Deep-Learning Algorithm Diagnostic Support for Usual Interstitial Pneumonia Pattern Recognition in Fibrotic...

High resolution computed tomography (HRCT) scan diagnostic classification for usual interstitial pneumonia (UIP) plays a critical role in therapeutic decision-making and clinical trial eligibility for interstitial lung disease (ILD) patients, but is subject to variability. A deep learning algorithm, the Systematic Objective Fibrotic Imaging Analysis Algorithm (SOFIA), has been validated to assist classification of HRCTs based on current guidelines. In this study, we evaluate the impact of SOFIA on inter-observer agreement for UIP classification and prognostic accuracy of clinicians' assessment of ILD HRCTs.

Radiologists and pulmonologists (reviewers) were invited to evaluate 203 HRCTs from a national fibrotic ILD registry, scoring each of four UIP categories (definite UIP, probable UIP, indeterminate, or alternative diagnosis). SOFIA outputs were then provided, and reviewers were able to reevaluate their scores. Changes in interobserver agreement for UIP classification and prognostic accuracy were calculated.

Three hundred twelve reviewers (120 radiologists, 192 pulmonologists) from 49 countries evaluated 203 HRCT scans. Following SOFIA, inter-observer diagnostic agreement improved for definite UIP from moderate to good (ICC_pre = 0.54[0.50-0.60]; ICC_post = 0.70[0.66-0.74]), and for probable UIP from fair to moderate (ICC_pre = 0.30[0.27-0.35]; ICC_post = 0.53[0.49-0.58]). Following SOFIA, there was improved prognostic accuracy for reviewers' definite UIP, probable UIP, and indeterminate scores (significant change in c-index), and the proportion of reviewers whose probable UIP scores were significantly predictive of transplant-free survival increased by 42%.

Providing SOFIA algorithm output to clinicians reviewing HRCT scans improved diagnostic agreement and prognostic accuracy for fibrotic ILD. SOFIA may be a useful automated assistive tool to support improved diagnostic consistency.