No abstract

Abstract. This paper presents a method to automatically assess the accuracy of image registration. It is applicable to images in which vessels are the main landmarks such as fundus images and angiography. The method simultaneously exploits not only the position, but also the intensity profile across the vasculatures. The accuracy measure is defined as the energy of the odd component of the 1D vessel profile in the difference image divided by the total energy of the corresponding vessels in the constituting images. Scale and orientation-selective quadrature filter banks have been employed to analyze the 1D signal profiles. Subsequently, the relative energy measure has been calibrated such that the measure translates to a spatial misalignment in pixels. The method was validated on a fundus image dataset from a diabetic retinopathy screening program at the Rotterdam Eye Hospital. An evaluation showed that the proposed measure assesses the registration accuracy with a bias of -0.1 pixels and a precision (standard deviation) of 0.9 pixels. The small Fourier footprint of the orientation selective quadrature filters makes the method robust against noise.