Robust Hippocampus Localization From Structured Magnetic Resonance Imaging Using Similarity Metric Learning

Accurate demarcation of anatomical landmarks in 3D medical imaging is a safety-critical and challenging task. State-of-the-art approaches formulate landmark localization either as a classification or as a regression problem. In this study, feature classification is performed as a verification step in a cascaded Hough regression networks (HRNs) for hippocampus localization in the structured magnetic resonance images of the brain. Global and local features of the landmarks are learned with coarse prediction and fine-tuning convolutional neural networks for coarse-to-fine localization. Siamese network was trained to learn a deep metric for verifying the roughly estimated locations. Feature verification with the Siamese network drops the outlier predictions and increase the robustness in prediction. Three-view patches(TVPs) with a size of <inline-formula> <tex-math notation="LaTeX">$64\times 64\times 3$ </tex-math></inline-formula> are fed for rough estimation while the TVP sizes for Siamese-based verification and Hough regression network (HRN)-based fine-grained estimations are <inline-formula> <tex-math notation="LaTeX">$32\times 32\times 3$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$16\times 16\times 3$ </tex-math></inline-formula>, respectively. The experiment was performed on the Gwangju Alzheimer&#x2019;s and Related Dementia&#x2019;s (GARD) cohort data set. The proposed approach demonstrated better performance with the errors of 1.70&#x00B1;0.50 millimeters(mm) and 1.66&#x00B1;0.49 mm for localizing the left and right hippocampi in the GARD data set. In Alzheimer&#x2019;s Disease Neuroimaging Initiative (ADNI) data set, the observed errors were 1.79 &#x00B1; 0.83 mm and 1.55 &#x00B1; 0.61 mm for localizing left and right hippocampus, respectively. Our results are comparable to those obtained by the state-of-the-art methods.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.