Frontiers of Medical Imaging

The following sections are included:

• Preface
• Contents

Coronary artery disease (CAD) is one of the main causes of human deaths in the western world. Usually, patients are being diagnosed by ECG analysis first, after which for asymptomatic patients additional tests are necessary. Ultimately an imaging procedure might be necessary to image the coronary arteries. Until a decade ago invasive coronary angiography was the standard imaging method for this. Since then, minimal invasive coronary angiography by multi-slice computed tomography (MSCT-CA) became available and has been a highly investigated imaging technique and may become a standard imaging tool for the diagnosis of coronary artery disease (CAD). The three-dimensional (3D) nature of MSCT-CA allows 3D image reconstruction from which quantitative parameters of the coronary arteries and plaques can be derived.

Iterative Positron Emission Tomography (PET) reconstruction computes projections between the voxel space and the LOR space, which are mathematically equivalent to the evaluation of multi-dimensional integrals. These integrals are elements of the System Matrix (SM) and can be obtained either by deterministic quadrature or Monte Carlo (MC) methods. Due to the enormous size of the SM, it cannot be stored but integral estimation should be repeated whenever matrix elements are needed. In this paper we show that it is worth using random SM estimates, because this way errors caused by the projectors of different iteration steps can compensate each other, and do not accumulate to unacceptable values, which can happen in case of deterministic approximation. MC quadrature, however, introduces a random noise in the iteration process, which will not be convergent but iterated values will fluctuate around the real solution.We also propose methods to reduce the scale of this fluctuation and force the iteration to concentrate on the real solution.

In dynamic MRI reconstruction, the problem is to recover the sequence given partially sampled K-space of each frame. Traditionally, the reconstruction is offline. However such offline reconstruction precludes application of these techniques where real-time reconstruction is necessary, viz. image guided surgery, catheter tracking etc. To enable such applications, there is an interest in online reconstruction of dynamic MRI. This chapter reviews existing techniques in this area.

Longitudinal brain image registration is a technique that is crucial to investigate trajectories of brain change in normal aging and dementia. The goal of producing statistically robust profiles of change is challenging, confounded by image artifacts and algorithm bias. Tensor-based morphometry (TBM) computes deformation mappings between pairs of images of the same subject at different scan times, using log-determinants of the mapping jacobians to record localized change. These algorithms typically generate forces derived from image mismatch at edges to drive change computation, adding a penalty term to dampen spurious indications of change from bias and image artifacts. Mismatch and penalty terms necessitate a tradeoff between sensitivity and specificity. Recent work has focused on techniques to minimize algorithm bias and ensure longitudinal consistency. Little has been done to improve the tradeoff using prior information, i.e. derived from the images themselves prior to matching. In the present chapter, we survey the recent literature and then propose an approach for enhancing specificity using prior information. Our experiments suggest that the resulting method has enhanced statistical power for discriminating between different trajectories of brain change over time.

X-ray fluorescence computed tomography (XFCT) is a promising technique able to identify and quantify features in small samples of high-atomic-number (Z) elements such as iodine, gadolinium and gold. In this chapter, as a proof-of-concept, we investigated the feasibility of simultaneously imaging multiple elements (multiplexing) using XFCT. A polychromatic X-ray source was used to stimulate emission of X-ray fluorescence photons from multiple high-Z elements with low concentrations [2% (weight/volume) gold (Au), gadolinium (Gd) and barium (Ba)] embedded within a water phantom. The water phantom was used to mimic biological tissue targeted with high-Z nanoprobes. The emitted X-ray energy spectra from three elements were collected and then used to isolate the K shell X-ray fluorescence peaks and to generate sinograms for the three elements of interest. The tomographic distribution of the high Z elements within the phantom was reconstructed. A linear relationship between the X-ray fluorescence intensity of tested elements and their concentrations was observed, suggesting that XFCT is capable of quantitative molecular imaging.

How to reduce radiation dose while maintaining the diagnostic performance is a major challenge in the computed tomography (CT) field. Inspired by the compressive sensing theory, the sparse constraint in terms of total variation (TV) minimization has already led to promising results for low-dose CT reconstruction. Compared to the discrete gradient transform used in the TV minimization method, dictionary learning is proven to be an effective way for sparse representation. On the other hand, it is important to consider the statistical property of projection data in low-dose CT cases. In this chapter, we present a dictionary learning based approach for low dose X-ray CT. In our method, the sparse constraint in terms of a redundant dictionary is incorporated into an objective function in a statistical iterative reconstruction framework. The dictionary can be either predetermined before an image reconstruction task or adaptively defined during the reconstruction process. An alternating minimization scheme is developed to minimize the objective function. Our approach is evaluated with low dose X-ray projections collected in an animal study. The results show that the proposed approach has potential to produce better images with lower noise and more detailed structural features.

Digital images combined with computer aided diagnosis techniques offer dermatologists a simple, robust, objective and portable means for assessing many kinds of dermatological conditions. Some of the conditions that are amenable to computer aided diagnosis include acne, psoriasis and rosacea. Computer aided diagnosis can act as an external reference point to verify clinicians' diagnoses or can act as another input to the diagnosis making process. In this chapter we examine machine-learning methods for the computer aided diagnosis of psoriasis severity by developing an algorithm for the analysis of psoriasis severity. Psoriatic lesions are analysed by first segmenting erythema, the inflamed red skin in lesions, and scaling, the flaky white skin in lesions. Erythema segmentation is achieved by decomposing the skin colour into its melanin and haemoglobin colour components, and a Support Vector Machine (SVM) is designed to identify erythema pixels. Scaling segmentation is done by first applying a contrast map to heighten the contrast between scaling and non-scaling pixels, and then using textures and neighbourhoods to identify the scaling pixels. Markov random fields (MRFs) are used in conjunction with SVMs to first identify candidate scaling pixels and then use the neighbourhood in which a pixel is located to determine its final classification. We demonstrate the effectiveness of the proposed segmentation and classification methods by testing them on a database of psoriasis images.

Our aim in this chapter is to study the conditions for the reasonably good performance of classifiers on brain functional magnetic resonance imaging. We propose a synthetic model for the systematic study of aspects such as dimensionality, sample size, subject variability and noise. Our simulations highlight the key factors that affect generalization accuracy.

Functional magnetic resonance imaging (fMRI) has become a novel technique for studying the human brain and obtaining maps of neuronal activity. An important goal in fMRI studies is to decompose the observed series of brain images in order either to detect activation when a stimulus is presented to the subject, or to identify and characterize underlying brain functional networks when the subject is at rest. In this chapter, a model class is presented for addressing this issue that consists of finite mixture of generalized linear regression models. The main building block of the method is the general linear model which constitutes a standard statistical framework for investigating relationships between variables of fMRI data. We extend this into a finite mixture framework that exploits enhanced modeling capabilities by incorporating some innovative sparse and spatial properties. In addition, a weighted multi-kernel scheme is employed dealing with the selection problem of kernel parameters where the weights are estimated during training. The proposed regression mixture model is trained using the maximum a posteriori approach, where the Expectation-Maximization (EM) algorithm is applied for constructing update equations for the model parameters. We provide comparative experimental results in both activation-based and resting state applications that illustrate the ability of the proposed method to produce improved performance and discrimination capabilities.

Cutaneous malignant melanoma is one of the most frequent types of cancer in the world; but if the malignancy is detected and treated early, it can be cured. Many dermatologists promote dermoscopy as an early detection tool; however, dermoscopy requires formal training with a steep learning curve. In this chapter, we introduce a novel tree-based framework to automate the melanoma detection from dermoscopic images. Inspired by the radial and vertical growth pattern of skin lesions, we designed a flexible and powerful framework by decomposing dermoscopic images recursively. Pixels are repeatedly clustered into sub-images according to the color information and spatial constraints. This framework allows us to extract features by examining the tree from a graphical aspect, or from a textural/geometrical aspect on the nodes. In order to demonstrate the effectiveness of the proposed framework, we applied the technique, in completely different manners, to two common tasks of a computer-aided diagnostic system: segmentation and classification. The former task achieved a per-pixel sensitivity and specificity of 0.89 and 0.90 respectively on a challenging data set of 116 pigmented skin lesions. The latter task was tested on a public data set of 112 malignant and 298 benign lesions. We obtained 0.86 and 0.85 for precision and recall, respectively, along with an F-measure of 0.83 using a 3-layer perceptron. These experiments testified the versatility and the power of the tree-structure framework for dermoscopic analyses.

The retinal vascular network and optic disc are two of the most important anatomical structures of the fundus. Two methods to detect them have been included in this chapter. Their goal is to form part of some disease screening system since the segmentation of these structures can be a key process to assist clinicians in different pathology diagnosis. For example, both of them can be used as reference to detect several retinal lesions or other ocular structures in addition to identifying some fundus features. This chapter presents an in-depth study about the use of mathematical morphology to locate these main retinal structures. In particular, the proposed methods are characterized by combining several morphological operators and by making use of a variant of the watershed transformation, the stochastic watershed. This transformation avoids sub-segmentation problems related to the classical watershed. The methods have been validated in several public databases by improving the results of other state-of-the-art algorithms. Therefore, it has been demonstrated that the performance of the stochastic watershed allows us to apply it for clinical purposes.

Glaucoma is a chronic eye disease in which the optic nerve is progressively damaged. As the disease often progresses silently without symptoms, early detection of glaucoma via screening is important. Cup to disc ratio (CDR) computed from monocular retinal fundus images may provide an option for low cost large-scale glaucoma screening programme. In the chapter, we introduce the Automatic RetinA cup to disc Ratio AssessmenT (ARARAT) system for glaucoma screening. ARARAT uses superpixel classification to segment the optic disc and optic cup from monocular retinal fundus images and computes the CDR values for the screening. The method is validated using two data sets from different races. The areas under curve of the receiver operating characteristic curves are 0.827 and 0.822 from the two data sets. From the discussion with clinicians, the results are promising for large-scale glaucoma screening.

Dynamic contrast enhanced (DCE)-MRI combined with pharmacokinetic (PK) modeling of a tumor tissue provides information about its perfusion and vascular permeability. Most PK models require the intravascular contrast agent concentration as an input, which is inseparable from DCE-MRI data. Thus, it is approximated with an arterial input function (AIF), measured outside of the tissue of interest. Variations and error in calculation of AIF is a major source of discrepancy between PK parameters reported in different studies. An algorithm is developed, using blind source separation, to identify and separate the intravascular signal from the DCE-MRI data at the tissue of interest. The performance of the algorithm is assessed using numerical and physical tissue-mimicking phantom as well as clinical DCE-MRI of prostate cancer. The results show the algorithm is capable of accurately separating the intravascular signal, and its PK parameters provide better separation of normal and tumor tissues. Thus, the proposed algorithm could replace AIF-based methods, and more importantly could be used in cases an AIF cannot be measured.

Automated breast ultrasound (ABUS) is developed to automatically scan the whole breast for breast imaging on clinical examination. Hundreds of slices compose an image volume in a scanning to establish the volumetric structure of breasts. Reviewing the image volumes is a time-consuming task for radiologists. In this chapter, a computer-aided detection (CADe) system is proposed to automatically detect suspicious abnormalities in ABUS images. The database used included 122 abnormal and 37 normal cases. In abnormal cases, 58 are benign and 78 are malignant lesions. A Hessian-based multi-scale blob detection was used to segment blob-like structures into tumor candidates. The blobness, echogenicity, and morphology features were then extracted and combined in a logistic regression model to classify tumors and nontumors. The CADe system achieved the sensitivity of 100%, 90%, and 70% with false positives per case of 17.4, 8.8, and 2.7, respectively. The performance provides a promising use in tumor detection of ABUS images.

This review utilizes prostate cancer (CaP) as a case study to assess the application of quantitative histomorphometry to characterize the aggressive phenotype and also to make predictions of outcome like recurrence, metastasis and survival. Dr. Robert Veltri describes the use of a microspectrophotometry microscope and novel software to capture nuclear morphometry of Feulgen (DNA) stained features and successfully identify indolent and aggressive CaP as well as predict outcomes such as biochemical recurrence, metastasis and survival. His research also indicates that quantitative nuclear morphometry by this method indicates a field effect nearby the can that also has predictive value. However, the original technology was bottle-necked by the fact that it could not be extended to whole slide images. Subsequently, the initiation of a collaboration with Dr. Anant Madabhushi who has developed high throughput quantitative image and histomorphometric tools that are amenable to work on whole slide digitized images, has allowed for confirmation of Dr. Veltri's prior and published observations that nuclear size, shape and texture as well as the glandular structure or architecture of prostate cancer are critical in predicting disease aggressiveness. Dr. Madabhushi's algorithms take advantage of advanced machine vision imaging images for diagnosis and prognosis. Additionally Dr. Madabhushi's group has also been developing machine learning tools to combine image based and molecular measurements for creating unified predictors of disease aggressiveness and patient outcome. It is hoped that the additional development and validation of these tools will set the stage for creation of decision tools to aid the pathologist to predict severe outcomes early so that appropriate interventions can be made by the urologist and patient.

Diabetic retinopathy (DR) is the first cause of blindness in people of working age in the developed countries. In spite of DR not being a curable illness, it can be treated if it is detected in its early stages. However, this early detection is not always possible since DR patients do not perceive symptoms until the illness is in advance stages, when treatment is less effective. This is why diabetic patients are periodically examined in the frame of preventive screening programs. Nevertheless, effectiveness of these preventive protocols is compromised due to the human and material resources needed to manage a huge number of patients needing revisions. This is why, especially over the last two decades, a great effort has been carried out by the scientific community to develop reliable and accurate systems for the automated detection of DR.

On the one hand, this chapter describes the medical and technical fundamentals of the automated detection of DR. On the other hand, outstanding comprehensive systems tested within the frame of screening programs as clinical trials are reviewed in order to assess the current state of art of the topic and its perspectives.

Magnetic resonance guided high intensity focused ultrasound (MRgHIFU) ablation is a promising technique for the treatment of hepatic and renal cancers. In this approach, the HIFU system is used for ablating the tumor using local heating while the MRI scanner provides online temperature measurements which enable continuous monitoring of the delivered thermal dose. The application of MRgHIFU ablation for abdominal organs is challenging. Many technical developments have addressed the major issues associated with this procedure such as the physiological motion of abdominal organs. In this chapter, we present the different techniques which have been developed to alleviate the problem of physiological motion for MRgHIFU ablation of abdominal organs.

Proton therapy is a key radiation therapy modality used in the treatment of cancer. With a low entrance dose and no exit dose, proton therapy has the desired physics properties to spare healthy tissues. However, uncertainties attributable to calibration of the computed tomography (CT) number-to-proton-relative stopping power, daily patient setup, and anatomical changes during a treatment course lead to proton-range uncertainties, which can be detrimental to patients. The intrinsic range uncertainty in CT-relative stopping power calibration can be minimized using proton CT; the random and systematic range variations from patient setup and anatomical changes can be monitored using positron emission tomography (PET) and prompted gamma ray imaging, which detect the byproducts from proton nuclear interactions with human tissues. This chapter reviews proton therapy physics, sources of proton-range uncertainties, and the advanced imaging technologies used to minimize or monitor these uncertainties before, during, and after proton therapy.

It is widely known that the state of a patient's coronary heart disease can be better assessed using intravascular ultrasound (IVUS) than with more conventional angiography. Recent work has shown that segmentation and 3D reconstruction of IVUS pullback sequence images can be used for computational fluid dynamic simulation of blood flow through the coronary arteries. This map of shear stress in the blood vessel walls can be used to predict susceptibility of a region of the arteries to future arteriosclerosis and disease. Manual segmentation of images is time consuming as well as cost prohibitive for routine diagnostic use. Current segmentation algorithms do not achieve a high enough accuracy because of the presence of speckle due to blood flow, relatively low resolution of images and presence of various artifacts including guide-wires, stents, vessel branches, and some other growth or inflammations. On the other hand, the image may be induced with additional blur due to movement distortions, as well as resolution-related mixing of closely resembling pixels thus forming a type of out-of-focus blur. Robust automated segmentation achieving high accuracy of 95% or above has been elusive despite work by a large community of researchers in the machine vision field. In this chapter, we present a comprehensive method, based on computer vision and pattern recognition, where a multitude of algorithms are applied simultaneously to the segmentation problem. The method presented is to combine algorithms using a meta-algorithmic approach. Each segmentation algorithm computes along with the segmentation a measure of confidence in the segmentation which can be biased on prior information about the presence of artifacts. A meta-algorithm then runs a library of algorithms on a sub-sequence of images to be segmented and chooses the segmentation based on computed confidence measures. Machine learning and testing is performed on a large database, that includes 2293 gated image frames that have been manually segmented for training and performance comparison, and a total of 57,098 image frames for testing the meta-algorithm to obtain reliable segmentation performance assessment.

Atherosclerosis is a condition in which the wall of the arteries hardens and thickens due to the accumulation of plaque. The primary cause of acute coronary events is related to the inflammation and disruption of coronary plaques. Characterization and detection of plaques vulnerable to complications (i.e., vulnerable plaque, VP) is one of the most active areas of research in cardiology. Intravascular ultrasound (IVUS) is an invasive technique that is capable of providing a real-time, high-resolution tomographic visualization of the coronary arteries which allows the accurate estimation of the morphological characteristics of the vessel. The extraction of valuable information from the IVUS data through the use of medical image analysis methods has played an important role in the detection and characterization of VP. In this chapter, we will present different computational methods developed for the analysis of IVUS data and the detection of VP.

Capacitive micromachined ultrasonic transducers (CMUTs) are micro-electromechanical devices (MEMS) fabricated using silicon micromachining techniques. In the past decade, their use has proved to be attractive mainly in the field of medical ultrasound imaging as active elements in ultrasound probes. The interest of this novel technology relies on its full compatibility with standard integrated circuit technology that makes it possible to integrate, on the same chip, the transducers and the electronics, thus enabling the realization of extremely low-cost and high-performance devices. From an operational point of view, CMUTs have been widely recognized as a valuable alternative to piezoelectric transducer technology in a variety of medical imaging applications, thanks to a higher sensitivity, a wider bandwidth, and an improved thermal efficiency.

In this chapter, the design and fabrication of a 192-element linear array CMUT probe operating in the range 6–18 MHz, designed for vascular, small parts, rheumatology and anesthesiology imaging applications, is reported. The CMUT array is microfabricated and packed using a novel fabrication concept specifically conceived for imaging transducer arrays. The performance optimization of the probe is performed by connecting the CMUT array with multichannel analog front-end electronic circuits housed into the probe body. Characterization and imaging results are used to assess the performance of CMUTs with respect to conventional piezoelectric transducers.

Rapidly growing volumes of medical data and studies, more imaging modalities, standardization of digital imaging and communications in medicine (DICOM) and healthcare level 7 (HL7), increasingly complicated workflow and structured reports, geographically distributed storage and access, data management, and security access have driven the evolution of Picture Archiving Communication System (PACS), adoption of web-based imaging cloud for healthcare information systems. Integrated enterprise PACS and imaging cloud services facilitate archiving, sharing and exchanging of imaging studies, ubiquitous and reliable network access anytime anywhere for clinical practice and healthcare provider, and achieve smooth workflow, balanced workloads, and optimized performance.

The following section is included:

• Index