Geodesic shape regression based deep learning segmentation for assessing longitudinal hippocampal atrophy in dementia progression

Abstract

Longitudinal hippocampal atrophy is commonly used as progressive marker assisting clinical diagnose of dementia. However, precise quantification of the atrophy is limited by longitudinal segmentation errors resulting from MRI artifacts across multiple independent scans. To accurately segment the hippocampal morphology from longitudinal 3T T1-weighted MR images, we propose a diffeomorphic geodesic guided deep learning method called the GeoLongSeg to mitigate the longitudinal variabilities that unrelated to diseases by enhancing intra-individual morphological consistency. Specifically, we integrate geodesic shape regression, an evolutional model that estimates smooth deformation process of anatomical shapes, into a two-stage segmentation network. We adopt a 3D U-Net in the first-stage network with an enhanced attention mechanism for independent segmentation. Then, a hippocampal shape evolutional trajectory is estimated by geodesic shape regression and fed into the second network to refine the independent segmentation. We verify that GeoLongSeg outperforms other four state-of-the-art segmentation pipelines in longitudinal morphological consistency evaluated by test–retest reliability, variance ratio and atrophy trajectories. When assessing hippocampal atrophy in longitudinal data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), results based on GeoLongSeg exhibit spatial and temporal local atrophy in bilateral hippocampi of dementia patients. These features derived from GeoLongSeg segmentation exhibit the greatest discriminatory capability compared to the outcomes of other methods in distinguishing between patients and normal controls. Overall, GeoLongSeg provides an accurate and efficient segmentation network for extracting hippocampal morphology from longitudinal MR images, which assist precise atrophy measurement of the hippocampus in early stage of dementia.