Boosting brain tumor segmentation: A novel 3D pooling approach with U-net 3D

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate brain tumor segmentation is vital for clinical decision-making.
• Existing methods face challenges with complex tumor structures and MRI intensity variations.
• U-Net 3D is a common architecture for medical image segmentation.

Purpose of the Study:

• To enhance brain tumor segmentation accuracy using multimodal MRI.
• To introduce a novel 3D pooling layer to improve U-Net 3D performance.
• To develop a robust segmentation method resilient to intensity variations.

Main Methods:

• A novel 3D pooling layer was integrated into the U-Net 3D architecture.
• Two complementary normalization pipelines were used for robustness.
• Ensemble learning by averaging predictions from independently trained networks was employed.
• The approach was evaluated on the BraTS2020 dataset using five-fold cross-validation.

Main Results:

• The proposed ensemble achieved Dice scores of 0.8299 (ET), 0.8882 (TC), and 0.8986 (WT).
• Hausdorff distance 95th percentile (HD95) values were 4.40 (ET), 4.95 (TC), and 11.14 (WT).
• The method demonstrated consistent gains over max-pooling variants and competitive performance against recent approaches.

Conclusions:

• The novel 3D pooling layer effectively improves brain tumor segmentation accuracy.
• The ensemble strategy enhances robustness to variations in MRI data.
• This approach offers a promising direction for automated and accurate brain tumor segmentation.