Deep learning classification of radiologic pattern is associated with progression of interstitial lung abnormalities and with survival

Abstract Rationale Interstitial lung abnormalities (ILA) are common, but not all progress, highlighting the need for objective computed tomography (CT)-based biomarkers for risk stratification. Objective To determine whether deep learning-based classification of usual interstitial pneumonia (UIP) is associated with fibrosis progression and survival in individuals with ILA. Methods Baseline and follow-up CT scans from participants without clinically diagnosed interstitial lung disease in two large observational cohorts (COPDGene and AGES-Reykjavík) were analyzed using data-driven textural analysis (DTA) to quantify fibrosis and a deep learning-based classifier (MIL-UIP) to estimate UIP likelihood. Associations of baseline MIL-UIP with DTA trajectory and survival were evaluated using linear mixed and multivariable Cox models, respectively. Measurements and Main Results Baseline MIL-UIP > 0.5 was associated with relative annual DTA increases of 13.41% (95% CI: 8.10%, 18.99%; p < 0.001) and 14.10% (95% CI: 5.06%, 23.91%; p = 0.002) in COPDGene and AGES-Reykjavík, respectively. Each 0.1-point increase in MIL-UIP was associated with relative rates of DTA change that were 1.10 percentage points higher (95% CI: 0.45, 1.74; p < 0.001) in COPDGene and 1.30 percentage points higher (95% CI: 0.17, 2.45; p = 0.025) in AGES-Reykjavík. MIL-UIP > 0.5 was associated with higher mortality in COPDGene (HR 1.61; 95% CI: 1.08, 2.38; p = 0.018) and AGES-Reykjavík (HR 1.62; 95% CI: 1.01, 2.61; p = 0.046). Baseline DTA and MIL-UIP were highly associated with visual assessments of ILA, UIP, and fibrosis progression. Conclusion Automated assessment of UIP-like CT features is associated with fibrosis progression and mortality in individuals with ILA, supporting its potential use for risk stratification and clinical trials enrichment.