Prognostic value of AI-derived probability of interstitial lung abnormalities and interstitial lung disease in patients with esophageal cancer.
Authors
Affiliations (5)
Affiliations (5)
- Department of Diagnostic and Interventional Radiology, Graduate School of Medicine, The University of Osaka, 2-2 Yamadaoka, Suita, Osaka, 565-0871, Japan. [email protected].
- Department of Diagnostic and Interventional Radiology, Graduate School of Medicine, The University of Osaka, 2-2 Yamadaoka, Suita, Osaka, 565-0871, Japan.
- Canon Inc, 30-2 Shimomaruko 3-chome, Ohta-ku, Tokyo, 146-8501, Japan.
- Artificial Intelligence in Diagnostic Radiology, Graduate School of Medicine, The University of Osaka, 2-2 Yamadaoka, Suita, Osaka, 565-0871, Japan.
- Center for Pulmonary Functional Imaging, Department of Radiology, Brigham and Women's Hospital, Harvard Medical School, 75 Francis Street, Boston, MA, 02115, USA.
Abstract
To investigate the association between AI-derived probability of interstitial lung abnormalities (ILA), including interstitial lung disease (ILD), and prognosis in an esophageal cancer cohort re-evaluated according to the 2025 American Thoracic Society (ATS) Clinical Statement. This retrospective study included patients with esophageal cancer who underwent pretreatment chest CT between January 2011 and December 2015. Chest CT images were visually assessed and classified as non-ILA, indeterminate for ILA, ILA, or ILD according to the 2025 ATS Clinical Statement. Twelve previously developed artificial intelligence (AI) models generated patient-level ILA probability scores ranging from 0 to 1. Associations between AI-derived ILA/ILD probability and five-year overall survival (OS) were evaluated using multivariable Cox proportional hazards models adjusted for age, sex, body mass index, smoking history, and histology, stratified by clinical stage. Associations between visual ILA/ILD classification and OS were also assessed. A total of 452 patients (median age, 68 years [interquartile range: 61-73 years]; 61 women) were included. Among them, 251 (56%) were classified as non-ILA, 118 (26%) as indeterminate for ILA, 75 (17%) as ILA, and 8 (2%) as ILD. All 12 AI models showed significant associations with shorter OS (all p < 0.05). The best-fit model yielded a hazard ratio (HR) for mortality of 7.18 (95% confidence interval [CI]: 2.70-19.1; p < 0.001). Visual ILA and ILD were associated with mortality (ILA: HR = 1.60, 95% CI: 1.02-2.51, p = 0.039; ILD: HR = 3.01, 95% CI: 1.07-8.44, p = 0.037). AI-derived probability scores differed significantly among visual categories (all p < 0.01). AI-derived ILA/ILD probability was associated with five-year OS in patients with esophageal cancer and may provide a reproducible quantitative marker for opportunistic risk stratification using chest CT.