Hierarchical multi-atlas label fusion with multi-scale feature representation and label-specific patch partition

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Computer Vision

Background:

• Multi-atlas label fusion methods enhance medical image segmentation accuracy.
• Current methods often use fixed-size image patches, limiting accuracy with complex anatomical structures.
• Accurate patch similarity measurement is crucial for effective label fusion.

Purpose of the Study:

• To improve the accuracy of multi-atlas patch-based label fusion methods.
• To address limitations of fixed-size image patches in capturing complex tissue appearance.
• To introduce a novel hierarchical label fusion approach with enhanced feature representations.

Main Methods:

• Developed a multi-scale feature representation for image patches to encode local and semi-local information.
• Partitioned atlas image patches into label-specific partial patches for increased specificity and flexibility.
• Implemented a coarse-to-fine iterative hierarchical approach to refine label fusion and correct mislabeled points.

Main Results:

• The proposed hierarchical label fusion method demonstrated superior segmentation accuracy across multiple datasets (ADNI, LBPA40, SATA, IXI).
• Multi-scale features and label-specific patches improved the fidelity of patch-based similarity measurements.
• Segmentation results consistently outperformed several state-of-the-art label fusion techniques.

Conclusions:

• The novel hierarchical label fusion approach with multi-scale features and label-specific patches significantly enhances medical image segmentation accuracy.
• This method effectively captures complex anatomical structures and improves robustness compared to traditional fixed-patch methods.
• The proposed technique offers a more accurate and flexible solution for various medical image analysis applications.