Fully automatic deep-learning based segmentation of hippocampal subfields from standard resolution T1W MRI in Alzheimer’s disease

Here we train an AI method to segment the hippocampus into subfields (CA1-3 CA4+DG, and Subiculum) from standard resolution T1W MRI alone to assess utility as biomarkers compared to whole hippocampal volume, in the absence of a high-resolution T2 MRI.

The method presented here enables the extraction of hippocampal subfield volumes from routinely acquired data, providing additional biomarkers that may not be possible for ongoing collection where patient burden prohibits collection of an additional high-resolution T2W MRI or for retrospective data without a high-resolution T2.

In the absence of a high-resolution T2W MRI, hippocampal subfields volumes, as estimated from T1W MRI, can provide an informative measure of disease progression better than the whole hippocampus alone. The subiculum and dentate gyrus reported the greatest sensitivity to group separation and correlation to cognitive scores, with the subiculum notably reporting sensitivity comparable to that of the ASHS method. Further development is required with validation on additional datasets.

Date: 26/07/2023