Detection of Interstitial Lung Abnormalities on Chest Radiographs: Diagnostic Performance of Radiologists and Artificial Intelligence.
Authors
Affiliations (11)
Affiliations (11)
- Department of Radiology, Seoul National University Hospital and Seoul National University College of Medicine, 101 Daehak-ro, Jongno-gu, Seoul 03080, Korea.
- Soombit.ai, 21 Pangyo-ro 255-gil, Bundanggu, Seongnam 13486, Korea.
- Department of Radiology and Research Institute of Radiology, University of Ulsan College of Medicine, Asan Medical Center, 88 Olympic-ro 43-gil, Songpa-gu, Seoul 05505, Korea.
- Department of Radiology, Research Institute of Clinical Medicine of Jeonbuk National University-Biomedical Research Institute of Jeonbuk National University Hospital, 20 Geonji-ro, Deokjin-gu, Jeonju 54907, Korea.
- Keimyung Unversity Dongsan Medical Center, 1035 Dalgubeol-daero, Dalseo-gu, Daegu 42601, Korea.
- Department of Radiology, Seoul Metropolitan Government-Seoul National University Boramae Medical Center, Seoul National University College of Medicine, 20 Boramae-ro 5-gil, Dongjak-gu, Seoul 07061, Korea.
- Department of Radiology, Chonnam National University Hwasun Hospital, Chonnam National University Medical School, 264 Seoyang-ro, Hwasuneup, Hwasun-gun, Jeollanam-do 58128, Korea.
- Department of Radiology, Chonnam National University Hospital, Chonnam National University Medical School, 42 Jebong-ro, Donggu, Gwangju 61469, Korea.
- Department of Medicine, Weill Cornell Medical College, 1305 York Ave, New York, NY 10021, USA.
- Department of Medicine and Centre for Heart Lung Innovation, University of British Columbia and St. Paul's Hospital, 1081 Burrard Street, Vancouver, British Columbia V6Z 1Y6, Canada.
- Department of Radiology, Pusan National University Hospital and Biomedical Research Institute, Pusan National University School of Medicine, 179 Gudeok-ro, Seo-gu, Busan, Korea.
Abstract
<b>Background:</b> Interstitial lung abnormalities (ILA) on chest CT are receiving growing attention given their association with progression to interstitial lung disease. Radiography's role in ILA detection is not well described. <b>Objective:</b> To evaluate the diagnostic performance of radiologists and an artificial intelligence (AI) model in detecting ILA on chest radiographs using CT as the reference. <b>Methods:</b> This retrospective study included adults who underwent both chest CT and chest radiography as part of health check-up programs at two institutions in Korea between January 2007 and December 2010. Five thoracic radiologists independently assessed ILA likelihood on radiographs using a 5-point Likert scale (positive, ≥4). A previously developed AI model (AIRead-CXR; Soombit.ai) processed radiographs to generate a probability (0 to 1) of reticular or reticulonodular opacities (positive, ≥0.4). CT served as the reference standard for fibrotic and nonfibrotic ILA. Radiologists' diagnostic performance for ILA detection was reported using mean performance metrics and compared with AI performance using generalized estimating equations. Associations of AI-based radiographic ILA, adjusting for age, sex, and smoking status, were assessed with all-cause mortality and respiratory disease-related mortality using Cox proportional hazard and Fine-Gray competing-risk regression models. <b>Results:</b> The analysis included 1168 individuals (median age, 56 years; 786 male, 382 female). Forty-one individuals had ILA on CT (fibrotic, 22; nonfibrotic, 19). For fibrotic ILA, radiologists and AI had AUC of 0.86 and 0.92 (P=.06), sensitivity of 62.7% and 68.2% (P=.43), specificity of 97.8% and 98.7% (P=.05), and accuracy of 97.2% and 98.1% (P=.04), respectively. For fibrotic or nonfibrotic ILA, radiologists and AI had AUC of 0.75 and 0.83 (P=.009), sensitivity of 38.5% and 41.5% (P=.48), specificity of 98.0% and 98.8% (P=.05), and accuracy of 95.9% and 96.8% (P=.03), respectively. During a median follow-up of 11.9 years, radiographic ILA was independently associated with respiratory disease-mortality (adjusted subdistribution HR, 8.72; P<.001) but not overall mortality (adjusted HR, 1.75; P=.17). <b>Conclusion:</b> Radiologists and AI achieved suboptimal sensitivity for ILA detection on radiography, albeit high specificity. <b>Clinical Impact:</b> Despite association of radiographic ILA with a clinically relevant outcome, the findings do not support radiographic screening for ILA, whether incorporating radiologist or AI interpretation.