Uncertainty-Aware Multi-Dimensional Mutual Learning for Brain and Brain Tumor Segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Neuroimaging

Background:

• Brain MRI segmentation commonly uses 3D Convolutional Neural Networks (CNNs) on volumes or 2D CNNs on slices.
• 3D CNNs capture spatial relationships, while 2D CNNs excel at local features, presenting complementary information.

Purpose of the Study:

• To develop a novel framework that leverages the complementary strengths of different dimensional CNNs for enhanced brain MRI segmentation.
• To improve the generalization ability of segmentation models by enabling simultaneous learning and mutual supervision between networks.

Main Methods:

• An Uncertainty-aware Multi-dimensional Mutual learning framework was developed, integrating 2D-CNN, 2.5D-CNN, and 3D-CNN.
• Each network learns simultaneously, providing soft labels as supervision to others.
• An uncertainty gating mechanism ensures the reliability of shared information by selecting qualified soft labels.

Main Results:

• The framework significantly enhances backbone network performance across multiple datasets.
• Achieved Dice metric improvements of 2.8% on MeniSeg, 1.4% on IBSR, and 1.3% on BraTS2020.
• Demonstrated the generalizability of the framework across different backbones.

Conclusions:

• The proposed Uncertainty-aware Multi-dimensional Mutual learning framework effectively integrates complementary information from different dimensional CNNs.
• This approach leads to substantial improvements in brain MRI segmentation accuracy.
• The method offers a reliable and generalizable solution for medical image segmentation tasks.