Learning-based 3T brain MRI segmentation with guidance from 7T MRI labeling.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Accurate brain magnetic resonance (MR) image segmentation into white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) is vital for neurological research and clinical diagnosis.
• Standard 3T MR images often lack sufficient contrast for reliable segmentation, hindering the training of learning-based methods.
• Ultrahigh field 7T MR imaging offers superior contrast and signal-to-noise ratio, providing high-quality data for generating reliable ground truth labels.

Purpose of the Study:

• To develop and validate a novel, fully automated method for segmenting 3T brain MR images.
• To leverage high-quality 7T MR images for creating reliable ground truth labels to train segmentation algorithms for lower-field 3T images.
• To enhance the accuracy of white matter, gray matter, and cerebrospinal fluid segmentation in 3T brain MR images.

Main Methods:

• A random forest-based algorithm was developed for segmenting 3T brain MR images.
• The algorithm was trained using semiautomatically derived labels from high-quality 7T MR images.
• A cascade of random forest classifiers was employed to iteratively refine probability maps for improved tissue segmentation.

Main Results:

• The proposed method achieved high segmentation accuracy on a local dataset (10 subjects) and the large Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
• For the local dataset, mean Dice ratios were 94.52% ± 0.9% for WM, 89.49% ± 1.83% for GM, and 79.97% ± 4.32% for CSF.
• The algorithm demonstrated statistically significant improvements over state-of-the-art methods on both datasets (p-values < 0.021).

Conclusions:

• A novel, fully automated method for 3T brain MR image segmentation has been successfully developed and validated.
• The proposed approach effectively utilizes 7T MR data to overcome the limitations of 3T image quality for accurate segmentation.
• This method offers a significant advancement for brain structural analysis and disease diagnosis using readily available 3T MR data.