Down‐sampling template curve to accelerate LDDMM‐curve with application to shape analysis of the corpus callosum

Abstract

Abstract Large deformation diffeomorphic metric mapping for curve (LDDMM‐curve) has been widely used in deformation based statistical shape analysis of the mid‐sagittal corpus callosum. A main limitation of LDDMM‐curve is that it is time‐consuming and computationally complex. In this study, down‐sampling strategies for accelerating LDDMM‐curve are investigated and tested on two large datasets, one on Alzheimer's disease (155 Alzheimer's disease, 325 mild cognitive impairment and 185 healthy controls) and the other on first‐episode schizophrenia (92 first‐episode schizophrenia and 106 healthy controls). For both datasets a variety of down‐sampling factors are tested in terms of registration accuracy, registration speed, and most importantly disease‐related patterns. Experimental results revealed that down‐sampling template curve by a factor of 2 can significantly reduce the running time of LDDMM‐curve without sacrificing the registration accuracy. Also, the disease‐induced patterns, or more specifically the group comparison results, were almost identical before and after down‐sampling. It is also shown that there was no need to down‐sample the target population curves but only the single template curve of the study of interest. Comprehensive analyses were conducted.