A study on medical image registration by mutual information with pyramid data structure

https://doi.org/10.1016/j.compbiomed.2006.02.005Get rights and content

Abstract

The registration method based on mutual information is currently a popular technique for medical image registration, but the computation of the mutual information is complex and the registration speed is slow. In this work, a new slice accumulation pyramid (SAP) data structure was proposed to expedite the registration process. A numerical comparative study between the new data structure and the existing wavelet pyramid (WP) data structure was given, and the results confirmed that the new pyramid data structure was superior to the WP in both the calculation efficiency and the optimizing performance. Finally, SAP was applied to remove the artifacts between CT and MRI data sets, and the results showed the validation of SAP to registration of mulmodality images.

Introduction

Medical image registration is a key technique in image processing and visualization, and it arises and develops with the expanded applications of medical imaging techniques. The registration goal is to combine the information in different modality images or to get rid of the artifacts in images, and it is now an indispensable technique for disease diagnosis and neuroscience research [15]. There are diverse medical image registration methods, including the traditional methods such as landmark based registration [1], [2], [14], [19], [23], feature based registration [3], [22], and some relatively new information based methods such as the one based on mutual information (MI) [4], [5], [17], [20]. MI can be taken to measure the similarities between two image sets and for convenience of calculation and analysis, the normalized mutual information (NMI) is usually adopted in image registration [16], [24]. In general, the accuracy and speed of registration are the two key criteria to evaluate whether a registration algorithm is practically used. Accuracy of sub-pixel can be achieved by registration with MI, and what prevents MI registration from universal clinical use is its huge calculation complexity and low registration speed [6], [5]. In recent years, such techniques as sub-sampling and wavelet pyramid (WP) data structure [7], [6] have been proposed to lower the calculation complexity and accelerate the registration speed. WP is more effective and of higher accuracy for registration than sub-sampling. For detail about sub-sampling one could refer to [8], [9], [10], [11], and in this paper, we mainly discuss the performance of the pyramid data structure used in image registration.

An image registration procedure based on pyramid data structure could be described as follows [21]. First, raw image data is processed to generate the pyramid grade data, which possesses more and more original information from the top to the bottom of the data structure. Second, more and more accurate result is achieved by subsequent registration on each pyramid grade from the top to the bottom with the registration result of the prior grade as the initial values of the current grade. In this paper, proposed is a new data structure called slice accumulation pyramid (SAP). Compared with the WP, SAP is generated by accumulation of slices and can totally keep the detail information of image at each grade, which avails for registration.

The details of WP and SAP was given in Section 2. A comparison study between these two data structures was conducted in Section 3. In Section 4, SAP was applied for the registration of two image sets with different modalities, CT and MRI. Finally, discussion concluded this paper in Section 5.

Section snippets

Methods

The detail of mutual information calculation is shown in the Appendix A; here the WP and SAP structures compared in the paper are shown in detail in the following.

Normalized MI distributions of WP and SAP

In this experiment, the data was a 3D magnetic resonance image (MRI) data set composed of 35 2D slices. We took the original data as the reference image set, and constructed the float one by shifting the reference set 25 and 18 pixels in X and Y directions, respectively. We added that for clarity and convenience for the results display and comparison, we only considered the above case with shifts in two directions and other cases had the similar results. All the experiments were implemented on

Application

One of the medical image registration purposes is to incorporate the information of multimodality images. CT and MRI are two widely applied imaging techniques in clinics for disease diagnosis and therapies such as tumor position seeking and cancer therapy, etc. Here, we would show the result of our method for registration between CT and MRI sets with artifacts. For the medical image registration, we could take the common obvious features in multimodality images as the makers for the qualitative

Discussion

In medical image registration, the adoption of pyramid structure can lower the computation complexity and accelerate the registration speed. The popular WP is based on a low-pass filter, which removes the high frequency component and only takes the remaining low frequency component as the approximation to the original data. The new SAP constructs the pyramid structure by accumulation in slices without sampling in X and Y, which means that no information is neglected on slices. So the WP may

Peng Xu received his B.S degree in Biomedical Engineering from the UESTC in 2000, and presently he is a Ph.D candidate in the UESTC. His research interest is the biological signal processing.

References (26)

  • H.M. Chen, P.K. Varshney, A pyramid approach for multimodality image registration based on mutual information, in:...
  • F. Maes et al.

    Multimodality image registration by maximization of mutual information

    IEEE Trans. MI

    (1997)
  • D.H. Laidlaw et al.

    Partial-volume Bayesian classification of material mixtures in MR volume data using voxel histograms

    IEEE Trans. MI

    (1998)
  • Cited by (23)

    • Volume sweeping and bodyline matching for automated prealignment in volumetric medical image registration

      2012, Computers in Biology and Medicine
      Citation Excerpt :

      For example, the MR-MR (magnetic resonance) registration has a larger capture range than the MR-single photon emission computed tomography (SPECT) registration. The results in [20] did not take into account the multi-resolution (coarse-to-fine) optimization strategy that is frequently used in implementation [21]. With that strategy, a capture range in the order of a few tens of millimeters is reported [19,22].

    • Atlas-based semiautomatic target volume definition (CTV) for head-and-neck tumors

      2010, International Journal of Radiation Oncology Biology Physics
      Citation Excerpt :

      Different and more complex methods regarding automatic registration for generation of the target volume have been proposed, based on statistical image information. However, our prerequisite of obtaining the target volume proposal in less than 3 minutes, for the sake of a low contouring time, is directly relevant to daily clinical practice and renders these methods (20–23) prohibitive in terms of computation time. In a recent study Zhang et al.(19) have shown a reduction in computation time down to 10 to 15 minutes by optimizing the hardware for a nonlinear image registration.

    • An automatic image registration scheme using Tsallis entropy

      2010, Biomedical Signal Processing and Control
    • Preoperative CT and Intraoperative CBCT Image Registration and Evaluation in Robotic Cochlear Implant Surgery

      2022, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    • Efficient approach of cardiac catheterization image enhancement

      2019, Journal of Theoretical and Applied Information Technology
    View all citing articles on Scopus

    Peng Xu received his B.S degree in Biomedical Engineering from the UESTC in 2000, and presently he is a Ph.D candidate in the UESTC. His research interest is the biological signal processing.

    DeZhong Yao received his Ph.D degree in Applied Geophysics from the Chengdu University of Technology, Chengdu, China, in 1991, and completed his postdoctoral fellowship in electromagnetic field at the UESTC in 1993. He has been a faculty member (1993), a Professor (1995) and the Dean of the School of Life Science and Technology of the UESTC (2001.1), Chengdu, China. He had been a visiting scholar at the University of Illinois at Chicago, USA, from September 1997 to August 1998, and a Visiting Professor at, the CMcMaster Universityanada, from November 2000 to May 2001, and at the Aalborg University, Denmark, from November 2003 to February 2004. His current interests include EEG, fMRI and their applications in cognitive science and neurological problems.

    Supported by NSFC #90208003 and #30200059; the 973 Project no. 2003CB716106; Key research project of science and technology of MOE, China; Sichuan youth Researcher Foundation, Doctor Training Fund of MOE, China; TRAPOYT.

    View full text