Correcting Class Imbalances with Self-Training for Improved Universal Lesion Detection and Tagging
Alexander Shieh, Tejas Sudharshan Mathai, Jianfei Liu, Angshuman Paul, Ronald M. Summers·April 07, 2025
Summary
A self-training pipeline using VFNet on the DeepLesion dataset for CT lesion detection and tagging was developed. The model, initially trained on a subset, incorporated novel lesion candidates, improving sensitivity across classes. Upsampling and a variable threshold policy further enhanced results, outperforming direct self-training. The method, employing self-training, upsampling, and confidence-based filtering, was the first for unsupervised lesion detection and tagging.
Introduction
Background
Overview of CT lesion detection and tagging challenges
Importance of accurate lesion detection in medical imaging
Introduction to the DeepLesion dataset
Objective
Development of a self-training pipeline for unsupervised lesion detection and tagging
Utilization of VFNet for lesion detection
Improvement of sensitivity across lesion classes through novel candidate incorporation
Method
Data Collection
Source of the DeepLesion dataset
Preprocessing steps for the dataset
Data Preprocessing
Techniques applied to the dataset for model training
Model Training
Introduction to VFNet architecture
Training process on the subset of the DeepLesion dataset
Novel Lesion Candidate Incorporation
Methodology for identifying and incorporating new lesion candidates
Upsampling
Importance and implementation of upsampling in the pipeline
Variable Threshold Policy
Description of the threshold policy and its role in enhancing detection results
Confidence-Based Filtering
Explanation of the filtering process based on model confidence
Results
Performance Evaluation
Metrics used for assessing the pipeline's effectiveness
Comparison with direct self-training methods
Sensitivity Improvement
Analysis of sensitivity across different lesion classes
Quantitative and Qualitative Results
Presentation of results in terms of accuracy, precision, recall, and F1-score
Conclusion
Unsupervised Lesion Detection and Tagging
Summary of the pipeline's contributions to the field
Future Work
Potential areas for further research and development
Practical Implications
Real-world applications and benefits of the developed pipeline
Basic info
papers
computer vision and pattern recognition
artificial intelligence
Advanced features