VOTENET+ : AN IMPROVED DEEP LEARNING LABEL FUSION METHOD FOR MULTI-ATLAS SEGMENTATION.

Area of Science:

• Medical image analysis
• Computational neuroscience
• Machine learning

Background:

• Multi-atlas segmentation (MAS) is crucial for analyzing medical images.
• Existing methods like Joint Label Fusion (JLF) often rely on image intensity differences.
• Deep learning models offer potential for improved atlas label prediction.

Purpose of the Study:

• To enhance multi-atlas segmentation (MAS) performance.
• To integrate the VoteNet model with the Joint Label Fusion (JLF) approach.
• To develop an improved deep network, VoteNet+, for more accurate brain image segmentation.

Main Methods:

• Utilized a deep convolutional neural network for atlas probability prediction.
• Developed VoteNet+, a novel deep network predicting atlas label differences.
• Employed Joint Label Fusion (JLF) as the label fusion method.
• Applied Platt scaling for probability calibration.

Main Results:

• Deep learning-based probability prediction outperforms traditional intensity-based methods in JLF.
• VoteNet+ demonstrated superior performance compared to the original VoteNet.
• JLF proved more effective than plurality voting within the VoteNet framework.
• The proposed method achieved better segmentation accuracy on LPBA40 3D MR brain images.

Conclusions:

• The integration of VoteNet with JLF, particularly using the enhanced VoteNet+ model, significantly improves MAS performance.
• Deep learning-based probability prediction is a more effective strategy for atlas label discrimination.
• The developed method offers a promising advancement for accurate 3D brain MR image segmentation.