Machine and deep learning based on magnetic resonance imaging to segment glioblastoma and predict the spread of...

Glioblastoma (GB) remains one of the most aggressive brain tumors, with limited survival and high recurrence rates. In most cases, GB recurrence occurs locally, either on the residual tumor after surgery or within 2 cm of the resection cavity-but in rarer cases, tumor cells can spread beyond this margin, leading to distant recurrence. By spread, we refer to the spatial dissemination of tumor cells beyond the typical local site, which can involve distant brain regions or even the leptomeninges, significantly impacting treatment planning and surgical decision-making. This study proposes the application of Machine Learning (ML) and Deep Learning (DL) approaches to MRI data from GB patients in the preoperative phases, aiming to develop predictive models to predict the extent of recurrence spread, on the integrated analysis of clinical, imaging, and instrumental data. Additionally, we plan to design a (semi) automatic segmentation tool for tumor delineation in MRI, which will support both the implementation of the study and serve as a standalone instrument to standardize volume measurement in neuroimaging.

A multicenter retrospective collection of clinical and radiological variables will be performed for all eligible GB patients. Variables will include demographic, surgical, pathological, and preoperative MRI features. Predictive modelling will use classical ML algorithms (e.g., Random Forest, SVM, Multilayer perceptron, etc.) and a 3D U-Net architecture for DL-based image segmentation. Dimensionality reduction (PCA, LASSO, etc.) will be used to prevent overfitting and improve model generalizability. Model performance will be assessed through Area Under the Curve (AUC), P-R curve, F-score, accuracy, sensitivity, specificity, confusion matrix, and Dice score for segmentation.

The development of Artificial Intelligence (AI)-based predictive models for GB is expected to provide a major contribution to outcome prediction, early targeted interventions, and personalized care. These tools may support optimized resource allocation, reduce healthcare costs, and improve patient and family outcomes. The findings from this study will serve as a foundation for a future prospective multicenter validation study.