Related Experiment Video
Updated: May 5, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.6K
Dual-SwinOrd: A Dual-Head Swin Transformer with Semantic Prior Injection for Ordinal Diabetic Retinopathy Grading
Wenjuan Yu1, Xiaonan Si2, Jingxiang Zhong1,3
1The First Affiliated Hospital of Jinan University, Guangzhou 510630, China.
Bioengineering (Basel, Switzerland)
|May 4, 2026
Summary
Dual-SwinOrd, a novel deep learning framework, enhances automated diabetic retinopathy (DR) grading by integrating Vision Transformers and semantic guidance. This approach improves diagnostic accuracy and severity prediction for better vision loss prevention.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Diabetic retinopathy (DR) is a leading cause of irreversible vision loss in working adults.
- Automated DR grading systems are crucial for early detection and intervention.
- Current deep learning models struggle with long-range dependencies and clinical relevance.
Purpose of the Study:
- To develop an advanced deep learning framework, Dual-SwinOrd, for more accurate and clinically relevant automated diabetic retinopathy grading.
- To address limitations of existing models in capturing global retinal structures and semantic understanding.
- To balance classification accuracy with severity-consistent predictions.
Main Methods:
- Proposed Dual-SwinOrd framework integrating a Swin Transformer backbone for hierarchical feature extraction.
- Incorporated Progressive Lesion-aware Kernel Attention (PLKA) and Semantic Prior Modulation (SPM) modules using PubMedCLIP for medical linguistic guidance.
- Implemented a Dual-Head learning strategy with parallel Classification and Ordinal Regression Heads to optimize accuracy and rank-consistency.
Main Results:
- Dual-SwinOrd achieved state-of-the-art performance on benchmark datasets.
- Achieved 87.98% accuracy and 0.9370 QWK on the APTOS 2019 dataset.
- Achieved 86.54% accuracy and 0.9040 QWK on the DDR dataset.
Conclusions:
- The Dual-SwinOrd framework effectively improves automated diabetic retinopathy grading.
- The novel approach enhances both diagnostic accuracy and severity prediction consistency.
- This method offers a promising solution for early detection and management of DR.
