The role of radiomics and delta-radiomic analysis for the prediction of prognostic and predictive factors in non-small cell lung cancer.

Background: Lung cancer is the leading cause of cancer-related death worldwide, despite the increase in therapeutic options and screening programs. Immunotherapy has significantly transformed the treatment of non-small cell lung cancer (NSCLC), however the long-term benefit is limited to a small subset of patients. However, apart from PD-L1 tissue expression—which has been repeatedly questioned—there are currently no reliable tools to predict treatment response. Study design and methods: This study included patients with advanced NSCLC from three different hospitals treated with anti-PD1 monotherapy as first or second line. All patients underwent a pre-treatment CT scan, and the lesion targets were semi-automatically segmented. Radiomic features were extracted and used to develop machine learning models to predict response to immunotherapy, labeled dichotomously as progressor or non-progressor. Results: A total of 125 patients were enrolled, and 165 parenchymal lung lesions and 86 lymph node lesions were considered. A support vector machine (SVM) model was developed for each lesion type, achieving an F1-score of 0.77 for parenchymal lesions and 0.86 for lymph node lesions in the internal dataset. In two external validation cohorts, the model achieved an F1-score of 0.69 for parenchymal lesions and 0.77 for lymph nodes, demonstrating strong generalizability across different scanners, protocols, and readers. Conclusions: These radiomics-based machine learning models showed good reproducibility 3 across external datasets and outperformed tissue-assessed PD-L1 expression in response prediction accuracy. These findings support the growing role of radiomics as a reliable and feasible tool in clinical decision making for NSCLC immunotherapy.