Abstract
AIM
The diffuse infiltration of glioblastoma (GBM) along the intricate network of white matter (WM) tracts poses a major challenge. Our aim is to investigate the potential of axons intersecting tumor enhancement (AXITE)-derived radiomic features in predicting one-year overall survival (OS), in comparison to conventional tumor-based radiomics.
METHODS
This study included data from the publicly available UCSF-PDGM-v2 (n=97). T1+C images and their corresponding tumor mask were rigidly registered to a common brain template (MNI152) using ANTs. WM tracts within the tumor-defined region of interest were generated using the HCP1065-1mm tractography atlas in DSI Studio Software, which mapped the AXITEs. Radiomic features were then extracted using PyRadiomics from subsequently masked AXITE maps as well as contrast-enhancing tumor masks for a comparative analysis. Four machine learning models, including a Random Forest, Support Vector Machine (SVM), Decision Tree, and Logistic Regression, were employed to predict OS using an 80/20 train/test split.
RESULTS
Among all models, SVM was the most effective in predicting OS. AXITE radiomics demonstrated remarkable precision (0.73 for OS < 1 yr and 0.78 for OS >1yr), accuracy (0.75), F1-Score (0.75), recall (0.75), and AUC of 0.84 in ROC analysis outclassing the performance of traditional tumor radiomics which had an accuracy (0.60), precision (0.66), F1-Score (0.56) and AUC of 0.60.
CONCLUSIONS
AXITE Radiomics holds promising potential in capturing tumoral heterogeneity and white matter tract infiltration of GBM, demonstrating superior prediction over traditional tumor radiomics. Our approach presents a more accurate representation of tumor invasiveness and tumor resistance associated with OS more accurately and promises a potential tool for improving patient prognosis and patient management.