• Jonathan H Chung, Lydia Chelala, Janelle Vu Pugashetti, Jennifer M Wang, Ayodeji Adegunsoye, Alexander W Matyga, Lauren Keith, Kai Ludwig, Sahar Zafari, Sahand Ghodrati, Ahmadreza Ghasemiesfe, Henry Guo, Eleanor Soo, Stephen Lyen, Charles Sayer, Charles Hatt, and Justin M Oldham.
• Department of Radiology, University of Chicago, Chicago, IL.
• Chest. 2024 Feb 1; 165 (2): 371380371-380.

BackgroundBecause chest CT scan has largely supplanted surgical lung biopsy for diagnosing most cases of interstitial lung disease (ILD), tools to standardize CT scan interpretation are urgently needed.Research QuestionDoes a deep learning (DL)-based classifier for usual interstitial pneumonia (UIP) derived using CT scan features accurately discriminate radiologist-determined visual UIP?Study Design And MethodsA retrospective cohort study was performed. Chest CT scans acquired in individuals with and without ILD were drawn from a variety of public and private data sources. Using radiologist-determined visual UIP as ground truth, a convolutional neural network was used to learn discrete CT scan features of UIP, with outputs used to predict the likelihood of UIP using a linear support vector machine. Test performance characteristics were assessed in an independent performance cohort and multicenter ILD clinical cohort. Transplant-free survival was compared between UIP classification approaches using the Kaplan-Meier estimator and Cox proportional hazards regression.ResultsA total of 2,907 chest CT scans were included in the training (n = 1,934), validation (n = 408), and performance (n = 565) data sets. The prevalence of radiologist-determined visual UIP was 12.4% and 37.1% in the performance and ILD clinical cohorts, respectively. The DL-based UIP classifier predicted visual UIP in the performance cohort with sensitivity and specificity of 93% and 86%, respectively, and in the multicenter ILD clinical cohort with 81% and 77%, respectively. DL-based and visual UIP classification similarly discriminated survival, and outcomes were consistent among cases with positive DL-based UIP classification irrespective of visual classification.InterpretationA DL-based classifier for UIP demonstrated good test performance across a wide range of UIP prevalence and similarly discriminated survival when compared with radiologist-determined UIP. This automated tool could efficiently screen for UIP in patients undergoing chest CT scan and identify a high-risk phenotype among those with known ILD.Copyright © 2023 American College of Chest Physicians. Published by Elsevier Inc. All rights reserved.

14 keywords...

hide…