Comparison of 10 brain tissue segmentation methods using revisited IBSR annotations.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate brain tissue segmentation is crucial for neurological research and clinical applications.
• Existing segmentation methods may exhibit performance bias due to inconsistent ground-truth annotations.
• The Internet Brain Segmentation Repository (IBSR) dataset's ground truth includes Sulcal cerebrospinal fluid (SCSF) voxels as gray matter, potentially affecting accuracy assessments.

Purpose of the Study:

• To evaluate the impact of Sulcal cerebrospinal fluid (SCSF) voxel inclusion in ground-truth annotations on the accuracy of 10 brain tissue segmentation methods.
• To compare the performance of various segmentation algorithms under different ground-truth labeling conditions.

Main Methods:

• Ten brain tissue segmentation methods (FAST, SPM5, SPM8, GAMIXTURE, ANN, FCM, KNN, SVPASEG, FANTASM, PVC) were evaluated.
• Performance was assessed using original IBSR ground-truth and subsequently with SCSF voxels excluded.
• Pairwise comparisons and permutation tests were employed for method ranking.

Main Results:

• Segmentation accuracy, measured by Dice coefficient, was affected by SCSF annotation changes for all tested methods, particularly SPM5, SPM8, and FAST.
• When SCSF voxels were excluded, SVPASEG (0.90 ± 0.01) and SPM8 (0.91 ± 0.01) showed superior gray matter segmentation.
• FAST (0.89 ± 0.02) demonstrated the best performance for white matter segmentation after excluding SCSF.

Conclusions:

• The inclusion or exclusion of SCSF voxels in ground-truth data significantly alters the performance and accuracy of brain tissue segmentation methods on IBSR images.
• The notable accuracy changes observed for common methods like FAST, SPM5, and SPM8 underscore the importance of addressing SCSF labeling inconsistencies in future comparative studies.