Prediction of intracranial response to PD-1/PD-L1 inhibitors therapy in brain metastases originating from non-small cell lung cancer using habitat imaging and peritumoral radiomics: a multicenter study

Min Ding¹Tianrui He²Jing Yu¹Jian Zheng³Song Wei⁴Yuan Yuan¹Chunhui Yang¹Ning Luo¹Xin Qi⁵Liting Liu¹Yiyang Sun¹Dailun Hou⁴Chao Yang⁵Hongxu Liu³Wenwen Liu⁶Qi Wang^1*

• ¹Second Affiliated Hospital of Dalian Medical University, Dalian, China
• ²Zhongshan Hospital Fudan University, Shanghai, China
• ³Liaoning Cancer Hospital and Institute, Shenyang, China
• ⁴Beijing Chest Hospital Affiliated to Capital Medical University, Beijing, China
• ⁵The First Affiliated Hospital of Dalian Medical University, Dalian, China
• ⁶The Second Hospital of Dalian Medical University, Dalian, China

Abstract Background: Predicting the intracranial efficacy of programmed death-1/programmed death-ligand 1 (PD-1/PD-L1) inhibitors in non-small cell lung cancer (NSCLC) patients with brain metastasis (BM) remains challenging. The objective of this study was to construct a habitat-peritumoral radiomics framework for immunotherapy response prediction, concurrently identifying the optimal peritumoral extent. Methods: This retrospective multicenter study analyzed 378 NSCLC-BM patients receiving PD-1/PD-L1 inhibitors. Participants were stratified into training (n=146), internal validation (n=63), and two external test cohorts (test 1: n=57; test 2: n=112). Logistic regression was conducted to determine significant clinical predictors. Habitat subregion segmentation was performed using K-means clustering with peritumoral extensions at incremental distances (1, 2, and 3 mm). Predictive models were developed using radiomic features extracted from intratumoral cores, habitat subregions, and peritumoral zones through machine learning approaches. A combined model integrated habitat signatures, peritumoral features, and clinical predictors. Model performance assessment employed the area under the curve (AUC), calibration curves, and decision curve analysis (DCA). Results: The habitat-based XGBoost model demonstrated superior predictive performance across all cohorts compared to alternative models, achieving AUCs of 0.900 (training), 0.886 (internal validation), 0.820 (test 1), and 0.804 (test 2). For peritumoral analysis, the peri-1 mm RandomForest model exceeded other regional configurations. Integrating peri-1 mm features and clinical factors yielded a marginal performance enhancement in the combined model, with corresponding AUCs of 0.898, 0.894, 0.837, and 0.814. The combined model demonstrated optimal calibration and significant clinical utility, as evidenced by calibration curves and DCA. Conclusion: The validated habitat-peritumoral radiomics framework, optimized at a 1-mm peritumoral extent, demonstrates robust predictive accuracy for intracranial immunotherapy response in NSCLC-BM patients and offers significant clinical utility.