Deep Learning-based Outcome Prediction in Progressive Fibrotic Lung Disease Using High-Resolution Computed Tomography

Rationale: Reliable outcome prediction in patients with fibrotic lung disease using baseline high-resolution computed tomography (HRCT) data remains challenging. Objectives: To evaluate the prognostic accuracy of a deep learning algorithm (SOFIA [Systematic Objective Fibrotic Imaging Analysis Algorithm]), trained and validated in the identification of usual interstitial pneumonia (UIP)-like features on HRCT (UIP probability), in a large cohort of well-characterized patients with progressive fibrotic lung disease drawn from a national registry. Methods: SOFIA and radiologist UIP probabilities were converted to Prospective Investigation of Pulmonary Embolism Diagnosis (PIOPED)-based UIP probability categories (UIP not included in the differential, 0-4%; low probability of UIP, 5-29%; intermediate probability of UIP, 30-69%; high probability of UIP, 70-94%; and pathognomonic for UIP, 95-100%), and their prognostic utility was assessed using Cox proportional hazards modeling. Measurements and Main Results: In multivariable analysis adjusting for age, sex, guideline-based radiologic diagnosis, and disease severity (using total interstitial lung disease [ILD] extent on HRCT, percent predicted FVC, DLCO, or the composite physiologic index), only SOFIA UIP probability PIOPED categories predicted survival. SOFIA-PIOPED UIP probability categories remained prognostically significant in patients considered indeterminate (n = 83) by expert radiologist consensus (hazard ratio, 1.73; P < 0.0001; 95% confidence interval, 1.40-2.14). In patients undergoing surgical lung biopsy (n = 86), after adjusting for guideline-based histologic pattern and total ILD extent on HRCT, only SOFIA-PIOPED probabilities were predictive of mortality (hazard ratio, 1.75; P < 0.0001; 95% confidence interval, 1.37-2.25). Conclusions: Deep learning-based UIP probability on HRCT provides enhanced outcome prediction in patients with progressive fibrotic lung disease when compared with expert radiologist evaluation or guideline-based histologic pattern. In principle, this tool may be useful in multidisciplinary characterization of fibrotic lung disease. The utility of this technology as a decision support system when ILD expertise is unavailable requires further investigation.