Class Imbalance Correction for Improved Universal Lesion Detection and Tagging in CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Radiologists use CT scans to detect and measure lesions for cancer staging.
• The DeepLesion dataset aids lesion detection algorithm development but has data imbalances.
• Existing algorithms struggle with missing measurements and imbalanced lesion categories.

Purpose of the Study:

• To address class imbalance in the DeepLesion dataset for improved lesion detection and tagging.
• To evaluate the impact of data balancing strategies on lesion detection model performance.
• To propose structured reporting guidelines for lesion findings in radiology reports.

Main Methods:

• Utilized a subset of the DeepLesion dataset (1331 lesions) to train a VFNet model.
• Implemented three data balancing strategies: by body part, by patient, and by lesion size.
• Compared performance against an unbalanced dataset and other models (FasterRCNN, RetinaNet, FoveaBox).

Main Results:

• Balancing body part labels improved sensitivity for lesions ≥ 1cm in underrepresented classes.
• Data balancing by lesion size enhanced recall across all classes for the VFNet model.
• Similar positive trends were observed with FasterRCNN, RetinaNet, and FoveaBox models.

Conclusions:

• Data-driven balancing strategies effectively mitigate class imbalance in lesion detection datasets.
• Improved lesion detection and tagging accuracy can aid radiologists in cancer staging and assessment.
• Structured reporting guidelines can enhance the clarity and consistency of radiology reports.