Assessing body composition using auto-segmentations of muscle and subcutaneous adipose tissue in prostate cancer patients receiving magnetic resonance-guided radiotherapy

Background and purpose: Sarcopenia negatively impacts treatment outcomes in prostate cancer (PCa) patients. With the growing use of magnetic resonance-guided radiotherapy (MRgRT), automated body composition (BC) analysis using deep learning frameworks may allow for early detection of at-risk patients and intervention during treatment. Methods: T2-weighted images were obtained from 71 patients with intermediate-risk PCa who underwent MRgRT using one of three treatment regimens: androgen deprivation therapy (ADT) with 60 Gy in 20 fractions (20# + ADT), ADT with 36.25 Gy in 5 fractions (5# + ADT), or 36.25 Gy in 5 fractions without ADT (5# + noADT). An auto-segmentation model was developed using a 36/9 train–test split with 5-fold cross-validation, implemented with and without a bowel and bone (BB) mask, to delineate muscle and subcutaneous adipose tissue (SAT). Performance was assessed using mean surface distance, 95th percentile Hausdorff distance and Dice similarity coefficient. The model was then applied to quantify muscle, SAT, and intramuscular adipose tissue from images acquired during MRgRT. Results: Performance metrics were similar with and without the BB mask; however, using the mask resulted in a 24 % reduction in failures. At the start of MRgRT, SAT area differed significantly between 20# + ADT and 5# + noADT (p = 0.04), with significant changes across all ROIs observed by treatment end. Conclusion: Routinely acquired MRgRT images can be used for reliable auto-segmentation of BC metrics. Significant differences in BC metrics were found between treatment regimens. This supports their use for sarcopenia detection and adaptive treatment planning.