Delta-Enhanced CT-Based 2.5D Deep Learning Model for Noninvasive Prediction of Leptomeningeal Metastasis in Lung Adenocarcinoma

Leptomeningeal metastasis (LM) is a devastating complication of lung adenocarcinoma, the current diagnosis of which relies on invasive cerebrospinal fluid cytology or costly MRI, both of which are limited in sensitivity and accessibility. To enable early, noninvasive LM risk prediction, we developed a 2.5D delta-radiomics deep learning framework based on routinely acquired contrast-enhanced CT scans. The proposed architecture integrates multi-instance learning to aggregate slice-level features into patient-level representations while preserving spatial context and controlling computational complexity. Delta features, derived by comparing pre-LM and post-LM scans, quantified subtle longitudinal changes within intratumoral and peritumoral regions. Multiple classifiers were evaluated, with XGBoost selected for its optimal performance, and Grad-CAM visualization was employed to assess spatial attention, revealing biologically plausible regions that drove model predictions. A nomogram was constructed to facilitate individualized clinical application, and decision curve analysis (DCA) demonstrated a consistent net benefit across a broad range of decision thresholds. In the test set, the delta-based multi-instance learning signature (MILDelta) achieved an AUC of 0.871, outperforming models trained solely on pre-LM (AUC 0.574) or post-LM (AUC 0.661) images, whereas a combined model incorporating imaging signatures and clinical variables, including EGFR mutation status, yielded the highest predictive accuracy, with an AUC of 0.910. This interpretable delta-informed model leverages standard contrast-enhanced CT to achieve individualized, noninvasive LM risk stratification with demonstrated generalizability and clinical utility, offering a scalable tool for timely intervention in precision oncology workflows.