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

Giulia Pasello^1*Harel Kotler²Alessandra Ferro¹Luca Bergamin³Elena Scagliori⁴Angela Grassi⁵Fabio Aiolli³Mattia De Nuzzo¹Marco Schiavon⁶Matteo Sepulcri⁷Marco Krengli⁷Valentina Guarneri⁸Francesca Caumo²Gisella Gennaro²

• ¹Medical Oncology 2, Veneto Institute of Oncology IOV – IRCCS, Padova, Italy
• ²Breast Radiology Unit, Veneto Institute of Oncology IOV – IRCCS, Padova, Italy
• ³Department of Mathematics, University of Padova, Padova, Italy
• ⁴Radiology Unit, Veneto Institute of Oncology IOV – IRCCS, Padova, Italy
• ⁵Clinical Research Unit, Veneto Institute of Oncology IOV-IRCCS, Padova, Italy
• ⁶Thoracic Surgery Unit, Department of Cardiac, Thoracic and Vascular Sciences and Public Health, University of Padova, Padova, Italy
• ⁷Radiotherapy Unit, Veneto Institute of Oncology IOV – IRCCS, Padova, Italy
• ⁸Department of Surgery, Oncology and Gastroenterology, University of Padova, Padova, Italy

Background and Objective Selecting the optimal treatment for locally advanced non-small cell lung cancer (LA-NSCLC) is complex and typically requires multidisciplinary tumor board (MTB) evaluation. This study investigated whether machine learning (ML) models trained on MTB decisions could support treatment selection by integrating clinicopathological characteristics with radiomic features from both the primary tumor and mediastinal lymph nodes (LN). Materials and methods We retrospectively analyzed patients with LA-NSCLC whose treatments had been decided by an expert MTB. Patients were categorized into three pathways: (A) upfront surgery, (B) neoadjuvant systemic treatment followed by surgery, (C) concurrent chemoradiotherapy. Baseline CT scans were segmented to extract radiomic features from primary tumors and mediastinal LNs. Two ML models were developed based on clinicopathological and radiomic data, using MTB decisions as ground truth: (1) A vs. Rest and (2) B vs. C. Performance was assessed in independent training and test cohorts using the area under the receiver operating characteristic curve (AUC) and accuracy. Results In the training cohort, the A vs. Rest achieved an AUC of 0.847 and accuracy of 0.795 with 13 features, while the B vs. C model reached an AUC of 0.740 and accuracy of 0.700 with 9 features. In the test cohort, results remained robust, with an AUC of 0.808 (accuracy 0.700) for A vs. Rest and an AUC of 0.754 (accuracy 0.740) for B vs. C. Conclusions ML models combining clinicopathological and radiomic features can reproduce MTB treatment recommendations for LA-NSCLC with good accuracy. This approach may provide decision in settings with limited MTB expertise and promote more consistent treatment allocation.