Quantifiable brain atrophy synthesis for benchmarking of cortical thickness estimation methods

Area of Science:

• Neuroimaging
• Biomarker Discovery
• Machine Learning in Medicine

Background:

• Cortical thickness (CTh) is a key biomarker for neurodegenerative diseases, typically measured using MRI software.
• Existing CTh estimation methods lack sensitivity to subtle, sub-voxel atrophy and a framework for region-wise evaluation.
• Machine learning approaches offer faster CTh estimation but require rigorous validation.

Purpose of the Study:

• To develop a brain MRI synthesis technique for generating quantifiable sub-voxel atrophy levels.
• To create a synthetic dataset for evaluating and comparing CTh estimation methods.
• To assess the sensitivity of different CTh estimation techniques to subtle neurodegenerative changes.

Main Methods:

• Proposed a novel MRI synthesis method to simulate global and local sub-voxel atrophy.
• Generated a synthetic dataset with controlled atrophy levels.
• Evaluated four CTh estimation methods: FreeSurfer (cross-sectional and longitudinal), DL+DiReCT, and HerstonNet.

Main Results:

• The synthetic dataset effectively benchmarks CTh estimation methods at global and local scales.
• DL+DiReCT exhibited superior sensitivity to sub-voxel atrophy compared to FreeSurfer and HerstonNet.
• The framework allows for regional inter- and intra-method performance comparisons.

Conclusions:

• The developed MRI synthesis approach provides a robust platform for evaluating CTh estimation tools.
• DL+DiReCT shows promise for sensitive detection of early neurodegeneration.
• This work facilitates more accurate benchmarking and comparison of CTh estimation software.