High-order Feature Learning for Multi-atlas based Label Fusion: Application to Brain Segmentation with MRI

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Multi-atlas segmentation methods are effective for brain ROIs segmentation.
• Current methods often rely on low-level image intensity features, which are insufficient for complex MR image patterns.
• High-order relationships within image patches are crucial for accurate segmentation.

Purpose of the Study:

• To develop a high-order feature learning framework for multi-atlas based label fusion.
• To improve the segmentation accuracy of structural brain MR images.
• To address the limitations of using only low-level image intensity features.

Main Methods:

• Employed an unsupervised feature learning method (means-covariances restricted Boltzmann machine, mcRBM) to extract high-order features (mean and covariance) from brain MR image patches.
• Proposed a group-fused sparsity dictionary learning method for joint calculation of voting weights in label fusion.
• Integrated learned high-order features with original image intensity features.

Main Results:

• The proposed method achieved superior Dice ratios compared to state-of-the-art methods on ADNI, NIREP, and LONI-LPBA40 datasets.
• Achieved Dice ratios of 88.30% and 88.83% for left and right hippocampus on ADNI and NIREP datasets, respectively.
• Outperformed existing methods with Dice ratios of 79.54% and 81.02% on LONI-LPBA40 dataset.

Conclusions:

• The developed high-order feature learning framework significantly enhances multi-atlas based brain MR image segmentation.
• The integration of mcRBM and group-fused sparsity dictionary learning offers a robust approach for ROI segmentation.
• The method demonstrates improved performance across multiple datasets, highlighting its generalizability.