Radiomic approach to support Multidisciplinary Tumor Board decision-making in Locally Advanced Non-Small Cell Lung Cancer

Background The optimal therapeutic management of locally advanced non-small cell lung cancer (LA-NSCLC) remains a major clinical challenge due to its biological and anatomical heterogeneity. Multidisciplinary tumor boards (MTBs) play a crucial role in integrating clinical, pathological, and radiological information to define the most appropriate treatment strategy. In this complex decision-making setting, radiomics has emerged as a quantitative imaging approach capable of capturing tumor heterogeneity and biological behavior beyond visual assessment. This study aimed to integrate clinicopathological data with radiomic features derived from both the primary tumor and mediastinal lymph nodes (LNs) to develop an AI-based model supporting MTB decisions in LA-NSCLC. Materials and methods We retrospectively included patients with stage III NSCLC discussed at the Thoracic Oncology MTB of the Veneto Institute of Oncology and the University Hospital of Padua. We classified them into three treatment groups: A) upfront surgery; B) neoadjuvant systemic treatment followed by surgery; C) chemo-radiotherapy. Patients were split into train and test cohorts. Clinical, pathological, and radiomic variables were extracted from baseline contrast-enhanced CT scans, following standardized segmentation of both primary lesions and mediastinal LNs. Radiomic features were pre-processed to remove noise, harmonized across scanners using the NeuroComBat algorithm, and combined with clinical variables for model development. Two machine-learning classifiers were trained: (1) to discriminate patients referred to upfront surgery from those undergoing multimodal treatments (A vs. B+C), and (2) to differentiate patients receiving neoadjuvant systemic therapy followed by surgery from those treated with definitive chemoradiotherapy (B vs. C). Model performance was evaluated using ten train-test splits with five-fold cross-validation, expressed as the area under the receiver operating characteristic curve (AUC) and accuracy, and further validated on an independent test cohort. Results Among 131 eligible patients, 76 were included in the training cohort and 19 in the independent test cohort. The best model distinguishing upfront surgery (A) from other treatment pathways achieved an AUC of 0.847 and accuracy of 0.795 in the training set, and 0.808 and 0.700 in the test set, respectively. The model distinguishing between neoadjuvant treatment plus surgery (B) and definitive chemoradiotherapy (C) achieved an AUC of 0.740 and accuracy of 0.700 in the training cohort, and 0.754 and 0.740 in the test cohort. Radiomic features derived from mediastinal lymph nodes, including sphericity, intensity variability, and surface-to-volume ratio, contributed significantly to model performance, highlighting their potential role as imaging biomarkers reflecting nodal biology. Conclusions This proof-of-concept study demonstrates the feasibility of an integrated clinico-radiomic approach to support MTB treatment decisions in stage III NSCLC. By combining clinical and imaging features extracted from both primary tumors and mediastinal lymph nodes, our machine-learning model reproduced real MTB recommendations with high accuracy and generalizability. If prospectively validated, this decision-support framework could enhance the objectivity, reproducibility, and personalization of therapeutic strategies for patients with LA-NSCLC, bridging the gap between quantitative imaging science and multidisciplinary clinical practice.