Loading...

Robust Similarity Measure in Medical Image Registration

Ghaffari, Aboozar | 2015

1389 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 47187 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Fatemizadeh, Emadeddin
  7. Abstract:
  8. Image Registration is spatially alignment of two images in a wide range of applications such as remote sensing, computer assisted surgery, and medical image analysis and processing. In general, registration algorithms can be categorized as either intensity based or feature based. The feature based methods use the alignment between the extracted features in two images. The simplest feature is images intensity which is directly used in the intensity based method via similarity measure. This similarity measure quantifies the matching of two images.Similarity measure is main core of image registration algorithms. Spatially varying intensity dis-tortion is an important challenge in a wide range of image processing fields such as segmentation, face recognition, object detection and image registration. Two examples of this distortion are bias field in magnetic resonance imaging (MRI) and illumination variations in geometric images. Correlation among pixels is the main characteristic of this distortion. Similarity measures such as sum-of-squared-differences (SSD) and mutual information (MI) ignore this correlation; Hence, perfect registration cannot be achieved in the presence of this distortion. To resolve this issue, the idea is to model this distortion appropriately.
    In this thesis, we suppose that spatially varying intensity distortion has a sparse representation in transform domain, i.e. its distribution has high peak at origin and a long tail. Based on this,two similarity measures are proposed in mono-modal setting. By this consideration and the concept
    of image decomposition, this distortion is corrected, and sparse induced similarity measure (SISM) is proposed. SISM performs the correction of spatially varying intensity distortion and registration simultaneously. We also use two viewpoints of Maximum Likelihood (ML) and Robust M-estimator.Based on these views, we propose robust Huber similarity measure (RHSM) in spatial transform domain as a new similarity measure. To demonstrate robustness of RHSM, image registration is treated as a nonlinear regression problem. In this view, with minimizing Fisher information function,robust similarity measure of RHSM is introduced.Recently, the low rank matrix recovery became one of the famous tools in signal and image process-ing. As the pixels of the spatially varying intensity distortion are correlated, this correlation of pixels is modelled by a low rank matrix in this thesis for the first time. Based on this model, we compensate this distortion analytically and introduce two new similarity measures named rank-induced similarity measure (RISM) and rank-regularized SSD (RRSSD). Image registration and distortion correction are performed simultaneously in these measures.In the last part of this thesis, we model multi-modal image registration as linear regression in the presence of spatially varying intensity distortion. In fact, this proposed model is a problem of image decomposition. We show uniqueness of this decomposition with considering sparse model. Based on this model, we propose a new similarity measure named affine-sparse induced similarity measure (ASISM). This measure is a generalized CC in multi-modal setting. In summary, we propose five similarity measures of SISM, RHSM, RISM, RRSSD and ASISM for mono-modal image registration. ASISM is also a useful measure in multi-modal setting. Based on the experiments and obtained results, the proposed similarity measures achieve clinically acceptable registration results, and outperform other state-of-the-art similarity measures such as the well-known method of residual complexity
  9. Keywords:
  10. Sparse Representation ; Image Registration ; Similarity Measure ; Bias ; Regression Analysis ; Low-Rank Matrix ; Maximum Likelihood Estimation ; Matrix Rank ; Spatially Varying Intensity Distortion

 Digital Object List

 Bookmark

...see more