Treatment-wise Glioblastoma Survival Inference with Multi-parametric Preoperative MRI

Area of Science:

• Neuro-oncology
• Medical imaging
• Machine learning

Background:

• Glioblastoma (GBM) survival time (ST) prediction is crucial for personalized treatment planning.
• Previous studies primarily mapped MRI scans to ST, neglecting the causal impact of treatment choice.
• A unified approach considering both patient status (MRI) and treatment is needed.

Purpose of the Study:

• To develop a novel treatment-conditioned regression model for predicting glioblastoma survival time.
• To integrate preoperative MRI data with treatment information for more accurate ST prediction.
• To enable personalized treatment planning by comparing predicted ST across different therapeutic options.

Main Methods:

• A treatment-conditioned regression model was proposed, incorporating both MRI scans and treatment data.
• Adaptive instance normalization was employed to inject treatment information into convolutional layers.
• The framework was evaluated on the BraTS20 dataset, considering Gross Total Resection (GTR), Subtotal Resection (STR), and no resection.

Main Results:

• The proposed model effectively utilizes data from all treatments in a unified manner.
• Injecting treatment information significantly improved the accuracy of glioblastoma survival time estimation.
• The framework demonstrated the effectiveness of considering treatment as a causal factor in ST prediction.

Conclusions:

• The developed treatment-conditioned model offers a more accurate approach to predicting glioblastoma survival.
• Integrating treatment information alongside MRI data enhances personalized treatment planning for GBM patients.
• This method provides a unified framework for utilizing diverse treatment data in survival prediction.