Robust Machine Learning-Based Correction on Automatic Segmentation of the Cerebellum and Brainstem.

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Image Analysis

Background:

• Automated segmentation aids in studying large brain structures like the cerebellum and brainstem.
• Existing automated methods can produce inaccuracies and undesirable boundaries.

Purpose of the Study:

• To evaluate SegAdapter, a machine learning method, for correcting automated segmentation errors in the cerebellum and brainstem.
• To assess SegAdapter's robustness across varying training set sizes, head coil usage, and brain atrophy levels.

Main Methods:

• High-resolution T1-weighted MRI from 30 healthy controls and 10 patients with fragile X-associated tremor/ataxia syndrome were used.
• Initial segmentations by Freesurfer were manually corrected (gold standard) and automatically corrected by SegAdapter.
• SegAdapter's performance was evaluated using Dice coefficient, assessing robustness with different training set sizes, head coil types, and brain atrophy.

Main Results:

• SegAdapter significantly improved spatial overlap (Dice coefficient) from 0.956 to 0.978 for the cerebellum and 0.821 to 0.954 for the brainstem, using only 5 training scans.
• Reducing the training set to 2 scans minimally impacted Dice coefficients (≤0.002 for cerebellum, ≤0.005 for brainstem).
• SegAdapter demonstrated robustness against variations in head coil usage and brain atrophy, with minimal reduction in spatial overlap (<0.01).

Conclusions:

• Combining automated segmentation with SegAdapter's corrective learning offers an accurate and efficient solution for cerebellum and brainstem segmentation.
• This method is particularly valuable for large-scale neuroimaging studies.
• SegAdapter shows potential for segmenting other neural regions accurately.