Abstracts

Rationale: Hippocampal atrophy is a potentially localizing feature in temporal lobe epilepsy, and robust, quantitative measures of hippocampal volume from MRI are diagnostically useful. Traditional manual hippocampal tracing is time consuming and costly, but automated atlas-based segmentation algorithms, if accurate and robust, could expedite these measurements.  Methods: We retrospectively identified 50 patients with varying degrees of hippocampal asymmetry noted in clinical records. Four patients with notable anatomic deformations were included to assess segmentation robustness. Manual tracing of hippocampi was performed by experienced technologists on 3T MPRAGE images, which measured hippocampal volume up to the tectal plate. The same images were automatically processed using NeuroQuant (NQ) volumetric analysis software (Cortechs Labs Inc.), and open-source FreeSurfer (FS) software (Vers 4.2-dev5). Left and right raw hippocampal volumes, and the volume asymmetry index (AI) were statistically compared.Ten subjects were chosen from the study cohort and had repeated manual hippocampal tracings by two additional technologists blind to previous results. A group of ten patients with multiple NQ protocol MRIs on different scanners was identified, and these images were processed using NQ and FS.  Results:  In total 70 segmentation attempts were performed by each of the three methods. Manual tracing was successfully completed in all cases. NQ segmentation failed to complete in the four cases with anatomic abnormalities, and FS segmentation failed in two. On review, segmentation errors were visually identified in 25 NQ and 27 FS segmentations.Manual tracing produced significantly smaller left and right hippocampal volumes than either of the automated methods (pIntraclass correlation coefficients (ICC) for the group of repeated segmentations were: Manual Tracing 0.564 (LH),  0.578 (RH), 0.982 (AI); NeuroQuant 0.935 (LH), 0.965 (RH), and 0.499 (AI); FreeSurfer 0.937 (LH), 0.852 (RH), 0.758 (AI). ICC values < 0.4 = poor reproducibility, 0.4 - 0.75 = fair, > 0.75 = excellent. Conclusions: Absolute hippocampal volume measurements vary significantly with the segmentation method used with manual being least reproducible.  In the presence of anatomic abnormalities, automated algorithm approaches can potentially fail , while human tracing segmentation remains robust . Algorithm-derived hippocampal volumes produced high reproducibility in absolute hippocampal volumes but poorer reproducibility in asymmetry index measurements.  Overall these results show the importance of visual review of automatically segmented images and highlight the need for improvements in segmentation algorithm accuracy and reliability.  Funding: This study was supported by the Mayo Clinic and a gift from Mr. and Mrs. David Hawk.