Abstracts

Rationale: The most frequent drug-resistant epilepsy is temporal lobe epilepsy (TLE) associated with hippocampal atrophy on MRI. In addition to atrophy, about 40% of TLE patients show atypical shape and positioning of the hippocampus, commonly referred to as malrotation (M Baulac et al., 1998; N Bernasconi et al., 2005). Our purpose was to evaluate the impact of malrotation on the performance of two state-of-the-art automated hippocampal segmentation algorithms. Methods: We segmented automatically the hippocampus in 35 controls and 66 TLE patients using SACHA (M Chupin et al., 2009), a region-growing algorithm constrained by anatomical priors, and FreeSurfer (B Fischl et al., 2002), a freely available atlas-based software. Segmentation accuracy was evaluated relative to manual labeling using overlap indices and surface-based shape mapping, for which we computed displacement vectors between automated and manual segmentations at each vertex. We quantified hippocampal malrotation characteristics (medial positioning, vertical orientation and collateral sulcus depth) using our previously developed computational models (NL Voets et al., 2011). We then correlated segmentation accuracies with malrotation features. We also assessed the sensitivity of each method to detect atrophy in TLE relative to controls by computing the effect size of a between group difference (Cohen's d). Finally, we evaluated the ability of the automated algorithms to lateralize the seizure focus using linear discriminant analysis.Results: Segmentation performances are detailed in the Table. In cases of malrotation, SACHA over-segmented the lateral convexity of the hippocampus (r=0.61; p<0.0001), but performed relatively well in the presence of atrophy (|r|<0.34; p>0.2). The deeper the collateral sulcus, the more it comes in contact with the lateral border of the hippocampus. In this scenario, SACHA tended to erroneously include the sulcal fundus into the tracing of the hippocampus (Figure). The performance of FreeSurfer was affected by both hippocampal malrotation (r=0.57; p<0.02) and atrophy (r=0.78, p<0.0001). Compared to manual volumetry (Cohen s d=1.68), the automated procedures detected smaller effect sizes of hippocampal atrophy (SACHA: 1.1, p<0.0001; FreeSurfer: 0.9, p<0.0001) and tended to be less accurate for seizure focus lateralization in the presence of malrotation (manual: 64%; SACHA: 50%, p=0.1; FreeSurfer: 41%, p=0.05).Conclusions: The accuracy and clinical utility of the algorithms we evaluated were suboptimal in the presence of hippocampal malrotation and severe atrophy. Since a decrease in sensitivity and lateralization performance compared to manual volumetry is disadvantageous in a pre-surgical setting, such morphological anomalies should be taken into account when designing automated segmentation methods.