MVFusFra: A Multi-View Dynamic Fusion Framework for Multimodal Brain Tumor Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain tumor diagnosis relies on multimodal imaging, often analyzing axial, coronal, and sagittal views.
• Existing methods may not fully leverage the 3D spatial information within these datasets.

Purpose of the Study:

• To propose and evaluate a multi-view dynamic fusion framework (MVFusFra) for improved brain tumor segmentation.
• To enhance the utilization of 3D information from multimodal brain images.

Main Methods:

• Developed a multi-view deep neural network architecture for segmenting tumors from individual views.
• Implemented a dynamic decision fusion method (voting and weighted averaging) to integrate multi-view segmentation results.
• Introduced a multi-view fusion loss (segmentation, transition, decision loss) to ensure consistency and facilitate training.

Main Results:

• Multi-view fusion significantly improved brain tumor segmentation performance compared to single-view methods.
• The proposed multi-view fusion loss demonstrated effectiveness in training multi-view networks.
• MVFusFra achieved superior segmentation efficiency compared to existing approaches on BRATS 2015 and 2018 datasets.

Conclusions:

• The MVFusFra framework effectively enhances brain tumor segmentation by integrating multi-view imaging data.
• The developed fusion loss is crucial for achieving consistent and accurate segmentation results.
• MVFusFra offers a promising, efficient solution for improving brain tumor diagnosis through advanced image analysis.