Deep-learning pipeline for automated skeletal muscle segmentation and sarcopenia detection

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Anatomy

Background:

• Sarcopenia, progressive skeletal muscle loss, is linked to poor outcomes in various diseases.
• Traditional computed tomography (CT) based sarcopenia assessment is manual, time-consuming, and variable.

Purpose of the Study:

• To validate a deep learning (DL) pipeline for accurate and reproducible sarcopenia detection using CT.
• To assess DL performance across diverse abdominal conditions and imaging protocols.

Main Methods:

• Utilized the SAROS CT dataset for training and diverse multi-center datasets for testing.
• Integrated TotalSegmentator and nnU-Net for L3 vertebral and skeletal muscle segmentation.
• Evaluated performance using Dice scores, IoU, and diagnostic accuracy metrics for sarcopenia detection.

Main Results:

• DL pipeline showed consistent segmentation accuracy (Dice 0.9287-0.9701, IoU up to 0.9423).
• Expert evaluation confirmed reliable segmentation (78%-92.6%).
• Sarcopenia detection was consistent (sensitivity 0.94-0.97, specificity 0.84-0.97, AUC up to 0.92) across varied conditions.

Conclusions:

• A deep learning pipeline provides consistent and reliable skeletal muscle segmentation and sarcopenia detection.
• The DL approach is effective across heterogeneous abdominal CT protocols and diverse clinical conditions.