OVERALL SURVIVAL PREDICTION OF BRAIN TUMOR PATIENTS WITH MULTIMODAL MRI USING SWIN UNETR

Area of Science:

• Neuro-oncology
• Medical Imaging Analysis
• Machine Learning for Healthcare

Background:

• Accurate glioblastoma (GBM) survival prediction is crucial for personalized treatment.
• Current methods using multimodal magnetic resonance imaging (MRI) are limited by data scarcity and reliance on handcrafted features.
• Advanced deep learning models can potentially improve prognostic accuracy.

Purpose of the Study:

• To develop a data-efficient, multi-task framework for glioblastoma survival prediction using Swin UNETR.
• To integrate hierarchical segmentation features and patient age for enhanced prognostic performance.
• To address data scarcity through segmentation pre-training and feature refinement.

Main Methods:

• Proposed a data-efficient multi-task learning framework utilizing Swin UNETR for glioblastoma survival prediction.
• Integrated multi-scale and hierarchical features from multimodal MRI, incorporating patient age.
• Employed segmentation pre-training and feature fine-tuning, refined by statistical F-values, to address data scarcity.

Main Results:

• Achieved superior segmentation accuracy on the BraTS dataset.
• Demonstrated state-of-the-art performance in glioblastoma survival prediction.
• The proposed model shows robust clinical prognostic capabilities.

Conclusions:

• The developed framework offers a robust and accurate solution for glioblastoma survival prediction.
• Integrating hierarchical segmentation features and patient age significantly improves prognostic accuracy.
• This approach holds promise for personalized treatment planning in glioblastoma patients.