A Fully Automated Deep Learning Pipeline for Multi–Vertebral Level Quantification and Characterization of Muscle and Adipose Tissue on Chest CT Scans

• DOI: 10.1148/ryai.210080
• Publication Date: 2022-01-05T14:50:45Z
• Authors: Christopher P. Br...; Till D. Best; Maria M. Wrobel; J. Peter Marquardt; Kirti Magudia; Cylen Javidan; Jonathan H. Chung; Jayashree Kalpath...; Katherine P. Andr...; Florian J. Fintel...
• Keywords: Hounsfield scale; Interquartile range

Body composition on chest CT scans encompasses a set of important imaging biomarkers. This study developed and validated fully automated analysis pipeline for multi-vertebral level assessment muscle adipose tissue routine scans. retrospectively trained two convolutional neural networks 629 from patients (55% women; mean age, 67 years ± 10 [standard deviation]) obtained between 2014 2017 prior to lobectomy primary lung cancer at three institutions. A slice-selection network was identify an axial image the fifth, eighth, 10th thoracic vertebral bodies. segmentation (U-Net) segment image. Radiologist-guided manual-level selection generated ground truth. The authors then assessed predictive performance their approach cross-sectional area (CSA) (in centimeters squared) attenuation Hounsfield units) independent test set. For pipeline, median absolute error intraclass correlation coefficients both tissues were 3.6% (interquartile range, 1.3%-7.0%) 0.959-0.998 CSA 1.0 HU 0.0-2.0 HU) 0.95-0.99 attenuation. demonstrates accurate reliable quantification characterization Keywords: Skeletal Muscle, Adipose Tissue, CT, Chest, Composition Analysis, Convolutional Neural Network (CNN), Supervised Learning Supplemental material is available this article. © RSNA, 2022.