RESEARCH ON ROBUSTNESS OF THE MUTUAL INFORMATION SIMILARITY METRIC FOR REGISTRATION OF MEDICAL IMAGES

Abstract

For the past few years, the medical image registration technology has made rapid and significant progress, and has been extensively applied for the 2D/3D medical image processing. However, the robustness of the similarity metric in the medical image registration is rarely studied. In this paper, the mutual information-based registration technology is introduced and the concept of the robustness of the similarity metric is defined. The robustness of the mutual information similarity metric is analyzed and discussed from three aspects such as interpolation methods, image data loss and noise corruption after the linear, quadratic spline, cubic spline, quadratic B-spline, and cubic B-spline interpolations are elaborated and studied. The robustness experiments reveal that the mutual information similarity metric can obtain good robustness in the case of the use of various interpolation methods in the medical image registration; the mutual information similarity metric can also keep good robustness in the case of a slight loss of image data. However, the metric will fail to register the images on the condition that the medical images are seriously incomplete. In addition, we find, if the medical images corrupted by the salt & pepper noise, then the metric basically can succeed in aligning the medical images; but regretfully, the metric fully fail under the condition of the medical images corrupted by the Gaussian noise.