Elastic registration of 2D abdominal CT images using hybrid feature point selection for liver lesions

Abstract

Abdominal CT images have distinct intensity distribution. This feature is used to correct local deformations in the image. Reference and study images are decomposed using wavelet decomposition. Global deformations are first corrected applying rigid registration by use of maximization of Mutual Information as the similarity measure at each level of registration hierarchy. Initially registered image and reference image are further elastically registered using landmark based elastic registration. Here landmarks or feature points are obtained by first intensity thresholding the images followed by boundary selection to obtain lesion boundaries and finally obtaining the centroid and convex hull points of lesions within the images. Convex hull points that lie on the boundary of lesions coupled with centroids of lesions are helpful in precisely identifying the lesions. An advantage of this is that lesions are enhanced to allow for deformations to be precisely determined. This is useful in improving diagnostic accuracy. The performance of algorithm is tested on a real case study of abdominal CT images with liver abscess. Considerable improvement in correlation coefficient and Signal to Noise ratio of the two images is observed.