Local label learning (LLL) for subcortical structure segmentation: application to hippocampus segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate segmentation of subcortical structures in brain images is crucial for quantitative analysis but remains challenging.
• Multi-atlas based segmentation methods are widely used but often rely on suboptimal predefined weighting models for label fusion.

Purpose of the Study:

• To develop a novel local label learning strategy for improved automatic and reliable segmentation of subcortical brain structures.
• To enhance the accuracy of segmentation by utilizing statistical machine learning techniques for label fusion.

Main Methods:

• A L1-regularized support vector machine (SVM) classifier was employed for voxel-wise segmentation.
• A k-nearest neighbor (kNN) based strategy was used for selecting training samples based on intensity and texture features.
• The method was validated on over 100 magnetic resonance (MR) images for hippocampal segmentation.

Main Results:

• The proposed local label learning strategy significantly outperformed state-of-the-art label fusion methods in segmentation accuracy.
• Validation experiments demonstrated consistent improvements in segmenting hippocampal structures across diverse datasets.
• Volumetric analysis confirmed the method's effectiveness in identifying hippocampal volume changes associated with Alzheimer's disease.

Conclusions:

• The novel local label learning approach offers a more accurate and reliable method for subcortical brain structure segmentation.
• This technique has the potential to advance quantitative brain image analysis and the study of neurodegenerative diseases like Alzheimer's.