Related Experiment Video
Updated: Sep 13, 2025
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
2.6K
Correcting Class Imbalances with Self-Training for Improved Universal Lesion Detection and Tagging
Alexander Shieh1, Tejas Sudharshan Mathai1, Jianfei Liu1
1Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, Bethesda MD, USA.
Arxiv
|July 30, 2025
Summary
A novel self-training pipeline enhances universal lesion detection and tagging (ULDT) in CT scans by iteratively improving model accuracy. This method effectively addresses data limitations and class imbalances, boosting lesion detection sensitivity.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Universal Lesion Detection and Tagging (ULDT) is crucial for cancer assessment using CT studies.
- Existing datasets like DeepLesion have annotation gaps and class imbalances, limiting ULDT algorithm development.
- Accurate lesion detection and tracking are vital for monitoring tumor burden and treatment response.
Purpose of the Study:
- To develop an effective self-training pipeline for ULDT in CT studies.
- To overcome limitations of incomplete annotations and class imbalances in existing datasets.
- To improve the sensitivity and accuracy of lesion detection algorithms.
Main Methods:
- A VFNet model was initially trained on a small annotated subset of the DeepLesion dataset.
- A self-training approach was implemented, incorporating novel lesion candidates from larger unseen datasets over multiple rounds.
- Experiments explored different threshold policies and upsampling strategies to manage class imbalances and improve lesion selection quality.
Main Results:
- Direct self-training improved sensitivity for over-represented lesion classes but decreased it for under-represented ones.
- A combination of upsampling mined lesions and a variable threshold policy increased sensitivity by 6.5% (72% vs 78.5%) at 4 False Positives (FP) compared to standard self-training.
- This optimized approach showed an 11.7% sensitivity increase compared to self-training without upsampling.
- The method maintained or improved sensitivity at 4 FP across all 8 lesion classes.
Conclusions:
- The developed self-training pipeline effectively enhances ULDT performance in CT studies.
- Addressing class imbalance through upsampling and variable thresholds is critical for robust lesion detection.
- This approach offers a promising solution for improving tumor burden assessment and lesion progression tracking in clinical settings.
