外文文献资料
翻译：李睿钦  指导老师：刘文军
transform英文
Medical image registration with partial data
Senthil Periaswamy， Hany Farid
The goal of image registration is to find a transformation that aligns one image to another. Medical image registration has emerged from this broad area of research as a particularly active field. This activity is due in part to the many clinical applications including diagnosis, longitudinal studies, and surgical planning, and to the need for registration across different imaging modalities (e.g., MRI, CT, PET, X-ray, etc.). Medical image registration, however, still presents many challenges. Several notable difficulties are (1) the transformation between images can vary widely and be highly non-rigid in nature; (2) images acquired from different modalities may differ significantly in overall appearance and resolution; (3) there may not be a one-to-one correspondence between the images (missing/partial data);
and (4) each imaging modality introduces its own unique challenges, making it difficult to develop a single generic registration algorithm.
In estimating the transformation that aligns two images we must choose: (1) to estimate the transformation between a small number of extracted features, or between the complete unprocessed intensity images; (2) a model that describes the geometric transformation; (3) whether to and how to explicitly model intensity changes; (4) an error metric that incorporates the previous three choices; and (5) a minimization technique for minimizing the error metric, yielding the desired transformation.
Feature-based approaches extract a (typically small) number of corresponding landmarks or features between the pair of images to be registered. The overall transformation is estimated from these features. Common features include corresponding points, edges, contours or surfaces. These features may be specified manually or extracted automatically. Fiducial markers may also be used as features; these markers are usually selected to be visible in different modalities. Feature-based approaches have the advanta
ge of greatly reducing computational complexity. Depending on the feature extraction process, these approaches may also be more robust to intensity variations that arise during, for example, cross modality registration. Also, features may be chosen to help reduce sensor noise. These approaches can be, however, highly sensitive to the accuracy of the feature extraction. Intensity-based approaches, on the other hand, estimate the transformation between the entire intensity images. Such an approach is typically more computationally demanding, but avoids the difficulties of a feature extraction stage.
Independent of the choice of a feature- or intensity-based technique, a model describing the geometric transform is required. A common and straightforward choice is a model that embodies a single global transformation. The problem of estimating a global translation and rotation parameter has been studied in detail, and a closed form solution was proposed by Schonemann. Other closed-form solutions include methods based on singular value decomposition (SVD), eigenvalue-eigenvector decomposition and unit quaternions. One idea for a global transformation model is to use polynomials. For examp
le, a zeroth-order polynomial limits the transformation to simple translations, a first-order polynomial allows for an affine transformation, and, of course, higher-order polynomials can be employed yielding progressively more flexible transformations. For example, the registration package Automated Image Registration (AIR) can employ (as an option) a fifth-order polynomial consisting of 168 parameters (for 3-D registration). The global approach has the advantage that the model consists of a relatively small number of parameters to be estimated, and the global nature of the model ensures a consistent transformation across the entire image. The disadvantage of this approach is that estimation of higher-order polynomials can lead to an unstable transformation, especially near the image boundaries. In addition, a relatively small and local perturbation can cause disproportionate and unpredictable changes in the overall transformation. An alternative to these global approaches are techniques that model the global transformation as a piecewise collection of local transformations. For example, the transformation between each local region may be modeled with a low-order polynomial, and global consistency is enforced via some form of a smoothness constraint. The advant
age of such an approach is that it is capable of modeling highly nonlinear transformations without the numerical instability of high-order global models. The disadvantage is one of computational inefficiency due to the significantly larger number of model parameters that need to be estimated, and the need to guarantee global consistency. Low-order polynomials are, of course, only one of many possible local models that may be employed. Other local models include B-splines, thin-plate splines, and a multitude of related techniques. The package Statistical Parametric Mapping (SPM) uses the low-frequency discrete cosine basis functions, where a bending-energy function is used to ensure global consistency. Physics-based techniques that compute a local geometric transform include those based on the Navier–Stokes equilibrium equations for linear elastici and those based on viscous fluid approaches.
Under certain conditions a purely geometric transformation is sufficient to model the transformation between a pair of images. Under many real-world conditions, however, the images undergo changes in both geometry and intensity (e.g., brightness and contrast). Many registration techniques attempt to remove these intensity differences with a pre-pro
cessing stage, such as histogram matching or homomorphic filtering. The issues involved with modeling intensity differences are similar to those involved in choosing a geometric model. Because the simultaneous estimation of geometric and intensity changes can be difficult, few techniques build explicit models of intensity differences. A few notable exceptions include AIR, in which global intensity differences are modeled with a single multiplicative contrast term, and SPM in which local intensity differences are modeled with a basis function approach.
Having decided upon a transformation model, the task of estimating the model parameters begins. As a first step, an error function in the model parameters must be chosen. This error function should embody some notion of what is meant for a pair of images to be registered. Perhaps the most common choice is a mean square error (MSE), defined as the mean of the square of the differences (in either feature distance or intensity) between the pair of images. This metric is easy to compute and often affords simple minimization techniques. A variation of this metric is the unnormalized correlation coefficient applicable to intensity-based techniques. This error metric is defined as the su
m of the point-wise products of the image intensities, and can be efficiently computed using Fourier techniques. A disadvantage of these error metrics is that images that would qualitatively be considered to be in good registration may still have large errors due to, for example, intensity variations, or slight misalignments. Another error metric (included in AIR) is the ratio of image uniformity (RIU) defined as the normalized standard deviation of the ratio of image intensities. Such a metric is invariant to overall intensity scale differences, but typically leads to nonlinear minimization schemes. Mutual information, entropy  and the Pearson product moment cross correlation are just a few examples of other possible error functions. Such error metrics are often adopted to deal with the lack of an explicit model of intensity transformations .

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。