Noninvasive Prediction of Bone Metastasis-Free Survival in Lung Adenocarcinoma Using Interpretable CT-based Deep Learning Model.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Shandong Cancer Hospital and Institute,Shandong First Medical University and Shandong Academy of Medical Sciences, China (J.G.,Y.H.); Department of Radiology, The Affiliated Hospital of Qingdao University, Qingdao, China (J.G., T.W., P.N., W.X.).
- Department of Radiology, Qilu Hospital of Shandong University, Jinan, Shandong, China (W.S.).
- Department of Radiology, The Affiliated Hospital of Qingdao University, Qingdao, China (J.G., T.W., P.N., W.X.).
- School of Engineering Medicine, Beihang University, Beijing, China (J.M.).
- Department of Radiology, Shandong Cancer Hospital and Institute,Shandong First Medical University and Shandong Academy of Medical Sciences, China (J.G.,Y.H.).
- Department of Radiology, The Affiliated Hospital of Qingdao University, Qingdao, China (J.G., T.W., P.N., W.X.). Electronic address: [email protected].
Abstract
Preoperatively identifying patients at high risk of bone metastasis (BM) remains challenging in resectable lung adenocarcinoma (LUAD), limiting early risk-adapted surveillance. We aimed to develop and validate an interpretable CT-based deep learning (DL) model-DL bone metastasis-free survival (BMFS) prediction signatures (DBPs)-to predict BMFS and provide time-dependent BM risk probabilities across follow-up. In this retrospective multicohort study, 1042 patients with preoperative computed tomography (CT) were included (training n = 594, internal validation n = 262, external validation n = 186). DBPs were evaluated using the area under the receiver operating characteristic curve (AUC) and Brier score. Interpretability was assessed by associating CT-derived DL features with histopathologic risk factors through unsupervised clustering and gradient boosting (GB) with SHapley Additive exPlanations (SHAP). DBPs predicted BMFS with AUCs of 0.822 (95% confidence interval [CI]: 0.744-0.901) and a Brier score of 0.08 in internal validation, and an AUC of 0.800 (95% CI: 0.698-0.888) and a Brier score of 0.10 in external validation. Unsupervised clusters derived from CT-based DL features were significantly associated with histopathologic risk factors (all P < 0.05), except epidermal growth factor receptor (EGFR) mutation status (P = 0.08). DL features also showed moderate discrimination for vascular invasion (VI) and visceral pleural invasion (VPI), with AUCs of 0.738 (95% CI: 0.656-0.809) and 0.708 (95% CI: 0.624-0.782), respectively, supporting the biological interpretability of the imaging signature. An interpretable CT-based DL model enables preoperative BMFS prediction in resectable LUAD, and supports interpretability by relating imaging-derived representations to established histopathologic aggressiveness factors.