Precise and Rapid Whole-Head Segmentation from Magnetic Resonance Images of Older Adults using Deep Learning

Skylar E Stolte ¹, Aprinda Indahlastari ^2,3, Jason Chen ⁴

⁰J. Crayton Pruitt Family Department of Biomedical Engineering, Herbert Wertheim College of Engineering, University of Florida, Gainesville, USA.

Imaging Neuroscience (cambridge, Mass.) +

|

March 11, 2024

View abstract on PubMed

Summary

GRACE, a new deep learning method, accurately segments whole-head MRI tissues in seconds, outperforming existing tools. This advancement is crucial for computational modeling, especially for older adults with age-related brain changes.

Area Of Science

Medical Imaging
Computational Neuroscience
Artificial Intelligence

Background

Accurate whole-head segmentation from Magnetic Resonance Images (MRI) is essential for creating individualized computational models, particularly for non-invasive brain stimulation.
Existing automatic segmentation tools often fail to adequately address the older population, who experience age-related structural changes like brain atrophy.

Purpose Of The Study

To introduce GRACE (General, Rapid, And Comprehensive whole-head segmentation), a novel deep learning method for whole-head MRI tissue segmentation.
To address the limitations of current tools in segmenting MRIs from older adults, enabling high-precision modeling for age-related brain disorders.

Main Methods

GRACE was trained and validated on a unique dataset of 177 manually corrected, T1-weighted MRI volumes.
The dataset includes meticulous manual review and segmentation into 11 distinct tissue types.
The method was compared against five freely available software tools and a traditional 3D U-Net.

Main Results

GRACE achieved superior performance in a five-tissue segmentation task, with an average Hausdorff Distance of 0.21, compared to the runner-up's 0.36.
GRACE segmented a whole-head MRI in approximately 3 seconds, significantly faster than the fastest existing tool at 3 minutes.
The study utilized the largest manually corrected dataset to date for MRI segmentation.

Conclusions

GRACE offers a highly accurate and rapid solution for segmenting diverse tissue types from older adults' T1-MRI scans.
The optimized model for older adult heads supports high-precision computational modeling in age-related brain disorders.
The GRACE code and trained weights are publicly available to promote open science and further research.

Keywords:

Artificial Intelligence Deep learning MRI Non-invasive brain stimulation Whole-head segmentation