Deep Learning Based Automated Platform for Sarcopenia Assessment and Outcomes Analysis in Head and Neck Cancer

Overview

Women’s Hospital The body composition status, which characterizes the levels of muscle and adipose tissue in a person’s body, is closely linked to cancer treatment risks and survival. Sarcopenia, describing muscle depletion, is a well-studied prognostic factor in patients diagnosed with head and neck cancers (HNC). Sarcopenia is typically assessed by measuring skeletal muscle index (SMI) derived from muscle segmentation of Computer Tomography (CT) imaging. However, manual segmentation is time-consuming, prone to human variability, and not practical for routine clinical use. In this study, a fully automated deep learning (DL) based platform was developed to accurately segment the skeletal muscle and adipose tissue at the third cervical vertebrae level (C3) from CT scans. This platform enables precise sarcopenia measurement and an evaluation of its relationships with overall -survival and treatment induced toxicity outcomes. A multi - institutional study was conducted using de -identified data from patients undergoing primary radia tion therapy for HNC at three major North American comprehensive cancer centers. Median Dice Similarity Coefficient (DSC), which measures pixel-wide agreement with ground truth, was 0.91 for predicted skeletal muscle segmentations and 0.86 for adipose on t he internal test set, with a 95.5% acceptable rate on external validation testing, indicating excellent predictive ability and generalizability of DL models. Predicted SMI values were highly correlated with manually annotated values, with Pearson r = 0.98 (p < 0.0001) for patients across datasets. In multivariable Cox -regression analysis (n=342), SMI -derived sarcopenia was associated with worse survival and longer PEG tube duration. This platform can integrate sarcopenia assessment into clinical treatment decision-making for individuals with HNC, ultimately leading to improved outcomes.