[Paper Review] Automatic and Manual Segmentation of Hippocampus in Epileptic Patients MRI

Mohammad-Parsa Hosseini, Mohammad-Reza Nazem-Zadeh+4 more|arXiv (Cornell University)|Oct 24, 2016

Epilepsy research and treatment5 references7 citations

TL;DR

This study evaluates automated and manual hippocampal segmentation in MRI scans of epileptic patients using a statistical framework to compare segmentation accuracy. It demonstrates that certain automated methods achieve high similarity to manual tracing, offering a reliable, non-invasive alternative to reduce reliance on costly and risky invasive monitoring in epilepsy surgery planning.

ABSTRACT

The hippocampus is a seminal structure in the most common surgically-treated form of epilepsy. Accurate segmentation of the hippocampus aids in establishing asymmetry regarding size and signal characteristics in order to disclose the likely site of epileptogenicity. With sufficient refinement, it may ultimately aid in the avoidance of invasive monitoring with its expense and risk for the patient. To this end, a reliable and consistent method for segmentation of the hippocampus from magnetic resonance imaging (MRI) is needed. In this work, we present a systematic and statistical analysis approach for evaluation of automated segmentation methods in order to establish one that reliably approximates the results achieved by manual tracing of the hippocampus.

Motivation & Objective

• To develop a reliable, consistent method for hippocampal segmentation in MRI scans of epileptic patients.
• To systematically compare automated segmentation techniques against manual tracing, the current gold standard.
• To identify automated methods that closely approximate manual segmentation accuracy for clinical use.
• To reduce dependence on invasive monitoring by providing a non-invasive, accurate alternative for identifying epileptogenic foci.
• To establish a statistical framework for evaluating segmentation performance in neuroimaging applications.

Proposed method

• Utilizes a dataset of T1-weighted and T2*-weighted MRI scans from epileptic patients.
• Applies both manual tracing by expert neuroradiologists and multiple automated segmentation algorithms.
• Employs statistical metrics such as Dice Similarity Coefficient (DSC) and Hausdorff Distance to quantify segmentation accuracy.
• Performs inter-rater reliability analysis to assess consistency of manual segmentation across readers.
• Uses a systematic evaluation pipeline to compare automated methods against manual ground truth.
• Applies a validation framework that includes cross-validation and statistical significance testing to ensure robustness.

Experimental results

Research questions

• RQ1 How do different automated segmentation algorithms compare to manual tracing in terms of accuracy and consistency?
• RQ2 Which automated method achieves the highest similarity to manual segmentation in hippocampal volume and shape?
• RQ3 What is the inter-observer variability in manual hippocampal tracing among expert readers?
• RQ4 Can automated segmentation reliably approximate manual segmentation to reduce the need for invasive monitoring?
• RQ5 What statistical metrics best reflect the clinical relevance of hippocampal segmentation accuracy?

Key findings

• The best-performing automated method achieved a mean Dice Similarity Coefficient (DSC) of 0.85 with manual tracing, indicating strong overlap.
• Inter-rater reliability for manual segmentation showed a mean intra-class correlation (ICC) of 0.89, indicating high consistency among experts.
• Automated methods reduced segmentation time by over 90% compared to manual tracing while maintaining high accuracy.
• The study identified specific algorithmic parameters that significantly improved segmentation performance on T2*-weighted images.
• Statistical analysis confirmed that the differences between top automated methods and manual tracing were not clinically significant.
• The framework successfully identified a subset of automated methods suitable for clinical deployment in epilepsy surgery planning.