Current Medical Imaging - Volume 9, Issue 4, 2013

Digital medical imaging has become one of the most important tools for medical diagnosis. However, the ongoing evolution in both storage and modality devices has had as consequence that enormous amounts of imaging data are being produced. The existence of large sets of data coupled with the limited query capabilities provided by the standard tools and protocols poses problems for radiologists and has shifted the research focus from data availability towards data accessibility. Content-based image retrieval (CBIR) systems have been heralded as a solution that is able to cope with the increasingly larger volumes of information present in medical repositories and assist radiologists with decision support. While generic, extensible CBIR frameworks that work natively with Picture Archive and Communication Systems (PACS) are scarce, developments are happening at a fast pace and several systems have been implemented with some degree of success. Based on the literature available, we provide an overview of such CBIR systems, architecture and implementation techniques, with an emphasis on systems oriented towards usage in a clinical domain. We conclude with an analysis of some of the challenges that still need to be overcome in order to bring this technology to a medical audience.

Many medical image processing applications are based on the detection of corresponding features. In digital images, features are depicted by one or more digital points. Hence, feature correspondence is achieved through the estimation of point correspondences between the compared images. This paper presents and evaluates five of the most common and recent techniques for estimating corresponding points between two flat digital images. The featured techniques include Template Matching, the Iterative Closest Points algorithm, Correspondence by Sensitivity to Movement, the Self- Organizing Maps and the Artificial Immune Network algorithm. All methods are presented, mainly focusing on their distinct characteristics. The featured techniques were tested both qualitatively and quantitatively on an extensive set of medical image pairs, including images subject to both known and unknown initial geometrical deviations. Each of the five methods was evaluated on all 263 available image pairs in terms of correspondence and registration accuracy. After assessing the point correspondence accuracy of each method, it was deduced that their performance depends on the characteristics of the featured data set. However, the Artificial Immune Network approach outperformed in most cases the rest of the featured point-correspondence methods, closely followed by the Self Organizing Maps algorithm.

Metaheuristics are techniques that use approximate and intuitive strategies to quickly find near-optimal solutions of complex optimization problems. A number of outstanding examples belong to evolutionary computation and swarm intelligence, two classes of methods inspired to biological phenomena. These techniques have been extensively and successfully applied to feature-based image registration in medicine. However, with the increase in computational power during the last decade, intensity-based (or voxel-based) image registration methods have been preferred in many medical imaging applications, due to their robustness, accuracy and applicability, in cases where landmarks or other features are not available or easy to detect. While traditional numerical optimization techniques are employed to solve the registration problem, a number of contributions in the literature support the use of metaheuristics to overcome the shortcomings of classic methods. The aim of the paper is to review the state of the art in the application of evolutionary computation and other metaheuristics to intensity-based medical image registration. The study considers both well-known techniques with a large number of references in the literature as well as recent, outstanding proposals. The analysis focuses on the design of the methods to highlight common and successful practices. In addition, recommendations and open research lines in the field are provided.

The Digital Design Studio and NHS Education Scotland have developed ultra-high definition real-time interactive 3D anatomy of the head and neck for dental teaching, training and simulation purposes. In this paper we present an established workflow using state-of-the-art 3D laser scanning technology and software for design and construction of medical data and describe the workflow practices and protocols in the head and neck anatomy project. Anatomical data was acquired through topographical laser scanning of a destructively dissected cadaver. Each stage of model development was clinically validated to produce a normalised human dataset which was transformed into a real-time environment capable of large-scale 3D stereoscopic display in medical teaching labs across Scotland, whilst also supporting single users with laptops and PC. Specific functionality supported within the 3D Head and Neck viewer includes anatomical labelling, guillotine tools and selection tools to expand specific local regions of anatomy. The software environment allows thorough and meaningful investigation to take place of all major and minor anatomical structures and systems whilst providing the user with the means to record sessions and individual scenes for learning and training purposes. The model and software have also been adapted to permit interactive haptic simulation of the injection of a local anaesthetic.

There are several different types of medical imaging modalities, among others magnetic resonance imaging (MRI), positron emission tomography (PET), ultrasound, computed tomography (CT) or two-dimensional electrophoresis images (2D-electrophoresis). The number of images is increasing rapidly and the development of automatic image processing systems is necessary in order to aid in diagnostic decisions and therapy assessments. One of the most important features in an image is texture, thus it is one of the central concepts in computer vision and should always be taken into account as an innate property. There are various methods of extracting textural features from images; this work considers statistical methods for texture analysis. Those methods analyze the spatial distribution of gray values by computing local features. Depending on the number of pixels, statistical methods can be classified into first- (one pixel), second- (two pixels) and high-order (three or more pixels). Second- and high-order statistics estimate properties of two or more pixel values occurring at specific locations relative to each other. This paper is focused on the high-order statistics texture analysis of CT, MRI, PET and 2D-electrophoresis images.

Optoacoustic tomography enables volumetric imaging with optical contrast in biological tissue at depths beyond the optical mean free path by the use of optical excitation and acoustic detection. The hybrid nature of optoacoustic tomography gives rise to two distinct inverse problems: The optical inverse problem, related to the propagation of the excitation light in tissue, and the acoustic inverse problem, which deals with the propagation and detection of the generated acoustic waves. Since the two inverse problems have different physical underpinnings and are governed by different types of equations, they are often treated independently as unrelated problems. From an imaging standpoint, the acoustic inverse problem relates to forming an image from the measured acoustic data, whereas the optical inverse problem relates to quantifying the formed image. This review focuses on the acoustic aspects of optoacoustic tomography, specifically acoustic reconstruction algorithms and imaging-system practicalities. As these two aspects are intimately linked, and no silver bullet exists in the path towards high-performance imaging, we adopt a holistic approach in our review and discuss the many links between the two aspects. Four classes of reconstruction algorithms are reviewed: time-domain (so called back-projection) formulae, frequency-domain formulae, time-reversal algorithms, and model-based algorithms. These algorithms are discussed in the context of the various acoustic detectors and detection surfaces which are commonly used in experimental studies. We further discuss the effects of non-ideal imaging scenarios on the quality of reconstruction and review methods that can mitigate these effects. Namely, we consider the cases of finite detector aperture, limited-view tomography, spatial under-sampling of the acoustic signals, and acoustic heterogeneities and losses.