CLELNet: A continual learning network for esophageal lesion analysis on endoscopic images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Esophageal lesion diagnosis relies heavily on endoscopist expertise.
• Intelligent diagnosis systems can alleviate endoscopist workload.
• Incremental learning is crucial for expanding diagnostic capabilities.

Purpose of the Study:

• To develop a deep learning model for incremental learning in esophageal lesion diagnosis.
• To create a system capable of classifying and segmenting esophageal lesions.
• To reduce the burden on endoscopists through an AI-powered tool.

Main Methods:

• Proposed a continual learning-based esophageal lesion network (CLELNet).
• Employed a convolutional autoencoder for feature extraction.
• Implemented shared and task-specific layers for multi-task learning (classification and segmentation).

Main Results:

• Achieved 95.96% classification accuracy for task 1.
• Obtained 65.66% Intersection over Union (IoU) for segmentation (task 2).
• Reached 78.08% Dice Similarity Coefficient (DSC) for segmentation (task 2).

Conclusions:

• CLELNet enables task-incremental learning without catastrophic forgetting.
• The model serves as a valuable computer-aided diagnosis system for esophageal lesions.
• This approach supports the development of adaptable AI diagnostic tools.