TabPFN-driven ternary classification of stage IA lung adenocarcinoma subtypes using AI-derived histogram features a retrospective multicenter cohort study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Thoracic Surgery, Beijing Haidian Hospital (Haidian Section of Peking University Third Hospital), Beijing, China.
- Department of Thoracic Surgery, Peking Union Medical College Hospital, Beijing, China.
- Institute of Advanced Research, Infervision Medical Technology Co. Ltd., Beijing, China.
- Department of Pathology, Peking University People's Hospital, Beijing, China.
- Department of Thoracic Surgery, Peking University People's Hospital, Beijing, China.
Abstract
Preoperative differentiation of precursor glandular lesions (PGL), minimally invasive (MIA), and invasive adenocarcinoma (IAC) in stage IA lung adenocarcinoma (LUAD) is critical for surgical planning but remains challenging due to overlapping CT features and interobserver variability. While existing artificial intelligence (AI) models focus predominantly on binary classification with limited multicenter validation, this study developed and validated a ternary classification framework using pretrained TabPFN and traditional machine learning (ML) algorithms based on AI-derived histogram features, benchmarking against intraoperative frozen section analysis. This multicenter retrospective study utilized preoperative CT scans from three institutions between September 2014 and October 2023. Data were divided into training, internal validation, and external test sets. Histogram features (n = 26) were automatically extracted using a commercial AI system (InferRead CT Lung). TabPFN and five ML algorithms were trained with selected clinical and histogram features. Performance was evaluated by accuracy, macro-AUC, sensitivity, specificity, and Cohen's Kappa. Statistical comparisons included DeLong tests for AUC and chi-square for categorical variables. The cohort comprised 584 stage IA LUAD patients (mean age 57.9 ± 11.0 years; 386 female), divided into training/validation sets (n = 412, center 1) and external test sets (n = 114, center 2; n = 58, center 3). TabPFN achieved macro-AUC of 0.781-0.911 and accuracy of 67.2-78.9% across external test sets, outperforming other ML algorithms. Of note, TabPFN achieved an overall better prediction accuracy compared to frozen section analysis on all test sets (internal: 92.3% vs 84.6%, P = 0.503; external 1: 87.5% vs 75%, P = 1.000; external 2: 67.2% vs 43.1%, P < 0.001). Subgroup analysis revealed superior performance for mGGN lesions (85%) on both external test sets. TabPFN enables robust, generalizable ternary classification of LUAD subtypes, surpassing conventional ML and frozen section analysis. Its integration with automated histogram analysis offers a scalable solution for preoperative stratification of early-stage lung cancer.