Evaluating Skellytour for Automated Skeleton Segmentation from Whole-Body CT Images.

Area of Science:

• Medical Imaging and Informatics
• Machine Learning in Radiology
• Skeletal Imaging Analysis

Background:

• Accurate bone segmentation in CT imaging is crucial for diagnosing and monitoring skeletal conditions.
• Existing segmentation models may have limitations in accuracy and generalizability across diverse datasets.

Purpose of the Study:

• To develop and evaluate Skellytour, a novel machine learning model for precise bone segmentation and subsegmentation using whole-body CT images.
• To benchmark Skellytour's performance against established models like TotalSegmentator.

Main Methods:

• Retrospective analysis of 90 whole-body CT scans from multiple myeloma patients, with manual segmentation into 60 labels, including cortical and trabecular bone.
• Assessment of segmentation performance using Dice Similarity Coefficient (DSC) and Normalized Surface Distance (NSD) on internal and external test datasets (362 scans).
• Evaluation of factors such as isotropy, resolution, labeling schemes, and postprocessing on model performance.

Main Results:

• Skellytour demonstrated high and consistent segmentation performance (DSC: 0.94-0.96, NSD: 0.99-1.0) across internal and external datasets.
• The model outperformed TotalSegmentator on two external datasets, indicating superior generalizability.
• High subsegmentation performance (DSC: 0.95, NSD: 0.995) was achieved, with detailed segmentations even in low-density bone regions.

Conclusions:

• Skellytour is an accurate and generalizable machine learning model for bone segmentation and subsegmentation in CT data.
• The model's performance and detailed output offer significant potential for skeletal imaging analysis.
• Skellytour is publicly available as a Python package on GitHub.