Deep learning–based quantitative CT assessment of interstitial lung abnormalities: prognostic risk thresholds in a health screening population

This retrospective cohort study investigated quantitative prognostic thresholds for interstitial lung abnormalities (ILA) in a health screening population. A total of 3,363 asymptomatic individuals aged ≥40 who underwent chest CT from 2007 to 2013 were followed for approximately 10 years to assess associations between ILA extent and long-term outcomes including all-cause mortality, ILD diagnosis, and lung cancer. ILA extents (total ILA% and fibrotic ILA%) were quantified across the whole lung using AVIEW Lung Texture ILA (version 1.1.39.14; Coreline Soft), a commercially available deep learning-based software that employs GAN-based image conversion to harmonize across different CT protocols and vendors. Two radiologists independently reviewed baseline CT scans, and optimal thresholds were determined using the minimum P-value method. Optimal cutoffs for all-cause mortality were 2.89% for total ILA (HR 5.15; P<0.001) and 0.26% for fibrotic ILA (HR 2.71; P<0.001). Deep learning-based quantitative ILA assessment was independently associated with long-term mortality, with practical prognostic thresholds of ~3% for total ILA and ~0.3% for fibrotic ILA in a health screening population.