Whole brain segmentation with full volume neural network.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Whole brain segmentation is crucial for neuroimaging, labeling anatomical regions within brain volumes.
• Current methods often use partial information (voxels, slices, sub-volumes), potentially leading to suboptimal segmentation.
• Convolutional neural networks show promise but can struggle with incomplete data representation.

Purpose of the Study:

• To develop a full volume framework for whole brain segmentation that utilizes complete volumetric information.
• To improve the accuracy and efficiency of brain segmentation compared to existing partial-volume approaches.
• To introduce an effective implementation using 3D HRNet and mixed precision training.

Main Methods:

• Proposed a full volume framework that processes the entire brain image at once.
• Implemented the framework using a 3D high-resolution network (HRNet) for detailed spatial representation learning.
• Utilized a mixed precision training scheme to optimize memory usage during training.

Main Results:

• The full volume framework demonstrated superior segmentation performance on a 3D MRI brain dataset.
• The proposed model outperformed existing state-of-the-art methods in whole brain segmentation accuracy.
• The approach effectively leverages complete volumetric information for more robust results.

Conclusions:

• The full volume framework offers a more effective approach to whole brain segmentation.
• Utilizing complete volumetric data and advanced network architectures like 3D HRNet enhances segmentation quality.
• This method represents a significant advancement in automated neuroimaging analysis.