Out‐of‐the box automated computation of cortical thickness of the medial perirhinal cortex does not replace segmentation by hand

Abstract

AbstractBackgroundCortical thickness of the medial perirhinal cortex (mPRC) is a promising marker of early, potentially pre‐clinical Alzheimer’s disease (AD) because atrophy of this region is causally linked to intracellular neurofibrillary tangles. Neurofibrillary pathology, a major neuropathological characteristic of AD, typically starts in the mPRC, also referred to as the transentorhinal cortex. From there it spreads to the entorhinal cortex and eventually to the hippocampus. To date, no tools exist that automatically compute the cortical thickness of the mPRC, most likely because of the novelty and complexity of the manual segmentation protocol and the large variability of the morphology of the mPRC. Herein, we aimed to develop an automated segmentation algorithm for the mPRC.MethodAn experienced rater (SK) manually segmented the mPRC from two data sets (train: n = 103; test: n = 44) using the FreeSurfer segmentation tools. The train data set (mean age = 76.4 ± 7.0 years; normal controls and patients with AD) was used to train the native FreeSurfer segmentation algorithm with the custom region as well as a voxel‐based deep convolutional neural network based on nnUNet with minimal adaptations to the training protocol. In a next step, the trained models were applied to the test data set (mean age = 69.2 ± 10.4 years; normal controls, AD, major depression, and mild cognitive impairment or dementia due to other etiologies than AD) and cortical thickness values were extracted. Finally, cortical thickness values of both segmentations were compared individually to the manual segmentation.ResultsIntraclass correlation coefficients (ICCs) between FreeSurfer and the manual segmentation reached .768 for the left hemisphere (lh) and .835 for the right hemisphere (rh). The deep learning method compared to the manual segmentation achieved ICCs of .840 for lh and .896 for rh.ConclusionThe deep learning based procedure outperformed the native FreeSurfer segmentation in this complicated segmentation task. Nevertheless, it did not reach inter‐human reliability (SK vs. unexperienced rater (NAH): lh = .953 and rh = .986). Next steps are customization of the deep learning algorithm and validation of clinical usefulness.