Testing a convolutional neural network‐based hippocampal segmentation method in a stroke population

As stroke mortality rates decrease, there has been a surge of effort to study poststroke dementia (PSD) to improve long‐term quality of life for stroke survivors. Hippocampal volume may be an important neuroimaging biomarker in poststroke dementia, as it has been associated with many other forms of dementia. However, studying hippocampal volume using MRI requires hippocampal segmentation. Advances in automated segmentation methods have allowed for studying the hippocampus on a large scale, which is important for robust results in the heterogeneous stroke population. However, most of these automated methods use a single atlas‐based approach and may fail in the presence of severe structural abnormalities common in stroke. Hippodeep, a new convolutional neural network‐based hippocampal segmentation method, does not rely solely on a single atlas‐based approach and thus may be better suited for stroke populations. Here, we compared quality control and the accuracy of segmentations generated by Hippodeep and two well‐accepted hippocampal segmentation methods on stroke MRIs (FreeSurfer 6.0 whole hippocampus and FreeSurfer 6.0 sum of hippocampal subfields). Quality control was performed using a stringent protocol for visual inspection of the segmentations, and accuracy was measured as volumetric correlation with manual segmentations. Hippodeep performed significantly better than both FreeSurfer methods in terms of quality control. All three automated segmentation methods had good correlation with manual segmentations and no one method was significantly more correlated than the others. Overall, this study suggests that both Hippodeep and FreeSurfer may be useful for hippocampal segmentation in stroke rehabilitation research, but Hippodeep may be more robust to stroke lesion anatomy.