Morphomic phenotyping of oncologic populations.

Abstract

e13572 Background: We have previously identified body and organ composition measures in pre-treatment CT scans that predict post-treatment mortality or disease recurrence in melanoma as well as esophageal, pancreatic, hepatocellular and lung cancers within single institution cohorts in the US. Similar trends can be identified in Asian cohorts although the absolute range of these body composition measures differ substantially from those in the US. We have also identified body and organ composition measures that predict drug tissue distribution and optimal dosing. The qualitative and relative nature of body and organ composition measures to date have limited their utility for precision treatment of oncologic populations. We have therefore refined the precision, granularity and reliability of these measures (morphomics) in oncologic and reference populations. Methods: Analytic Morphomics, our group’s imaging analysis platform, extracts granular, anatomically-indexed body and organ composition measures from medical imaging. This platform has been updated to use artificial intelligence and nested, supervised machine learning algorithms. Our digital morphomics “assays” are 100% reproducible due to our extensive use of deterministic software algorithms and scalable due to our use of efficient ML algorithms and software code. Our ML models have been trained using state-of-the art techniques in 2D and 3D convolutional neural networks (CNNs) for image segmentation, utilizing repeated randomization cross-validation, hold-out sets, and ensemble modeling methods, with large, broadly representable training image sets, and verified by extensive user testing on clinical scans from nearly 200,000 live subjects. Results: We analyzed curated cohorts reflecting the general United States and Taiwanese population (Ns=5,286; 1,459). In addition, we analyzed oncological cohorts reflecting patients with colon (N=95), ovarian (216), pancreatic (156), and esophageal cancers (Ns=333, 344). Validated morphomic measures include bone mineral density (BMD), visceral fat area (VAT, subcutaneous fat area (SAT), skeletal muscle area (SMA), skeletal muscle density (MD) at multiple vertebral levels. When referenced against measures taken from general non-oncologic (trauma) populations to adjust for non-specific changes due to age and gender, these referenced morphomic measures gain predictive power for mortality and cancer recurrence. Conclusions: The unprecedented level of granularity and precision of these raw and referenced morphomic measures shows clinical utility for improved patient selection, drug dosing as well as evaluation of treatment response and toxicity.