Related Experiment Video
Updated: Jun 11, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.3K
A hybrid Transformer-CNN framework for uncertainty-guided semi-supervised multiclass eye disease classification with
Muhammad Hammad Malik1, Zishuo Wan1, Yingying Ren2
1School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China.
Summary
This study introduces a novel AI model for classifying eye diseases from fundus images, achieving high accuracy and interpretability. The approach enhances early diagnosis and treatment to prevent vision loss.
Area of Science:
- Ophthalmology and Medical Imaging
- Artificial Intelligence in Healthcare
- Deep Learning for Disease Classification
Background:
- Accurate classification of eye diseases like cataract, diabetic retinopathy (DR), and glaucoma from fundus images is crucial for preventing vision loss.
- Existing deep learning methods face challenges with large labeled datasets, inefficient unlabeled data utilization, and limited interpretability, hindering clinical application.
Purpose of the Study:
- To develop a novel CNN-Transformer hybrid architecture for enhanced multiclass eye disease classification.
- To improve the utilization of both labeled and unlabeled data through innovative semi-supervised learning (SSL).
- To enhance model interpretability for clinical validation and trust.
Main Methods:
- A hybrid CNN-Transformer architecture (ConvNeXt backbone with Transformer modules) using multi-head attention for spatial and long-range feature capture.
- Uncertainty-Guided MixMatch (UG-MixMatch) SSL framework employing Monte Carlo (MC) dropout for uncertainty quantification and pseudo-label refinement.
- Gradient-based Integrated Attention Map (GIAM) for interpretable predictions, aggregating attention maps with adaptive channel-wise weighting.
Main Results:
- Achieved 95.27% classification accuracy with UG-MixMatch and 95.51% with MC dropout on the Ocular Imaging Health (OIH) dataset.
- Demonstrated near-perfect agreement with ground truth (Cohen's kappa score of 93.70).
- Exceptional class-wise performance, including 100% sensitivity/specificity for DR and high specificity for cataract and glaucoma, with robust AUC values.
Conclusions:
- The proposed framework effectively addresses data scarcity and enhances interpretability in eye disease classification.
- The hybrid model delivers clinically relevant performance, offering a promising step towards scalable, explainable, and accurate diagnostic tools.
- GIAM visualizations provide enhanced clinical interpretability, validating model predictions for potential use in Clinical Decision Support Systems (CDSS).
