Related Experiment Video
Updated: Jan 10, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Assessing deep learning models for multi-class upper endoscopic disease segmentation: A comprehensive comparative
In Neng Chan1, Pak Kin Wong2, Tao Yan3
1Department of Electromechanical Engineering, University of Macau, Macau 999078, China.
World Journal of Gastroenterology
|November 20, 2025
Summary
Deep learning models show promise for segmenting upper gastrointestinal diseases, with hierarchical architectures like Swin-UMamba and SegFormer demonstrating high accuracy. Further clinical validation is essential for real-world application in endoscopy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Upper gastrointestinal (UGI) diseases pose diagnostic challenges in endoscopy due to visual similarities and observer variability.
- Automated segmentation using deep learning (DL) can aid endoscopists, but multi-class UGI disease segmentation is underexplored.
- Clinical validation of DL models for UGI disease segmentation is limited, hindering real-world application.
Purpose of the Study:
- To evaluate 17 state-of-the-art deep learning (DL) models for multi-class upper gastrointestinal (UGI) disease segmentation.
- To assess the clinical translation and real-world applicability of different DL architectures for UGI endoscopy.
- To identify DL models that can reduce diagnostic errors and support clinical decision-making.
Main Methods:
- Evaluated 17 DL models (CNN, transformer, mamba-based) on a self-collected dataset (3313 images, 9 classes) and the EDD2020 dataset (386 images, 5 classes).
- Assessed segmentation performance (IoU) and the performance-efficiency trade-off for each model.
- Conducted statistical analyses and cross-dataset evaluations to measure performance differences and generalization capabilities.
Main Results:
- Swin-UMamba achieved the highest segmentation performance (IoU: 89.06% self-collected, 77.53% EDD2020), followed by SegFormer and ConvNeXt + UPerNet.
- SegFormer offered the best accuracy-efficiency balance (92.02%), suitable for real-time clinical use.
- Transformer-based models generalized better (64.78%-71.52% retention) in cross-dataset evaluations, despite overall performance drops.
Conclusions:
- Hierarchical DL architectures like Swin-UMamba and SegFormer show significant potential for UGI disease segmentation.
- These models can help reduce missed diagnoses and improve endoscopic workflow efficiency.
- Robust clinical validation is critical before widespread real-world deployment in endoscopic practice.
