AI-Driven Preoperative Chest Radiograph Analysis for Prognostic Stratification in Surgically Resected Pathological Stage 1 Non-Small Cell Lung Cancer.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Research Institute of Radiological Science and Center for Clinical Imaging Data Science, Yongin Severance Hospital, Yonsei University College of Medicine, Seoul, South Korea.
- Division of Pulmonology, Allergy and Critical Care Medicine, Department of Internal Medicine, Yongin Severance Hospital, Yonsei University College of Medicine, Seoul, South Korea.
- Department of Radiology, Research Institute of Radiological Science, Severance Hospital, Yonsei University College of Medicine, Seoul, South Korea.
- Department of Radiology, Research Institute of Radiological Science, Severance Hospital, Yonsei University College of Medicine, Seoul, South Korea. [email protected].
Abstract
The purpose of the study is to investigate the potential of artificial intelligence (AI)-driven analysis of preoperative chest radiograph (CXR) for predicting postoperative outcome in patients with early-stage non-small cell lung cancer (NSCLC). We retrospectively enrolled 416 consecutive patients (mean age, 65.6 years ± 9.9; 197 men) who underwent curative surgical resection for pathological stage 1 NSCLC at two referral hospitals between March 2020 and February 2021. AI-driven preoperative CXR analysis was performed for the detection of four abnormalities. The lesion detectability on CXR by AI analysis (abnormality score threshold of ≥ 15) was assessed. Cox proportional hazards regression analyses were performed to determine predictors of recurrence-free survival (RFS), and the performance of prognostic models based on clinical variables and AI results was compared to models based on clinical variables and pathologic tumor size/preoperative CT parameters. AI-based abnormality score was median 39.1% (interquartile range, 4.0-82.0). During a follow-up period of 1060 ± 200.7 days, 34 patients (8.2%) experienced recurrence. Both AI detectability and AI abnormality score were significant independent risk factors for poor RFS (hazard ratio 7.201 [95% CI 2.533-20.470] and 1.029 [95% CI 1.016-1.042] per 1% increase in score, p < 0.001). A multivariable prognostic model based on AI abnormality score showed comparable performance (c-index 0.795) to the models based on pathologic tumor size or CT parameters (c-index 0.794). Adding AI abnormality score to CT-derived tumor size significantly improved discrimination performance compared with the model using CT parameters (c-index 0.837 vs. 0.821, p < 0.001). AI-driven analysis of preoperative CXR can enhance the preoperative prediction of postoperative prognosis in patients undergoing surgical resection for early-stage NSCLC.