An Artificial Intelligence–Based Chest X-ray Model on Human Nodule Detection Accuracy From a Multicenter Study

• DOI: 10.1001/jamanetworkopen.2021.41096
• Publication Date: 2021-12-29T17:01:49Z
• Authors: Fatemeh Homayounieh; Subba Digumarthy; Shadi Ebrahimian; Johannes Rueckel; Boj Friedrich Hoppe; Bastian Oliver Sabel; Sailesh Conjeti; Karsten Ridder; Markus Sistermanns; Lei Wang; Alexander Preuhs; Florin Ghesu; Awais Mansoor; Mateen Moghbel; Ariel Botwin; Ramandeep Singh; Samuel Cartmell; John Patti; Christian Huemmer; Andreas Fieselmann; Clemens Joerger; Negar Mirshahzadeh; Victorine Muse; Mannudeep Kalra
• Keywords: Chest radiograph; Atelectasis; Solitary pulmonary nodule

<h3>Importance</h3> Most early lung cancers present as pulmonary nodules on imaging, but these can be easily missed chest radiographs. <h3>Objective</h3> To assess if a novel artificial intelligence (AI) algorithm help detect radiographs at different levels of detection difficulty. <h3>Design, Setting, and Participants</h3> This diagnostic study included 100 posteroanterior radiograph images taken between 2000 2010 adult patients from an ambulatory health care center in Germany image database the US. Included were selected to represent with difficulties (from easy difficult), comprised both normal nonnormal control. <h3>Exposures</h3> All processed AI algorithm, Rad Companion Chest X-ray. Two thoracic radiologists established ground truth 9 test US independently reviewed all 2 sessions (unaided AI-aided mode) least 1-month washout period. <h3>Main Outcomes Measures</h3> Each radiologist recorded presence 5 findings (pulmonary nodules, atelectasis, consolidation, pneumothorax, pleural effusion) their level confidence for detecting individual finding scale 1 10 (1 representing lowest confidence; 10, highest confidence). The analyzed metrics sensitivity, specificity, accuracy, receiver operating characteristics curve area under (AUC). <h3>Results</h3> Images included, mean (SD) age 55 (20) years including 64 men 36 women. Mean accuracy across improved by 6.4% (95% CI, 2.3% 10.6%) interpretation compared unaided interpretation. Partial AUCs within effective interval range 0 0.2 false positive rate 5.6% −1.4% 12.0%) Junior saw greater improvement sensitivity nodule senior counterparts (12%; 95% 4% 19% vs 9%; 1% 17%) while experienced similar specificity (4%; −2% 9%) junior −3% 5%). <h3>Conclusions Relevance</h3> In this study, was associated difficulty readers experience.