Signal subspace registration of time series medical imagery

Image registration is one of the crucial steps in detecting changes among the time series medical images. Due to variations in the imaging system over time, the impulse response of the imaging system, also known as its Point Spread Function (PSF), exhibits a time-varying behavior. The registration is further complicated due to the subtle coordinate changes introduced by the patient. In this work, the registration problem is approached via a spatially varying multi-dimensional adaptive filtering method that relates one image in terms of an unknown linear combination of the other image and its spatially transformed versions. Using this model, we develop a scheme, which we refer to as Signal Subspace Processing, to estimate a localized impulse response to calibrate relatively small regions. A criterion is designed to identify the localized PSF's that are not sensitive to the system noise or anatomical changes but accurately represent the spatially varying nature of the unknown miscalibration sources. Low order polynomials are used to sew the localized PSF together and construct a global spatially variant PSF. The anatomical changes between the time series images are achieved by calibrating the image with the global spatially variant PSF. Numerical experiments using MR images illustrate the effectiveness of the proposed algorithm.