Multi-Template Mesiotemporal Lobe Segmentation: Effects of Surface and Volume Feature Modeling

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Neurology

Background:

• Neurological disorders often involve atrophy of mesiotemporal lobe structures, including the hippocampus (HP), amygdala (AM), and entorhinal cortex (EC).
• Accurate segmentation of these structures is crucial for understanding disease progression and guiding patient management.
• Existing segmentation methods, particularly multi-template approaches, show varying performance based on their reliance on surface or volume data and template library size.

Purpose of the Study:

• To compare the performance of surface-based and volume-based segmentation approaches for mesiotemporal lobe structures.
• To develop and evaluate a novel hybrid algorithm (HybridMulti) that combines surface and volume similarity measures for improved template selection and segmentation accuracy.
• To assess the impact of template library size and pathological conditions on segmentation performance.

Main Methods:

• Development of the HybridMulti algorithm, integrating surface- and volume-derived similarity measures for optimal template selection.
• Introduction of non-linear registration driven by volume intensities and template surface features.
• Implementation of shape averaging using multi-scale regional weighting and icosahedron sampling.

Main Results:

• HybridMulti achieved high segmentation accuracy in healthy controls (HP/AM/EC Dice: 89.7/89.3/82.9%) and epilepsy patients (88.7/89.0/82.6%).
• Performance remained consistent across 1.5T and 3T MRI datasets.
• The algorithm outperformed existing surface- or volume-based multi-template methods, demonstrating superior accuracy and sensitivity to epilepsy-related atrophy.
• HybridMulti maintained accuracy even with a reduced template library (50% size), unlike purely surface-based methods.

Conclusions:

• The HybridMulti algorithm offers a robust and accurate method for segmenting mesiotemporal lobe structures, crucial for neurological disorder research.
• Its hybrid approach enhances template selection and registration, leading to improved performance compared to existing methods.
• The algorithm's resilience to template library size and its sensitivity to atrophy make it a valuable tool for clinical applications and disease research.