Surface-based multi-template automated hippocampal segmentation: Application to temporal lobe epilepsy

Abstract

In drug-resistant temporal lobe epilepsy (TLE), detecting hippocampal atrophy on MRI is crucial as it allows defining the surgical target. In addition to atrophy, about 40% of patients present with malrotation, a developmental anomaly characterized by atypical morphologies of the hippocampus and collateral sulcus. We have recently shown that both atrophy and malrotation impact negatively the performance of volume-based techniques. Here, we propose a novel hippocampal segmentation algorithm (SurfMulti) that integrates deformable parametric surfaces, vertex-wise modeling of locoregional texture and shape, and multiple templates in a unified framework. To account for inter-subject variability, including shape variants, we used a library derived from a large database of healthy (n = 80) and diseased (n = 288) hippocampi. To quantify malrotation, we generated 3D models from manual hippocampal labels and automatically extracted collateral sulci. The accuracy of SurfMulti was evaluated relative to manual labeling and segmentation obtained through a single atlas-based algorithm (FreeSurfer) and a volume-based multi-template approach (Vol-multi) using the Dice similarity index and surface-based shape mapping, for which we computed vertex-wise displacement vectors between automated and manual segmentations. We then correlated segmentation accuracy with malrotation features and atrophy. SurfMulti outperformed FreeSurfer and Vol-multi, and achieved a level of accuracy in TLE patients (Dice = 86.9%) virtually identical to healthy controls (Dice = 87.5%). Vertex-wise shape mapping showed that SurfMulti had an excellent overlap with manual labels, with sub-millimeter precision. Its performance was not influenced by atrophy or malrotation (|r| < 0.20, p > 0.2), while FreeSurfer (|r| > 0.35, p < 0.0001) and Vol-multi (|r| > 0.28, p < 0.05) were hampered by both anomalies. The magnitude of atrophy detected using SurfMulti was the closest to manual volumetry (Cohen’s d: manual = 1.71, t = 7.6; SurfMulti = 1.60, t = 7.0; Vol-multi = 1.38, t = 6.1; FreeSurfer = 0.91, t = 3.9). The high performance of SurfMulti regardless of cohort, atrophy and shape variants identifies this algorithm as a robust segmentation tool for hippocampal volumetry.

Introduction

Temporal lobe epilepsy (TLE) is the most frequent form of drug-resistant focal epilepsy. Hippocampal sclerosis, the histopathological hallmark of this condition, generally appears as atrophy on MRI (Bernasconi, 2006, Cascino, 2008, Cendes et al., 1993, Jackson et al., 1990). Volumetry is more sensitive in detecting hippocampal sclerosis than visual evaluation and allows defining the surgical target in the majority of patients (Bernasconi et al., 2003, Jackson et al., 1993, Kuzniecky et al., 1999). Removing the diseased hippocampus is an effective treatment, offering seizure freedom in a large proportion of patients (Cascino, 2004, Schramm and Clusmann, 2008). In addition to atrophy, about 40% of TLE patients show atypical shape and positioning of the hippocampus (Bernasconi et al., 2005, Voets et al., 2011). These features, commonly referred to as malrotation, are considered markers of neurodevelopmental anomalies (Baulac et al., 1998, Voets et al., 2011) and may contribute to the pathogenesis of this condition (Blumcke et al., 2002, Sloviter et al., 2004). They are mainly characterized by a rounder appearance of and vertical orientation of the hippocampus, and an abnormally deep collateral sulcus (Baulac et al., 1998, Bernasconi et al., 2005).

Given the clinical utility of volumetry in defining the side of the epileptic focus, automated hippocampal segmentation may constitute a valuable tool in the presurgical evaluation in TLE. However, to date performance in patients has been rather unsatisfactory (Akhondi-Asl et al., 2011, Avants et al., 2010, Chupin et al., 2009b, Hammers et al., 2007, Heckemann et al., 2010, Pardoe et al., 2009). We recently evaluated the influence of hippocampal malrotation on three state-of-the-art automatic segmentation algorithms (Kim et al., 2012) including a region growing approach that utilizes rule-based detection of anatomical landmarks (Chupin et al., 2009b), an algorithm based on the nonlinear warp of a target image to a probabilistic atlas (Fischl et al., 2002), and a multi-template approach (Collins and Pruessner, 2010). While the overall performance of the multi-template method was superior to the others, accuracy and clinical utility of all three algorithms were affected by malrotation. The presence of abnormal anatomical variants altering the morphology of the hippocampus likely modifies its spatial relationship with surrounding structures, leading algorithms that rely on template or prior-knowledge based on healthy subjects to fall into local minima (Chupin et al., 2009b, Khan et al., 2008, Pardoe et al., 2009). On the other hand, by building prior shape models and selecting automatically those that best fit the structure to segment, methods based on multi-template libraries and label fusion (Aljabar et al., 2009, Collins and Pruessner, 2010, Heckemann et al., 2006) have the potential to overcome the limitations of individual or averaged template techniques. However, the non-linear image registration may fail in case of atypical morphology (Kim et al., 2012).

Besides volume-based segmentation, the use of non-parametric deformable models based on level-set formulations have allowed for flexible deformations against morphological and topological variations (Leventon et al., 2000, Yang and Duncan, 2004). As this approach does not guarantee point-wise correspondence, sampling of local texture and shape at the boundary is not straightforward. Conversely, parametric models permit vertex-wise sampling (Kelemen et al., 1999, Klemencic et al., 2004, Pitiot et al., 2004). Nevertheless, both methods have provided so far relatively poor results in healthy controls, likely due to the initialization step through a single average surface or a seed point, which may not sufficiently account for shape variations of the structure to segment.

In this paper, we propose a novel hippocampal segmentation algorithm that integrates deformable parametric surfaces, vertex-wise modeling of locoregional texture and shape features, and multiple templates in a unified framework. We compared the performance of our method (henceforth named SurfMulti) to manual tracing and segmentation obtained through a volume-based single template (Fischl et al., 2002) and multi-template approach (Collins and Pruessner, 2010). This work is an extension of our previously published methodology (Kim et al., 2011). Here, we evaluated the contribution of single features to the performance of SurfMulti. In addition, we evaluated the impact of hippocampal malrotation and atrophy on the global performance of the automated algorithms and assessed local segmentation accuracy using surface-based shape analysis. Finally, we investigated the ability of each automated method to lateralize the seizure focus.