Predicting the risk of brain metastasis following prophylactic cranial irradiation in small cell lung cancer: an intratumoural and peritumoural CT radiomics model

A machine learning model based on intratumoural and 8-mm peritumoural CT radiomics features was developed to predict brain metastasis (BM) after prophylactic cranial irradiation (PCI) in patients with small cell lung cancer (SCLC). This retrospective study included 185 patients with SCLC who underwent PCI. The patients were randomly divided into a training cohort ( n = 148) and a test cohort ( n = 37). Radiomics features were extracted from intratumoural and 8-mm peritumoural regions using the PyRadiomics toolkit. Feature correlations were assessed using Spearman’s correlation analysis. Feature selection was then performed using least absolute shrinkage and selection operator (LASSO) regression. Five predictive models were constructed using machine learning algorithms, including logistic regression (LR), support vector machine (SVM), and k-nearest neighbours (KNN): (1) an intratumoural radiomics model; (2) an 8-mm peritumoural radiomics model; (3) an intratumoural–peritumoural feature fusion model; (4) an intratumoural–peritumoural image fusion model; and (5) a clinical model. Model performance was evaluated and compared using the area under the receiver operating characteristic curve (AUC). In the training cohort, the intratumoural and peritumoural feature fusion model demonstrated optimal performance (AUC = 0.922, 95% CI 0.8649–0.9793), significantly outperforming both the single-region model and the clinical model (AUC = 0.637, 95% CI 0.5369–0.7375). In the test cohort, this model maintained good predictive performance (AUC = 0.819, 95% CI 0.6410–0.9971), demonstrating strong generalisation ability. The peritumoural model showed performance similar to that of the feature fusion model. Lower predictive performance was observed for the image fusion and clinical models. DeLong’s test showed that the IntraPeri8mm model significantly outperformed selected comparison models in the training cohort. CT radiomics features derived from the tumour and the surrounding 8-mm peritumoural region can be used to predict the risk of BM after PCI in patients with SCLC. The feature fusion model achieved the best overall predictive performance. This model may provide a useful tool for personalised BM risk stratification.