Development of CT-based deep learning and radiomics models for the early prediction of microwave ablation efficacy in lung tumors

Accurate and early prediction of the efficacy of microwave ablation for lung tumors is crucial for optimizing personalized therapeutic strategies. However, non-invasive imaging biomarkers for immediate post-procedural assessment remain under-explored. The aim of this study was to develop and explore Computed Tomography-based deep learning and radiomics models to predict early treatment responses following microwave ablation. This retrospective study enrolled 167 patients with lung tumors who underwent unenhanced Computed Tomography scanning immediately after microwave ablation, with a minimum follow-up period of six months. Regions of interest were semi-automatically delineated on the immediate post-ablation Computed Tomography images. Radiomics models were constructed using AdaBoost, Gaussian Naive Bayes, and Stochastic Gradient Descent algorithms, while a deep learning model was developed based on the ResNet50 architecture. The predictive performance of the established models was evaluated using receiver operating characteristic curves, calibration plots, and decision curve analysis. The radiomics models (AdaBoost, Gaussian Naive Bayes, and Stochastic Gradient Descent) and the ResNet50-based deep learning model demonstrated favorable predictive capability, as reflected by their areas under the curve, sensitivity, specificity, and accuracy. Furthermore, combined models integrating these imaging signatures with clinical parameters exhibited good performance. The ResNet50-based deep learning model and the radiomics models derived from immediate post-ablation Computed Tomography scans show promising feasibility as non-invasive tools for the early prediction of microwave ablation efficacy in lung tumors, although further validation in larger, multi-center cohorts is warranted.