Brain Tumor Segmentation and Survival Prediction Using Multimodal MRI Scans With Deep Learning

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Gliomas are primary brain tumors requiring accurate segmentation and survival prediction for effective patient management.
• Current methods face challenges in robustly segmenting tumors and predicting patient outcomes.

Purpose of the Study:

• To develop and evaluate a deep learning framework for brain tumor segmentation and overall survival prediction in glioma patients.
• To leverage multimodal MRI data for improved diagnostic and prognostic capabilities.

Main Methods:

• Utilized an ensemble of three 3D Convolutional Neural Network (CNN) architectures for robust brain tumor segmentation via majority voting.
• Extracted 4,524 radiomic features from segmented tumor regions for survival prediction.
• Employed decision tree and cross-validation for feature selection, followed by a random forest model for survival prediction.

Main Results:

• The deep learning framework achieved 2nd place in survival prediction and 5th place in segmentation among 60+ teams at the 2018 MICCAI BraTS challenge.
• Demonstrated a promising 61.0% accuracy in classifying patients into short-survivor, mid-survivor, and long-survivor categories.
• The ensemble approach effectively reduced model bias and enhanced performance for both segmentation and prediction tasks.

Conclusions:

• The proposed deep learning framework offers a robust and accurate solution for brain tumor segmentation and survival prediction in gliomas.
• Multimodal MRI data combined with advanced machine learning techniques shows significant potential for improving glioma patient care.
• The method's strong performance in a competitive challenge highlights its clinical relevance and effectiveness.