Noninvasive Prediction of Bone Metastasis-Free Survival in Lung Adenocarcinoma Using Interpretable CT-based Deep...

Area of Science:

• Radiology and Oncology
• Artificial Intelligence in Medicine
• Cancer Imaging

Background:

• Identifying patients at high risk for bone metastasis (BM) in resectable lung adenocarcinoma (LUAD) preoperatively is crucial for timely, risk-adapted surveillance.
• Current methods for predicting BM risk lack sufficient accuracy and interpretability, hindering early intervention strategies.

Purpose of the Study:

• To develop and validate an interpretable computed tomography (CT)-based deep learning (DL) model, termed DL bone metastasis-free survival (BMFS) prediction signatures (DBPs).
• To predict BMFS and provide time-dependent BM risk probabilities for patients with resectable LUAD.

Main Methods:

• A retrospective multicohort study including 1042 patients with preoperative CT scans (training, internal, and external validation cohorts).
• Development and evaluation of DBPs using area under the receiver operating characteristic curve (AUC) and Brier score.
• Assessment of model interpretability by associating CT-derived DL features with histopathologic risk factors using unsupervised clustering and SHapley Additive exPlanations (SHAP).

Main Results:

• DBPs demonstrated strong performance in predicting BMFS, with AUCs of 0.822 (internal validation) and 0.800 (external validation).
• The model achieved Brier scores of 0.08 (internal) and 0.10 (external), indicating good predictive accuracy.
• CT-derived DL features showed significant associations with histopathologic risk factors and moderate discrimination for vascular and visceral pleural invasion, supporting biological interpretability.

Conclusions:

• An interpretable CT-based DL model (DBPs) can effectively predict BMFS in patients with resectable LUAD.
• The model's interpretability is supported by the correlation of imaging features with established histopathologic aggressiveness factors.
• This AI-driven approach offers a promising tool for preoperative risk stratification and personalized surveillance in LUAD management.