Radiomics-Based Prediction of Treatment Response in Non-Small Cell Lung Cancer Using Pre-Treatment CT Imaging...

Background/Objectives: Non-small cell lung cancer (NSCLC) is the most common type of lung cancer, representing nearly 85% of cases worldwide. Predicting patient response before treatment, however, remains a major clinical challenge. Radiomics enables non-invasive extraction of quantitative imaging data that reflects tumor characteristics and heterogeneity. In this study, we developed and externally validated a CT-derived radiomic signature for predicting outcomes in advanced NSCLC patients treated with first-line platinum-based chemotherapy. Methods: NSCLC across three tertiary care centers were included. Tumor lesions from baseline contrast-enhanced CT scans were semi-automatically outlined using 3D Slicer software (version 5.12.2), followed by the extraction of 851 quantitative imaging biomarkers through PyRadiomics version 3.0. The extracted parameters comprised histogram-based features, morphological measurements, texture-related variables, and wavelet-derived attributes. Reliability of feature extraction was evaluated using intraclass correlation coefficient analysis, whereas LASSO regression together with stability selection was applied to identify the most relevant predictors. Predictive capability was assessed across five machine learning techniques, including logistic regression, random forest, support vector machine, gradient boosting, and multilayer perceptron models. The resulting radiomics-based composite score was then tested in an independent external validation cohort. Response to treatment was determined according to RECIST 1.1 guidelines. Results: After feature reproducibility assessment and stability screening, 421 radiomic parameters were retained for further analysis and 14 feasible parameters were selected by LASSO. There was good predictive power of the integrated model that included both radiomic and clinical features, with AUC values of 0.876, 0.849, and 0.831 in the training set, internal validation set and external validation set, respectively. The radiomic signature successfully differentiated patients according to their probability of therapeutic response. Furthermore, by using decision-curve analysis, the combined radiomics nomogram was more clinically useful than models based on clinical characteristics within clinically relevant decision-threshold ranges with greater net benefit. Conclusions: This pre-treatment CT-based radiomics signature demonstrates potential as a decision-support tool for chemotherapy-response prediction in NSCLC. However, prospective and multi-ethnic validation is required before clinical application. Based on this data, larger multi-ethnic cohorts should be considered for prospective validation.