An Improved Full Convolutional Network Combined with Conditional Random Fields for Brain MR Image Segmentation...

Area of Science:

• Medical Imaging
• Computer Vision
• Neuroscience

Background:

• Deep convolutional neural networks (CNNs) show limitations in precise brain region boundary segmentation.
• Existing algorithms struggle with object boundary segmentation accuracy in brain tissue.

Purpose of the Study:

• To enhance brain tissue segmentation accuracy.
• To develop an object region segmentation algorithm integrating pixel-level and semantic information.

Main Methods:

• Utilized CNN with an attention module for semantic information extraction.
• Employed a pixel-level classifier for coarse segmentation.
• Applied conditional random fields (CRFs) to model pixel relationships and extract local features.
• Fused semantic and local pixel-level information as potentials in a Gibbs distribution for fine segmentation.

Main Results:

• The proposed algorithm achieves higher precision compared to state-of-the-art deep feature models.
• Successfully addresses the issue of rough edge segmentation in brain regions.
• Demonstrates good 3D visualization effects.

Conclusions:

• The fusion of pixel-level and semantic information offers a superior approach to brain region segmentation.
• The developed algorithm provides enhanced accuracy and finer segmentation boundaries.
• This method holds promise for improved neuroimaging analysis and visualization.