization and interpretation of model predictions.
Main Results:
- The proposed network achieved superior performance on two public datasets compared to state-of-the-art methods.
- Key performance metrics included Kappa: 83.46%, BACC: 60.44%, MCC: 65.18%, F1: 63.69%, and mAUC: 92.26%.
- The attention mechanism effectively captured fine details of lesions, and Grad-CAM provided interpretable localization maps.
Conclusions:
- The developed resampling-based cost loss attention network offers a significant advancement in explainable and accurate diabetic retinopathy grading.
- The proposed methods effectively mitigate challenges associated with imbalanced data and small lesion detection.
- The model's explainability enhances clinical utility and trust in automated DR grading systems.
