Abstract:

BACKGROUND: Traditional mutual information (MI) based multimodal medical image registration evaluates probability density function by two-dimensional histogram or Parzen-window function, and then plug this estimation into the expression of MI. In this process, it just considers images intensity, but ignore spatial information of image, which leading to image mismatch.
OBJECTIVE: To propose a new robust and fast registration method based on amplitude modulation-frequency modulation (AM-FM).
METHODS: Firstly, we decomposed the image using AM-FM model so as to obtain the AM-FM features. Therefore, we got a series of high dimension features consisting of AM-FM features and intensity features. Finally, we computed mutual information using these high dimension features and entropy graph. It was an extension to the mutual information framework which incorporated spatial information about image structure into the registration process and had the potential to improve the accuracy and robustness of image registration, entropy graph was used to compute the mutual information.
RESULTS AND CONCLUSION: In the experiment, we used 20 sets of data: T1-T2 weighted images and CT-PET images to compare the proposed method with MI. The results indicate that this algorithm is a more robust for image registration than conditional mutual information, even if the image in the context of low resolution and noise. This method is more accurate and robust while remaining computationally efficient. Hence, it is particularly suited to the clinical application.