Late Breaking Abstract - A Deep Learning Algorithm for Classifying Fibrotic Lung Disease on High Resolution Computed Tomography

We investigated the utility of a deep learning algorithm for providing automated classification of fibrotic lung disease on HRCT according to criteria specified in two international diagnostic guideline statements; 1) the ATS/ERS/JRS.ALAT guidelines for diagnosis and management of IPF and 2) the Fleischner Society diagnostic criteria for IPF. We benchmarked algorithm performance against a cohort of 91 thoracic radiologists. 1157 HRCT studies showing evidence of a fibrotic lung disease from 2 institutions were used to train the algorithm which was based on Google's InceptionV2 neural network. Algorithm performance, reported as accuracy, prognostic accuracy and Cohen's kappa coefficient of interobserver agreement, was evaluated on a cohort of 150 HRCTs with fibrotic lung disease against the majority vote of ninety-one specialist thoracic radiologists drawn from multiple international thoracic imaging societies. The median accuracy of the thoracic radiologists was 70.7±0.09% while accuracy of the algorithm was 73.3%, outperforming 60/91 of the thoracic radiologists. The algorithm's categorisation of UIP vs not UIP provided equal prognostic discrimination that the majority opinion of the thoracic radiologists (HR 2.88, p<0.0001 95%CI 1.79-4.61 versus HR 2.74, p<0.0001 95%CI 1.67-4.48 respectively). For the Fleischner Society HRCT criteria for UIP, median interobserver agreement between the radiologists was moderate (k=0.56±0.03) but good between the algorithm and the radiologists (k=0.64±0.17). HRCT evaluation by a deep learning algorithm may provide low-cost, reproducible, near-instantaneous classification of fibrotic lung disease on HRCT with human-level accuracy.