Evaluating Traditional, Deep Learning and Subfield Methods for Automatically Segmenting the Hippocampus From MRI.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Hippocampal atrophy correlates with cognitive impairment in various diseases.
• Manual hippocampus segmentation from MRI is crucial but laborious and prone to errors.
• Automatic segmentation methods are needed, but direct comparisons are lacking.

Purpose of the Study:

• To independently evaluate and compare the performance of 10 automatic hippocampus segmentation methods.
• To assess traditional, deep learning-based, and subfield segmentation techniques.
• To identify strengths and weaknesses of current automatic segmentation approaches.

Main Methods:

• Evaluated 10 automatic methods (FreeSurfer, SynthSeg, FastSurfer, FIRST, e2dhipseg, Hippmapper, Hippodeep, FreeSurfer-Subfields, HippUnfold, HSF).
• Utilized 3 datasets with manual hippocampus segmentations for ground truth.
• Assessed performance using overlap metrics, volume correlations, similarity, diagnostic differentiation, and error analysis.

Main Results:

• Most methods, particularly deep learning models trained on manual labels, performed well on public datasets.
• Methods exhibited increased error and variability when applied to clinical datasets.
• Over-segmentation, especially at the anterior hippocampus, was a common issue, yet diagnostic group differentiation was often successful.

Conclusions:

• Current automatic hippocampus segmentation methods face challenges, particularly with clinical data variability.
• Deep learning approaches show promise but require further refinement for robust clinical application.
• There is a need for more diverse, manually labeled datasets to improve and validate segmentation algorithms across various populations and pathologies.