3D Universal Lesion Detection and Tagging in CT with Self-Training

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Radiologists face challenges in lesion localization, classification, and measurement in CT studies.
• Universal lesion detection and tagging (ULDT) aims to streamline these tasks and enable tumor burden assessment.
• Existing ULDT methods struggle with incomplete volumetric data and class imbalance in datasets like DeepLesion.

Purpose of the Study:

• To develop a self-training pipeline for detecting and tagging 3D lesions based on their body part location.
• To overcome limitations of existing datasets and methods in 3D lesion analysis.

Main Methods:

• A VFNet model was trained on a 30% subset of the DeepLesion dataset for 2D lesion detection and tagging.
• A self-training approach was employed, expanding 2D lesion context to 3D and iteratively retraining the model.
• The pipeline integrates mined 3D lesion proposals back into the training data over multiple rounds.

Main Results:

• The VFNet model achieved an average sensitivity of 46.9% at [0.125:8] false positives using only 30% of the data.
• Performance is comparable to existing approaches that utilized the entire DeepLesion dataset.
• The model successfully detected lesions in 3D and tagged them by body part.

Conclusions:

• Self-training is an effective strategy for 3D lesion detection and tagging, even with limited data.
• This approach can improve efficiency and accuracy in radiological assessments.
• This work represents the first joint 3D lesion detection and body part tagging method.