Evaluating the effect of multiple sclerosis lesions on automatic brain structure segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Automatic brain structure segmentation is crucial for neurological research.
• Existing methods are typically validated on healthy brains, neglecting lesion effects.
• Multiple Sclerosis (MS) causes lesions that may compromise segmentation accuracy.

Purpose of the Study:

• To evaluate the impact of MS lesions on three distinct automatic brain segmentation methods: FreeSurfer, FIRST, and multi-atlas majority voting.
• To quantitatively assess performance variations using Dice Similarity Coefficient (DSC) and volume differences.
• To identify which segmentation strategies and brain structures are most susceptible to lesion presence.

Main Methods:

• Simulation of 2174 MS lesions across 100 synthetic patient images using public databases (IBSR18, MICCAI'12).
• Quantitative comparison of segmentation accuracy (DSC, volume) between healthy and lesioned brain images.
• Analysis of lesion impact on learning-based (FreeSurfer), deformable (FIRST), and atlas-based (majority voting) methods.

Main Results:

• All three segmentation methods showed performance changes due to MS lesions, with varied effects (under/over-segmentation).
• FreeSurfer was most affected (DSC diff: -0.11 ± 0.54 to 9.65 ± 9.87), while FIRST was most robust (-2.40 ± 5.54 to 0.44 ± 0.94).
• Nucleus accumbens was most affected; thalamus and brainstem showed less variation. Lesion location impacted FIRST more than global methods.

Conclusions:

• MS lesions significantly affect automatic deep gray matter (DGM) segmentation accuracy.
• Current segmentation tools require careful consideration of lesion effects for reliable disease progression monitoring in MS.
• Further development is needed to enhance the robustness of automated segmentation in the presence of brain pathology.