Hippocampus Localization Using a Two-Stage Ensemble Hough Convolutional Neural Network

In this paper, we present a two-stage ensemble-based approach to localize the anatomical structure of interest from magnetic resonance imaging (MRI) scans. We combine a Hough voting method with a convolutional neural network to automatically localize brain anatomical structures such as the hippocampus. The hippocampus is one of the regions that can be affected by the Alzheimer's disease, and this region is known to be related to memory loss. The structural changes of the hippocampus are important biomarkers for dementia. To analyze the structural changes, accurate localization plays a vital role. Furthermore, for segmentation and registration of anatomical structures, exact localization is desired. Our proposed models use a deep convolutional neural network (CNN) to calculate displacement vectors by exploiting the Hough voting strategy from multiple 3-viewpoint patch samples. The displacement vectors are added to the sample position to estimate the target position. To efficiently learn from samples, we employed a local and global strategy. The multiple global models were trained using randomly selected 3-viewpoint patches from the whole MRI scan. The results from global models are aggregated to obtain global predictions. Similarly, we trained multiple local models, extracting patches from the vicinity of the hippocampus location and assembling them to obtain a local prediction. The proposed models exploit the Alzheimer's disease neuroimaging initiative (ADNI) MRI dataset and the Gwangju Alzheimer's and related dementia (GARD) cohort MRI dataset for training, validating and testing. The average prediction error using the proposed two-stage ensemble Hough convolutional neural network (Hough-CNN) models are 2.32 and 2.25 mm for the left and right hippocampi, respectively, for 65 test MRIs from the GARD cohort dataset. Similarly, for the ADNI MRI dataset, the average prediction error for the left and right hippocampi are 2.31 and 2.04 mm, respectively, for 56 MRI scans.