Development and validation of a CT-based habitat radiomics model for predicting pathological grading in non-small cell lung cancer.

Objective: To develop and validate models for predicting pathological grading of non-small cell lung cancer (NSCLC) using habitat radiomics and clinical semantic features. Materials and methods: In this retrospective study of 800 NSCLC patients, a whole tumor volume (WTV) was delineated by applying a 3 mm expansion to the gross tumor volume (GTV) on non-contrast CT scans. Habitat subregions within the WTV were identified using K-means clustering. A two-step binary classification model was constructed to predict pathological grades: Model-1 distinguished Grade 3 from combined Grades 1-2, and Model-2 further differentiated Grade 1 from Grade 2. Predictive models were built with logistic regression based on four distinct feature sets: WTV radiomics (Clf WVOI), habitat radiomics (Clf Habitats), clinical features (Clf Clinical), and a combined feature set (Clf Total). Results: In both Model-1 and Model-2, the classification performance of Clf Habitats was generally superior to that of Clf WVOI and Clf Clinical, achieving an AUC of 0.89 and 0.87, specificity of 0.73 for both models, and BACC of 0.78 and 0.79, respectively, on the test set. The combined model, Clf Total, achieved the best predictive performance on the test set, with AUC values of 0.91 and 0.88, specificity of 0.84 and 0.77, and BACC of 0.82 and 0.81. Conclusion: Habitat radiomics significantly improves NSCLC pathological grading. The multimodal model offers robust performance and high specificity, aiding personalized treatment planning.