New Deformable Image Registration Methods Based on Object Edge Curves

Abstract

Image registration is a field widely used in artificial intelligence including medical image processing. Radiology is one of many disciplines in medicine that uses the benefit of the image registration in the treatment process and responsible for radiation therapy, that when applied through multiple treatment fractions, there are changes occurred not only on the lesion but on the normal organ as well. In this reason, the radiation dose is closely assessed, verified, and monitored between each fraction. Additionally, the deformable image registration is used to register the different time of the acquisition. The product of the registration is used to deform the radiation dose before the dose accumulation. Various methods have been proposed to solve the problem, but the limitation still remains. This study proposes new methods of the deformable image registration based on the edge curves. The edge curve of the object is used as the information of image registration provides more information and preserves the accuracy for image registration. The polynomial regression is used to determine the coefficients. The conversion to the Bspline equation is used to reduce the members of the coefficient. The alignment between the reference and source curves shows good performance of the proposed methods. For the registration, the deformable models are generated by the Gaussian function. The function is modified to create various types of the kernel. These kernels are connected to be the Gaussian line. Later, the Gaussian lines are combined. The pixels in an image are clustered with respect to the reference edge curve by using the nearest neighbor method.The deformed image is reconstructed by the multiplication between the combinations of Gaussian lines and the pixel clustered image. The performance of the proposed methods is compared to state-of-the-art image registration like the Free Form Deformation (FFD), Optical Flow (OF), Demons (DM) and Level Set (LS). The datasets of binary, grayscale, and medical images are used for the performance evaluation. The Normalized Cross Correlation (NCC) coefficient, Hausdorff Distance (HD) and Dice Similarity (DS) are used as the quantitative measurement of similarity. The evaluated curves are delineated by an expert and measured by HD and DS. The results show the proposed methods provide the accuracy against the state-of-the-art methods, but are comparable to the feature-based FFD in the dataset of multiple image modalities and inferior to OF in the dataset of different breathing cycle.