Deep Learning-based Fibrosis Extent on Computed Tomography Predicts Outcome of Fibrosing Interstitial Lung Disease Independent of Visually Assessed Computed Tomography Pattern

Abstract

Rationale: Radiologic pattern has been shown to predict survival in patients with fibrosing interstitial lung disease. The additional prognostic value of fibrosis extent by quantitative computed tomography (CT) is unknown. Objectives: We hypothesized that fibrosis extent provides information beyond visually assessed CT pattern that is useful for outcome prediction. Methods: We performed a retrospective analysis of chest CT, demographics, longitudinal pulmonary function, and transplantation-free survival among participants in the Pulmonary Fibrosis Foundation Patient Registry. CT pattern was classified visually according to the 2018 usual interstitial pneumonia criteria. Extent of fibrosis was objectively quantified using data-driven textural analysis. We used Kaplan-Meier plots and Cox proportional hazards and linear mixed-effects models to evaluate the relationships between CT-derived metrics and outcomes. Results: Visual assessment and quantitative analysis were performed on 979 enrollment CT scans. Linear mixed-effect modeling showed that greater baseline fibrosis extent was significantly associated with the annual rate of decline in forced vital capacity. In multivariable models that included CT pattern and fibrosis extent, quantitative fibrosis extent was strongly associated with transplantation-free survival independent of CT pattern (hazard ratio, 1.04; 95% confidence interval, 1.04-1.05; P < 0.001; C statistic = 0.73). Conclusions: The extent of lung fibrosis by quantitative CT is a strong predictor of physiologic progression and survival, independent of visually assessed CT pattern.

Keywords: deep learning; interstitial lung disease; radiology; usual interstitial pneumonia.