Medical Imaging Systems Technology

The following sections are included:

• Preface

• Contents

In this chapter, we discuss how medical image analysis can take advantage of computational modeling. First, a general framework of model-based approach is presented, whose objective is enabling to relate external descriptions to internal processes, which originate them. Two coupled models i.e. model of an organ under study and model of an imaging modality, constitute the main components of this framework. Then, a few applications of the presented approach are described. Two vital abdominal organs, namely liver and kidney, are considered and their typical pathological processes connected with significant vascular modifications are analyzed. Resulting patterns are observed by means of dynamic Computed Tomography, when contrast material injection and propagation are simulated. Texture features are extracted from regions of interest defined in simulated images in different acquisition conditions. Studying evolution of features allows us to better understand both advantages and limitations of texture analysis and is very helpful in proper medical image interpretation.

This chapter describes a methodology for creating a generic volumetric biomechani-cal model of an organ from different image modalities. The application detailed is the segmentation of time series of cardiac images using this model. Other applications on non-rigid registration and physiology simulation are also presented. The building of such a generic model consists of three stages: geometric meshing, non-rigid deformation of the mesh in images of various modalities, and image to mesh information mapping through rasterization. The non-rigid deformation stage, which relies on the combination of global and local deformations, can then be used to segment time series of images, e.g. cine MRI or gated SPECT cardiac images. We believe that this type of deformable biomechanical models will play an important role in the extraction of useful quantitative local parameters of the cardiac function. Moreover, it makes it possible to simulate physiological phenomena based on the information integrated in the model and to incorporate biomechanical information in image analysis methods like non-rigid registration.

In this chapter, we discuss fractal-based algorithms and their applications to identify tumors in brain magnetic resonance (MR) images. Considerable research has been pursued on the application of fractal geometry in various aspects of image analysis and pattern recognition. MR images typically have a degree of noise and randomness associated with the natural random texture of structure. Thus, fractal analysis is suitable for MR image analysis. For tumor detection, we propose three novel fractal analysis algorithms and one existing fractal-based technique using brain MR images. The first algorithm involves thresholding the pixel intensity values and hence, we call the technique piecewise-threshold-box-counting (PTBC) method. For the subsequent two methods such as piecewise modified box-counting (PMBC) and piecewise triangular prism surface area (PTPSA) methods respectively, the intensity is treated as the third dimension. Using the PMBC and PTPSA methods we may detect and locate the tumor in the brain MR images more accurately. The performance of the existing “Blanket” fractal analysis technique is also comparable to either PMBC or PTPSA algorithms in tumor detection. In all these four fractal analysis techniques, we find the tumor location by obtaining the differences in fractal dimension between normal and tumor images. Further, we propose a fifth novel fractal analysis technique that exploits fractional Brownian motion (fBm) for brain tumor analysis. The fBm method does not require any reference image for locating tumor in the brain MR image. The fBm method also offers satisfactory tumor detection.

Early diagnosis of cancer continues to be a major challenge in the field of cancer prevention and management as it can decrease the severity and also improve the life of patients. Rapid progress in the past few decades in molecular- and immunological-based methods has helped to alleviate the problem to a certain extent. However, a rapid, reagent free method to identify cancers in situ or ex vivo, to make screening objective and swift still remains unachieved. At this point, the utilization of methods based on IR spectroscopy is expected to provide the breakthrough and make diagnosis and grading of cancers/malignancies a simple and inexpensive affair. In the present chapter, we deal with the utilization of FTIR-microspectroscopy for the diagnosis of colon cancer, which is a preliminary step towards its implications for future oncology. We highlight a few technological aspects and also the enormous potential of IR spectroscopy in colon cancer specifically, and other cancers in general.

In the last decade, ultrasound along with digital mammography has come to be regarded as the gold standard for breast cancer diagnosis. While several computer-aided diagnostic methodologies have been proposed to aid radiologists in detecting and diagnosing suspicious regions on digital mammograms, only recently has there been a similar initiative for breast ultrasound. In this article, we review different computerized methods for analysis of lesions seen on breast ultrasound. We describe and compare approaches and algorithms pertaining to detection, segmentation, feature extraction, and classification of sonographic breast lesions. We also highlight the challenges in computer-aided detection and diagnosis of lesions seen on breast ultrasound and suggest future avenues of research.

There are several comprehensive reviews on blind deblurring. This paper is a supplement and update to the reviews by Kundur and Hatzinakos. Most of the methods reviewed in Kundur and Hatzinakos’ papers require that the PSF and the original image must be irreducible. However, this irreducibility assumption is not true in some important types of applications, such as when the PSF is Gaussian, which is a good model for many imaging systems and a good approximation for spiral CT images. In this chapter, the edge-to-noise ratio (ENR) principle for blind deblurring was reviewed and applied to deblurring spiral CT images with the EM algorithm and Wiener filtering technique, two representative image deblurring algorithms. Two algorithms, EM-ENR and Wiener-ENR, are formulated according to the general ENR principle. In the phantom studies, both algorithms reduce image blurring, according to our blurring residual measure. Also, both algorithms work well in patient studies. After fully automatic blind deblurring, the conspicuity of the submillimeter features of the cochlea is substantially improved. The EM-ENR method significantly improves image quality with a high computation overhead, while the Wiener-ENR method greatly reduces the computation cost with an acceptable improvement of image quality. Finally, future research directions are discussed, along with guidelines for practical applications of the algorithms.

This chapter presents a versatile model-based coding scheme for three-dimensional medical data. The potential of the proposed system is that it copes with many of the requirements typical of the medical imaging field without sacrificing compression efficiency. Among the most interesting functionalities are progressively refinable up-to-lossless quality of the decoded information, object-based processing and the possibility to decode a single 2D image of the dataset. Furthermore, such features can be combined enabling a swift access to any two-dimensional object of any image of interest with refinable quality. The price to pay for this is an overhead in bitstream which could degrade the compression performance. Though, the possibility to focus the decoding process on a specific region of a certain 2D image allows a very efficient access to the information of interest. We believe this is an important feature for a coding system meant to be used for medical applications, which largely compensates for the eventual loss in compression that could be implied.

For multimodality image registration, mutual information has become the similarity measure of choice, due to its flexibility, theoretical elegance and general robustness. An essential element in any image registration algorithm is interpolation, which is needed to evaluate voxel intensities at non-grid positions. Interpolation has been a topic of considerable study and the performance of many interpolators has been characterized. Nevertheless, interpolation has some unexpected influence on registration accuracy, causing artifactual fluctuations in the estimated value of mutual information. These “interpolation artifacts” are not reduced by using interpolators with higher accuracy. This surprising finding warranted further investigation into the role of interpolation methods in multimodality image registration. This chapter reviews several commonly used interpolation methods, the application of such methods and the associated problems. A theoretical analysis of the underlying cause of these interpolation artifacts is described. Finally, several strategies are outlined to reduce these artifacts and to improve registration robustness. Such strategies are also applicable to related similarity measures, including normalized mutual information, joint entropy, and Hill's third moment.

This chapter describes a methodology for the construction of 3D atlases of the heart suitable for the generation of statistical shape models from large image databases of dynamic cardiac MRI and SPECT studies. The presented approach makes use of non-rigid registration for the automatic establishing of landmark correspondences across a training set of shapes. The main properties of the constructed shape models are presented, and their potential in terms of the ultimate goal of image segmentation is also discussed.

Mutual information is a powerful similarity measure assuming no specific relationship between the intensities of corresponding points. It has been shown that for many applications, by maximizing the mutual information, two images can be aligned with high precision. To utilize mutual information as a similarity measure for image registration, the joint histogram of two images must be estimated. Under certain conditions, existing joint histogram estimation methods result in a phenomenon known as interpolation induced artifacts that hamper the subsequent optimization process and may deteriorate registration accuracy. In this chapter, we discuss this phenomenon in detail and review some plausible solutions to reduce this phenomenon. In addition, a robust global optimization scheme is introduced in this chapter to optimize the mutual information similarity measure between two images. The core of this scheme lies in a heuristic global optimum test algorithm to distinguish the global optimum of the mutual information registration function from the numerous local ones. Finally, the effectiveness of the techniques covered in this chapter is demonstrated by two clinical examples including 3D brain image registration using the data provided by Vanderbilt University and cryosection image registration using the Visible Human Project data sets from the National Library of Medicine.

Image reconstruction problems arise in many fields of applications. The iterative approach is important for image reconstruction and attracting more and more attention with the rapid evolution of computer technology. In this chapter, classic results and recent advances on iterative algorithms for image reconstruction are reviewed, with an emphasis on the algebraic algorithms in both of their simultaneous and ordered-subset formats. Several important issues are discussed: computational structures, convergence conditions, characterizations of the converged limits, relaxation strategies, etc., especially in the inconsistent cases. Because emerging applications demand the algorithms in an infinite dimensional setting, the algorithms are formulated in a general configuration in infinite dimensional Hilbert spaces, which generalizes the convention finite dimensional cases. It was demonstrated that several important algorithms such as ART, SART, DWE and CAV all belong to the same class of weighted lest-squares solvers with different weightings and are specific versions of the Landweber scheme. The general framework provides a mechanism to derive new algorithms from known ones. The application of the infinite dimensional version is presented for one emerging image modality: bioluminescence tomography.

The following sections are included:

• INDEX