Volumetric assessment of body composition from cardiac SPECT CT attenuation maps and association of sarcopenia measures with mortality

Jirong Yi, Anna Michalowska, Aakash Dhananjay Shanbhag, Robert Miller, Wenhao Zhang, Aditya Killekar, Nipun Manral, Mark Lemley, Jianhang Zhou, Bryan Bednarski, Joanna Liang, Valerie Builoff, Terrence Ruddy, Andrew Einstein, Attila Feher, Edward Miller, Albert Sinusas, Paul Kavanagh, Damini Dey, Daniel Berman et al. (1 additional authors not shown)

Abstract:

Introduction: Sarcopenia is well-established as an important clinical factor, and its quantification is typically performed using L3 slice from abdominal CT. We developed a novel method combining artificial intelligence (AI) and image processing to automatically and volumetrically analyze its quantification from SPECT/CT attenuation maps (CTAC). We aimed to evaluate the prognostic value for sarcopenia measurements to predict all-cause mortality (ACM) during follow-up.

Methods: The study involved 10826 patients [male 5989 (55.32%)] from the multicenter REFINE SPECT registry from 4 different institutes with available cardiac SPECT/CT scans [kVp 120-130] with over 2-year follow-up for mortality. We utilized a previously validated deep learning CT segmentation model (TotalSegmentor v2) combined with image processing steps to automatically outline the rib cage which surrounds the lungs and the heart, then threshold fat and muscle within these structures (Hounsfield Unit [HU] [-29, 150] for skeletal muscle (SM), and [-190,-30] for adipose tissue), and segment 5 body tissues (SM, bone, subcutaneous adipose tissue (SAT), intramuscular adipose tissue (IMAT), and visceral adipose tissue (VAT)) on CTAC. We then evaluated the prognostic value of SM indexed volume (SM volume/height2 [cm3/m2]). We applied a percentile-based grid search to establish sex-specific thresholds for risk stratification for the proposed SM measure. Kaplan-Meier curves and unadjusted Cox proportional hazard ratios (HRs) were used to evaluate associations with ACM. The log-rank test was used to evaluate statistical significance.

Results: The age (median, Interquartile range [IQR]) was 65 [57-73] years. During a follow-up of 870 [522-1296] days, 667 (6.2%) patients died. Total computing was 110 seconds per case (64 sec for deep learning segmentation NVIDIA RTX Titan GPU for and 46 sec for image processing with AMD 3960X CPU). The indexed volumes in population (median [IQR] in cm3/m2) were 1005.8 [798.5-1249.4] for SM, 311.2 [264.3-363.1] for bone, 268.2 [165.7-394.3] for VAT, 93.4 [68.7-130.7] for IMAT, and 1701.5 [ 1098.7-2545.2] for SAT. The SM thresholds were established at 681.8 cm3/m2 for male (5th percentile in male) and 507.6 cm3/m2 for female (5th percentile in female). The presence of low SM indexed volume was associated with an increased risk of death, i.e., HR: 2.51 [95% CI: 1.88,3.34], p<0.0001.

Conclusions: Automatic volumetric body composition including sarcopenia measures can be obtained from cardiac SPECT/CT attenuation maps. Chest skeletal muscle characteristics improve mortality risk assessment in cardiovascular patients undergoing myocardial perfusion imaging, without additional scanning or physician interaction.