Prediction of response using artificial intelligence in cases treated with stereotactic brain radiotherapy

Prediction of Response Using Artificial Intelligence in Cases Treated with Stereotactic Brain Radiotherapy. Eskisehir Osmangazi University Faculty of Medicine, Department of Radiation Oncology, Medical Specialization Thesis, Eskisehir, 2025. In this study, we aimed to evaluate the potential of RM features extracted from T1 and T2 sequence magnetic resonance imaging (MRI) obtained before stereotactic body radiation therapy (SBRT) to predict treatment response in patients with brain metastases (BM) using artificial intelligence (AI) models. Patients who underwent SBRT for BM between January 2015 and March 2023 were retrospectively reviewed. A total of 91 BM cases treated with SBRT were included in the study. After the diagnosis of BM, patients received SBRT with a median dose of 27 Gy in 3 fractions. Contrast-enhanced brain MRIs were performed at 3 and 6 months for response assessment. Response evaluation was conducted based on RANO-BM criteria and classified into four groups. Patients with complete and partial responses were grouped as responders, while those with stable or progressive disease were categorized as non-responders. The data were randomly divided into 70% training and 30% test sets. To detail the training results of the models, accuracy, recall, precision, and the F1 score, which represents the harmonic mean of precision and recall, were used. These selected radiomic features were assessed for their ability to predict response in models created with deep learning algorithms based on neural networks, such as MLP and ANFIS. As a result, while the performances of both ANFIS and MLP models were notably high in the T1 and T2 sequence training datasets, in the test group, the MLP-1 and ANFIS-1 models for the T1 sequence (F1 score: 0.74 vs. 0.72) effectively predicted significant radiomic features. For the T2 sequence, the MLP-1, ANFIS-1, and ANFIS-2 models demonstrated good performance (F1 score: 0.74 vs. 0.82 vs. 0.86). In conclusion, this study demonstrated that machine learning models using MRI-based radiomic features could predict early SBRT response in BM with distinctive accuracy. Our findings are anticipated to make a significant contribution to the literature by encouraging multicenter studies and increasing interest in personalized treatment approaches.