Abstract:

Image registration is used to reduce movement artifacts caused by contracting heart muscle in transmembrane voltage measurements using fluorescence microscopy. The applied registration methods include Thin-Plate Splines (TPS) and Gaussian Elastic Body Splines (GEBS). Landmarks are established automatically using regional cross-correlation. Then these landmarks are filtered for meaningful correspondences by requiring a minimum correlation coefficient and clustering adjacent and identical displacements. Registration of an image sequence showing a contracting muscle is realized by spatially aligning the images at maximum contraction and at rest. For the other images the movement of the muscle is interpolated using an analytical description of the contraction of heart muscle.TPS cause amplification of displacements at the image border, while GEBS restrict landmarks influence to a local region. Over a set of 81 images GEBS are shown to register images better and more robust than TPS, which in some cases cannot reduce movements. Validation through visualization of transmembrane voltages on contracting muscle reveals that GEBS registration removes movement artifacts better than TPS. Image regions with prominent structures are successfully tackled by GEBS registration.

Abstract:

This thesis explores image post-processing methods to reduce motion in images acquired with a microscope showing contracting muscles. The registration methods used include Thin-Plate Splines (TPS), Gaussian Elastic Body Splines (GEBS), and Fluid Flow.TPS are based on point correspondences between images and are widely used in image registra- tion. While GEBS are also based on landmarks they propose a better model for the displacements in body tissue. The registration algorithms spatially align two images without user interaction. Point correspondences are established using regional cross-correlation. The large amount of landmarks is filtered based on minimal normalized correlation-coefficients, identical displacements from dif- ferent template sizes (used in regional cross-correlation), and clustering of adjacent and identical displacements. The filtering parameters are optimized based on minimization of the sum of squared intensity difference (SSD) by testing 15 parameter variations. GEBS parameters are optimized se- quentially by using a golden-section search algorithm. Fluid Flow registration serves for compari- son of a voxel intensity based registration.The advantages of GEBS are: displacements in one direction affect other directions, landmarks have a locally restricted influence, and elasticity from physical model corresponds to tissue prop- erties. Compared to TPS these benefits lead to a better registration of images showing a contract- ing heart muscle. TPS cause amplification of landmark displacements at the image borders, while GEBS restrict landmark influence to a local region. Over a set of 87 images GEBS are shown to register images more robust than TPS, which in some cases cannot reduce movements. In all cases GEBS perform equally well or better than TPS. Compared to Fluid Flow registration GEBS work comparably well.Registration of an image sequence showing a contracting muscle is realized by spatially aligning the image with largest contraction with an image at rest and then interpolating the movement of the muscle for the other images. This interpolation is done by an analytical description of the contraction of heart muscle, fitted to the voxel similarity measurement SSD, which is assumed to be proportional to muscle movement.Validation through visualization of action potentials on contracting muscle reveals that GEBS and Fluid Flow registration best match the reference image without movement. Motion artifacts remain in regions with low contrast. Image post-processing cannot correctly align images in low-contrast regions due to lack of infomation. In image regions with prominent structures GEBS registration successfully removes motion artifacts in action potentials.