Related Experiment Video
Updated: Sep 8, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
A multi-task deep learning pipeline integrating vessel segmentation and radiomics for multiclass retinal disease
Feng Yan1, Yanxia Liu2, Qingsong Zhao3
1Department of Ophthalmology, Ningcheng Central Hospital, Chifeng City, Inner Mongolia Autonomous Region 024200, China.
Photodiagnosis and Photodynamic Therapy
|September 6, 2025
Summary
This study developed a deep learning framework for classifying diabetic retinopathy, hypertensive retinopathy, and papilledema from fundus images. The model achieved high accuracy, demonstrating potential for automated retinal disease diagnosis.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Accurate classification of retinal diseases like diabetic retinopathy (DR), hypertensive retinopathy (HR), and papilledema is crucial for timely intervention.
- Automated analysis of fundus images can aid in early detection and management of these conditions.
Purpose of the Study:
- To develop a multi-task deep learning framework integrating vessel segmentation and radiomic analysis for automated classification of four retinal conditions.
- To evaluate the performance of transformer-based segmentation models and radiomic features for disease classification.
Main Methods:
- Utilized a dataset of 2,165 patients from eight medical centers, with fundus images undergoing standardized preprocessing.
- Employed five deep learning models (U-Net, Attention U-Net, DeepLabV3+, HRNet, Swin-Unet) for whole vessel and artery-vein segmentation.
- Extracted 220 radiomic features and computed arteriovenous ratio (AVR), followed by feature selection and classification using XGBoost, CatBoost, RF, and Ensemble models.
Main Results:
- Swin-Unet achieved superior segmentation performance with external Dice Similarity Coefficient (DSC) of 92.4% (whole vessel) and 89.8% (artery-vein).
- The LASSO-Ensemble classifier combination yielded test accuracy of 93.7%, external test accuracy of 92.3%, and AUC of 95.2%.
- AVR estimates aligned with clinical expectations and significantly contributed to class discrimination.
Conclusions:
- The developed multi-task pipeline effectively combines transformer-based segmentation with radiomics for accurate and interpretable retinal disease classification.
- The framework demonstrates strong generalizability, showing significant potential for future clinical applications in automated retinal disease diagnosis.
