Brain Tumor Segmentation Network with Multi-View Ensemble Discrimination and Kernel-Sharing Dilated Convolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate brain tumor segmentation from 3D MRI is crucial for clinical applications.
• Traditional Convolutional Neural Network (CNN) models often process MRI views independently, limiting feature integration.
• Existing multi-branch CNNs struggle to precisely characterize tumor features across varying sizes and shapes.

Purpose of the Study:

• To develop an advanced method for accurate brain tumor segmentation using 3D MRI.
• To overcome limitations of single-view processing and feature characterization in existing CNN models.
• To improve the precision of brain tumor segmentation for better clinical decision-making and surgical planning.

Main Methods:

• Proposed a hierarchical multi-view convolution method decoupling 3D convolution into axial, coronal, and sagittal views.
• Implemented a multi-branch kernel-sharing mechanism with dilated rates for parameter-consistent kernels and varied receptive fields.
• Ensembled discriminant results from three views for pixel classification.

Main Results:

• Achieved average Dice coefficients of 78.16% (ET), 89.52% (WT), and 83.05% (TC) on the BraTS2020 dataset.
• The proposed network has a low parameter count of only 0.5 M.
• Demonstrated accuracy improvements of 1.74% (ET), 0.5% (WT), and 2.19% (TC) over baseline networks.

Conclusions:

• The hierarchical multi-view convolution method significantly enhances brain tumor segmentation accuracy in 3D MRI.
• The proposed approach offers an efficient and effective solution with a reduced parameter count.
• This method holds promise for improving diagnostic accuracy and treatment planning in neuro-oncology.