Geometric Connection Between Machine Learning And Image Registration

Image registration is a critical step in an enormous variety of problems in medical image computing. The field of medical imaging poses many challenges for image computing and the analysis of visual information. This is true for both practical medicine and theoretical science. In particular, non-invasive human neuroimaging has been extensively developed in recent decades and now produces data that are expected to shed the light on how the human brain works.This is why it is now extremely important to develop algorithms for the analysis of these very specific data. Our study addresses one of the very important questions of medical image computing, namely how to align (register) one image of an organ (e.g., brain) to another similar image. In the area of neuroscience, image registration is needed both to match scans of the same person taken with time steps (e.g., in case of functional MRI) and to compare brain scans of different individuals or even groups (e.g., healthy versus diseased). The affine transformation is known to produce poor results in this task. More sophisticated methods, for example, Large Deformation Diffeomorphic Metric Mapping (LDDMM) enable more accurate registration while lacking reasonable automatic methods for selecting the parameters. We propose an iterative procedure of optimizing LDDMM model parameters by a gradient descend within the supervised settings. Based on the synthetic dataset, we run three sets of experiments. We first confirm that the LDDMM with standard parameter selection procedure outperforms