Abstracts

Rationale: The most frequent drug-resistant epilepsy is temporal lobe epilepsy (TLE) related to hippocampal atrophy on MRI. Automatic hippocampal segmentation has provided unsatisfactory results [1-3], likely due to shape alterations caused by the combined effects of atrophy and atypical positioning [4]. We propose a novel surface-based segmentation method (SurfMulti) that statistically estimates texture and shape. To account for inter-subject variability, including shape variants, we used a multi-template library derived from a large database of controls and patients.Methods: Manual Labels of the hippocampus (40 healthy controls and 144 TLE patients) were converted into surface meshes using the spherical harmonic description [5]. Locoregional texture (tissue homogeneity, contrast, Gabor energy) and shape features were modeled on each hippocampal surface. We then built a template library derived from all subjects. To segment a new hippocampus, an optimal subset of training surfaces and features was selected from the library based on a similarity function. The final segmentation was obtained by evolving the averaged surface of the selected subset. We evaluated the performance of SurfMulti against the manual label using the Dice overlap index. In addition, we compared SurfMulti to two volume-based state-of-the-art single- (FreeSurfer) [6] and multi-template (Vol-multi) approaches [7]. We assessed the sensitivity of each algorithm to detect atrophy in TLE relative to controls by computing the effect size of a between group difference (Cohen's d).Results: Our method outperformed both volume-based approaches with performances in pathological hippocampi comparable to controls (Table). The sensitivity of SurfMulti to detect atrophy was similar to that of manual volumetry (Cohen's d: 1.60 vs.1.71), whereas it decreased significantly for the two other methods (1.38 and 0.91). Vertex-wise displacements between manual labels and SurfMulti segmentations yielded sub-voxel error (absolute mean error<0.51mm). The 3D individual analysis revealed that SurfMulti reliably segmented hippocampi with various shape abnormalities compared to the other methods (Figure).Conclusions: The proposed surface-based multi-template technique achieved a level of accuracy in TLE patients virtually identical to healthy controls, with a Dice index of 86.1%. Such performance has not yet been paralleled in epilepsy. Given that the hippocampal atrophy in TLE lateralizes the seizure focus and is a predictor of a favorable seizure outcome after surgery, our automated algorithm assures to be a robust surrogate tool for the time-demanding manual procedure in the presurgical evaluation. References 1. HR Pardoe et al., Epilepsia 50(12) 2009. 2. A Akhondi-Asl et al., Neuroimage 54(S1) 2011. 3. M Chupin et al., NeuroImage 46(3) 2009. 4. N Bernasconi., Brain 128(10) 2005. 5. M Styner et al., MICCAI Opensource workshop 2006. 6. B Fischl et al., Neuron 33(3) 2002. 7. DL Collins et al., Neuroimage 52(4) 2010.