Non-local Atlas-guided Multi-channel Forest Learning for Human Brain Labeling

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computer Vision

Background:

• Accurate labeling of anatomical regions in MR brain images is crucial for quantitative research.
• Existing label fusion methods often rely heavily on appearance information.
• Contextual features, proven effective in computer vision, offer potential for improved image labeling.

Purpose of the Study:

• To develop a novel learning-based label fusion method incorporating both appearance and context features for enhanced brain MR image segmentation.
• To address inter-subject variations in brain anatomy through a multi-atlas strategy.

Main Methods:

• A multi-channel random forest model was employed to learn the relationship between hybrid features (appearance and context) and target anatomical labels.
• The method was extended to a multi-atlas scenario, training a random forest per atlas and aggregating results for consensus.
• Evaluation was performed on the LONI-LBPA40 and IXI datasets.

Main Results:

• The proposed method achieved the highest labeling accuracy among evaluated state-of-the-art techniques.
• Integration of appearance and context features led to superior performance in anatomical region identification.
• The multi-atlas approach effectively handled inter-subject variability.

Conclusions:

• The novel learning-based label fusion method significantly advances the accuracy of anatomical brain MR image segmentation.
• Combining appearance and context features with a multi-atlas strategy offers a robust solution for quantitative neuroimaging research.
• This approach provides a valuable tool for precise analysis of brain structures.