Multiscale Multiparametric MRI Deep Learning for Short-Term Survival Assessment in Glioblastoma

Area of Science:

• Neuro-oncology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accurate preoperative prediction of short-term survival in glioblastoma (GBM) is crucial for treatment planning but challenging due to tumor heterogeneity.
• Existing methods struggle to capture the complex factors influencing GBM patient outcomes.

Purpose of the Study:

• To develop and externally validate a multiscale magnetic resonance imaging (MRI)-based deep learning model for predicting short-term survival in newly diagnosed glioblastoma.
• To investigate the association between the model's predictions and transcriptomic data.

Main Methods:

• Retrospective, multicenter study including 728 adult patients with newly diagnosed glioblastoma.
• A deep learning model integrated whole-brain, 3D tumor, and 2.5D tumor MRI inputs.
• Model performance was assessed using Area Under the Receiver Operating Characteristic Curve (AUC) and compared against clinical and conventional MRI baselines.

Main Results:

• The multiscale MRI deep learning model achieved high discrimination for short-term survival, with external validation AUCs ranging from 0.798 to 0.871.
• The model demonstrated significant improvements in AUC over combined clinical-MRI morphometric baselines in external cohorts.
• Model outputs were significantly associated with immune/inflammatory and cell-division/genome-maintenance transcriptomic pathways after FDR correction.

Conclusions:

• Multiscale MRI-based deep learning offers a promising tool for preoperative short-term survival assessment in glioblastoma.
• The model's performance is linked to underlying biological processes, specifically immune responses and cell cycle regulation.
• This approach may aid in personalized management strategies for glioblastoma patients.