ESTRO 2025 - Abstract Book

S3750

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

1449

Digital Poster A.I. generated prediction model for treatment response after SBRT in melanoma brain metastases Donato Pezzulla 1 , Rossella Di Franco 2 , Gabriella Macchia 1 , Valentina Borzillo 2 , Carmela Romano 3 , Esmeralda Scipilliti 2 , Gianluca Ametrano 2 , Francesco Deodato 1,4 , Paolo Muto 2 , Savino Cilla 3 1 Radiation Oncology Unit, Responsible Research Hospital, Campobasso, Italy. 2 Department of Radiation Oncology, Istituto Nazionale Tumori-IRCCS-Fondazione G. Pascale, Napoli, Italy. 3 Medical Physics Unit, Responsible Research Hospital, Campobasso, Italy. 4 Istituto di Radiologia, Università Cattolica del Sacro Cuore Roma, Roma, Italy Purpose/Objective: We aimed to develop and validate a machine-learning (ML) prediction model for the complete and partial response of oligometastatic brain lesions from melanoma after two months from SBRT. Material/Methods: 51 consecutive patients with a total 99 MBMs treated with CyberKnife® SRS/SRT from December 2012 to December 2018 were selected among the ones enrolled in the Rabbit study [1]. Local response was evaluated on brain MRI performed during follow-up using response evaluation criteria in solid tumors (RECIST) and revised using response criteria for brain metastases from the response assessment in neuro-oncology criteria (RANO) group. For each treated lesion we consider the following variables: number of treated lesion, the LDH pre-treatment level, previous treatments (surgery or whole brain RT), BRAF status, time from primary diagnosis to brain lesions discovery, lesion dimension and site, SBRT dose and fractionation, type of concomitant treatment with ipilimumab (IPI). The Least Absolute Shrinkage and Selection Operator (LASSO) was performed to select the clinical covariates that are strongly associated with the treatment response output. A classification and regression tree analysis (CART) model was trained and validated to predict the complete response of brain lesions after SBRT. The performance of models was assessed by receiver operating characteristic curves (ROC) and area under the curve (AUC). Results: The clinical characteristics of these patients are described in Table 1. 21.4% of lesions had a complete or partial response and were used as ground truth for the supervised models. LASSO strongly associated complete response with three variables, namely the type of concomitant treatment with ipilimumab, BRAF status, and time from primary diagnosis to brain lesions discovery, that were used as input for ML modeling. In the training and validation sets of the CART model, the AUCs for complete response were 0.776 (95% CI: 0.726-0.786) and 0.751 (95% CI: 0.719 0.802), respectively. CART analysis classified patients with concomitant treatment to have a probability of complete response of 75.0%. Other temporal choices for ipilimumab decreased the complete response down to 18.7%. In this last group of patients, a wild-type BRAF status still decreases the CR probability to 8.5%.