3D Deep Learning for Brain Tumor Segmentation and Survival Prediction: A Comprehensive Multi-Modal Analysis Using the...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Computational Neuroscience

Background:

• Three-dimensional deep learning (3D DL) shows potential for automated brain tumor segmentation and survival prediction.
• Robust validation across multiple MRI modalities is crucial for clinical implementation of 3D DL.

Purpose of the Study:

• To develop and validate a comprehensive 3D deep learning framework for brain tumor segmentation and survival prediction.
• To assess the performance of multi-modal MRI data in enhancing segmentation and classification accuracy.

Main Methods:

• A 3D U-Net architecture was developed using 369 cases from the BraTS2020 dataset for tumor segmentation.
• Segmentation outputs were used to train machine learning classifiers for 6-month and 12-month survival prediction.
• Models were evaluated using Dice Similarity Coefficient, Hausdorff Distance, AUC-ROC, and balanced accuracy.

Main Results:

• The 3D U-Net model achieved a mean validation Dice score of 0.8388 for segmentation.
• Survival classification for 12-month prediction yielded an AUC of 0.746 with 69% accuracy.
• Key predictors for 12-month survival included extent of resection and tumor medians from T1 contrast-enhanced and FLAIR MRI.

Conclusions:

• A multi-modal 3D deep learning approach significantly enhances performance in brain tumor analysis.
• Segmentation-derived features hold promise for accurate patient survival prediction.
• The developed framework offers a robust tool for clinical application in neuro-oncology.