Abstract
BACKGROUND: The T790M mutation is a significant mechanism of acquired resistance to EGFR-TKIs in non-small cell lung cancer (NSCLC). Its noninvasive detection in spinal metastases remains challenging due to tumour heterogeneity and limitations of current diagnostic methods. This study aimed to develop an MRI-based radiomics model derived from spinal metastases to non-invasively predict T790M resistance mutations in NSCLC patients, by incorporating intratumoral spatial heterogeneity. METHODS: One hundred ten EGFR-mutant NSCLC patients with spinal metastases (80 from Center 1, 30 from Center 2) underwent T1W and T2FS MRI scans. Spinal lesions were partitioned into phenotypically consistent subregions using patient- and population-level clustering based on local entropy to capture spatial heterogeneity. Radiomic features were extracted from each subregion, and reproducibility was assessed using the intraclass correlation coefficient (ICC >0.80). Significant features were selected via the Mann-Whitney U test and LASSO regression, and logistic regression models were constructed for each subregion and MRI sequence. A multi-sequence regional fusion model was subsequently developed based on the best-performing subregion. Model performance was evaluated by AUC, sensitivity, and specificity in both internal and external validation cohorts. SHAP analysis was conducted to interpret feature contributions. RESULTS: Models based on inner subregions with higher heterogeneity outperformed those from marginal or whole-tumor regions. The fusion model combining T1W and T2FS features achieved AUCs of 0.916 (training), 0.867 (internal validation), and 0.839 (external validation). SHAP analysis identified key textural features associated with the T790M mutation. CONCLUSION: Subregion-based MRI radiomics enables accurate, noninvasive prediction of T790M mutations in NSCLC spinal metastases. This subregion-based MRI radiomics model, to our knowledge, is the first to non-invasively predict T790M resistance mutations in spinal metastases by integrating spatial heterogeneity and SHAP interpretability. This subregion-based MRI radiomics model is exploratory and showed a consistent trend toward improved discrimination and net benefit.