With the rapid development of computer and information technologies, medical imaging has become one of the major sources of information for therapy and research in medicine, biology and other fields. Along with the advancement of medical imaging techniques, computer-aided detection and diagnosis (CAD/CADx) has recently emerged to become one of the major research subjects within the area of diagnostic radiology and medical image analysis. This thesis presents two multi-level learning-based approaches for medical image understanding with applications of CAD/CADx. The so-called "multi-level learning strategy" relies on that supervised and unsupervised statistical learning techniques are utilized to hierarchically model and analyze the medical image content in a "bottom up" way. As the first approach, a learning-based algorithm for automatic medical image classification based on sparse aggregation of learned local appearance cues is proposed. The algorithm starts with a number of landmark detectors to collect local appearance cues throughout the image, which are subsequently verified by a group of learned sparse spatial configuration models. In most cases, a decision could already be made at this stage by simply aggregating the verified detections. For the remaining cases, an additional global appearance filtering step is employed to provide complementary information to make the final decision. This approach is evaluated on a large-scale chest radiograph view identification task and a multi-class radiograph annotation task, demonstrating its improved performance in comparison with other state-of-the-art algorithms. It also achieves high accuracy and robustness against images with severe diseases, imaging artifacts, occlusion, or missing data. As the second approach, a learning-based approach for automatic segmentation of ill-defined and spiculated mammographic masses is presented. The algorithm starts with statistical modeling of exemplar-based image patches. Then, the segmentation problem is regarded as a pixel-wise labeling problem on the produced mass class-conditional probability image, where mass candidates and clutters are extracted. A multi-scale steerable ridge detection algorithm is further employed to detect spiculations. Finally, a graph-cuts technique is employed to unify all outputs from previous steps to generate the final segmentation mask. The proposed method specifically tackles the challenge of inclusion of mass margin and associated extension for segmentation, which is considered to be a very difficult task for many conventional methods. / Master of Science

Inflammatory muscle disease (myositis) is characterised by inflammation and a gradual increase in muscle weakness. Diagnosis typically requires a range of clinical tests, including magnetic resonance imaging of the thigh muscles to assess the disease severity. In the past, this has been measured by manually counting the number of muscles affected. In this work, a computer-aided analysis of inflammatory muscle disease is presented to help doctors diagnose and monitor the disease. Methods to quantify the level of oedema and fat infiltration from magnetic resonance scans are proposed and the disease quantities determined are shown to have positive correlation against expert medical opinion. The methods have been designed and tested on a database of clinically acquired T1 and STIR sequences, and are proven to be robust despite suboptimal image quality. General background information is first introduced, giving an overview of the medical, technical, and theoretical topics necessary to understand the problem domain. Next, a detailed introduction to the physics of magnetic resonance imaging is given. A review of important literature from similar and related domains is presented, with valuable insights that are utilised at a later stage. Scans are carefully pre-processed to bring all slices in to a common frame of reference and the methods to quantify the level of oedema and fat infiltration are defined and shown to have good positive correlation with expert medical opinion. A number of validation tests are performed with re-scanned subjects to indicate the level of repeatability. The disease quantities, together with statistical features from the T1-STIR joint histogram, are used for automatic classification of the disease severity. Automatic classification is shown to be successful on out of sample data for both the oedema and fat infiltration problems.

京都大学 / 新制・課程博士 / 博士(医学) / 甲第24473号 / 医博第4915号 / 新制||医||1062(附属図書館) / 京都大学大学院医学研究科医学専攻 / (主査)教授 溝脇 尚志, 教授 黒田 知宏, 教授 花川 隆 / 学位規則第4条第1項該当 / Doctor of Medical Science / Kyoto University / DFAM

This thesis presents techniques constructed to aid the radiologists in detecting breast cancer, the second largest cause of cancer deaths for western women. In the first part of the thesis, a computer-aided detection (CAD) system constructed for the detection of stellate lesions is presented. Different segmentation methods and an attempt to incorporate contra-lateral information are evaluated. In the second part, a new method for evaluating such CAD systems is presented based on constructing credible regions for the number of false positive marks per image at a certain desired target sensitivity. This method shows that the resulting regions are rather wide and this explains some of the difficulties encountered by other researchers when trying to compare CAD algorithms on different data sets. In this part an attempt to model the clinical use of CAD as a second look is also made and it shows that applying CAD in sequence to the radiologist in a routine manner, without duly altering the decision criterion of the radiologist, might very well result in suboptimal operating points. Finally, in the third part two dual-energy imaging methods optimized for contrast-enhanced imaging of breast tumors are presented. The first is based on applying an electronic threshold to a photon-counting digital detector to discriminate between high- and low-energy photons. This allows simultaneous acquisition of the high- and low-energy images. The second method is based on the geometry of a scanned multi-slit system and also allows single-shot contrast-enhanced dual-energy mammography by filtering the x-ray beam that reaches different detector lines differently. / QC 20100819

In this thesis, we present an algorithm for the automated detection of abnormalities (targets) in ultrasound images. The algorithm uses little a priori information and does not require training data. The proposed scheme is a combination of the CLEAN algorithm, originally proposed for radio astronomy, and constant false alarm rate (CFAR) processing, developed for use in radar systems. Neither of these algorithms appears to have been previously used for target detection in ultrasound images. The CLEAN algorithm identifies areas in the ultrasound image that stand out above a threshold in relation to the background; CFAR techniques allow for an automated and adaptive selection of the threshold. The algorithm was tested on simulated B-mode images. Using a contrast-detail analysis, probability of detection curves indicate that, depending on the contrast, the method has considerable promise for the automated detection of abnormalities with diameters greater than a few millimetres.

In this thesis, we present an algorithm for the automated detection of abnormalities (targets) in ultrasound images. The algorithm uses little a priori information and does not require training data. The proposed scheme is a combination of the CLEAN algorithm, originally proposed for radio astronomy, and constant false alarm rate (CFAR) processing, developed for use in radar systems. Neither of these algorithms appears to have been previously used for target detection in ultrasound images. The CLEAN algorithm identifies areas in the ultrasound image that stand out above a threshold in relation to the background; CFAR techniques allow for an automated and adaptive selection of the threshold. The algorithm was tested on simulated B-mode images. Using a contrast-detail analysis, probability of detection curves indicate that, depending on the contrast, the method has considerable promise for the automated detection of abnormalities with diameters greater than a few millimetres.

<p>Breast cancer screening is currently performed by mammography, which is limited by overlying anatomy and dense breast tissue. Computer aided detection (CADe) systems can serve as a double reader to improve radiologist performance. Tomosynthesis is a limited-angle cone-beam x-ray imaging modality that is currently being investigated to overcome mammography's limitations. CADe systems will play a crucial role to enhance workflow and performance for breast tomosynthesis.</p><p>The purpose of this work was to develop unique CADe algorithms for breast tomosynthesis reconstructed volumes. Unlike traditional CADe algorithms which rely on segmentation followed by feature extraction, selection and merging, this dissertation instead adopts information theory principles which are more robust. Information theory relies entirely on the statistical properties of an image and makes no assumptions about underlying distributions and is thus advantageous for smaller datasets such those currently used for all tomosynthesis CADe studies.</p><p>The proposed algorithm has two 2 stages (1) initial candidate generation of suspicious locations (2) false positive reduction. Images were accrued from 250 human subjects. In the first stage, initial suspicious locations were first isolated in the 25 projection images per subject acquired by the tomosynthesis system. Only these suspicious locations were reconstructed to yield 3D Volumes of Interest (VOI). For the second stage of the algorithm false positive reduction was then done in three ways: (1) using only the central slice of the VOI containing the largest cross-section of the mass, (2) using the entire volume, and (3) making decisions on a per slice basis and then combining those decisions using either a linear discriminant or decision fusion. A 92% sensitivity was achieved by all three approaches with 4.4 FPs / volume for approach 1, 3.9 for the second approach and 2.5 for the slice-by-slice based algorithm using decision fusion.</p><p>We have therefore developed a novel CADe algorithm for breast tomosynthesis. The techniques uses an information theory approach to achieve very high sensitivity for cancer detection while effectively minimizing false positives.</p> / Dissertation

Early detection of breast cancer on screening mammograms is crucial to reduce mortality rates. Computer-aided detection (CADe) systems for mammography are of great importance since they have been shown to positively assist radiologists in detecting early cancer. However, one area where CADe systems for mammography need improvement is in the early detection and annotation of spiculated lesions, which may represent invasive malignancies, and hence, early detection is crucial. Spicule annotation is important since it can yield useful discriminative information about the suspect lesion location on the mammogram and can also provide rich visual evidence to the interpreting radiologist to make the right follow-up decision. However, spicule annotation is a non-trivial task since spicules are fine scale curvilinear structures that are often not clearly visible amidst the surrounding breast parenchyma. The first contribution of this dissertation is an active contour algorithm called snakules for the annotation of spicules on mammography. Observer studies with experienced radiologists to evaluate the performance of snakules demonstrate the potential of the algorithm as an annotation tool that could be used to augment existing spiculated mass CADe systems. Mammography suffers from a major limitation: the 3-D to 2-D projection process results in anatomical noise due to overlapping of out of plane tissue structures, which hinders both radiologists and CADe systems in finding early cancers. This has motivated the development of 3-D breast imaging in the form of breast tomosynthesis, stereoscopic (stereo) mammography, and breast computed tomography (CT) to augment mammography for early cancer detection. Our second contribution is a novel computational stereo model for estimating a dense disparity map from a pair of stereo mammograms. This problem is very important since this is the first step towards elucidating 3-D information that is essential for conducting 3-D digital analysis on the stereo mammogram images. Nearly all of the 3-D structural information of interest on a stereo mammogram exists as a complex network of multi-layered, heavily occluded curvilinear structures, which is unlike what is seen on optical images of the real world. Our proposed stereo model employs a new singularity index as a constraint in a global optimization framework to obtain better estimates of disparity along critical curvilinear structures. The new singularity index is an important contribution of this work. In-depth theoretical analyses and experiments on several real world images demonstrate the efficacy of the index for detecting multi-scale curvilinear structures. Experiments on synthetic images with known ground truth and on real stereo mammograms highlight the advantages of the proposed stereo model over the canonical stereo model. The final contribution of this dissertation is an observer study, which demonstrates the feasibility of viewing breast tomosynthesis projection images stereoscopically. Unlike stereo mammogram images, each tomosynthesis projection image is acquired at a much lower dose. Stereo viewing of tomosynthesis projection images has the potential to reveal the 3-D structure of the breast, unlike the current cine or slice-by slice viewing modes. The results from our study suggest that stereo viewing could be a viable reading mode for breast tomosynthesis data in the future. / text

Ultrasound imaging suffers from severe speckle noise. We propose a novel approach for speckle reduction using 2D homogeneity and directional average filters to remove speckle noise. We transform speckle noise into additive noise using a logarithm transformation. Texture information is employed to describe the speckle characteristics of the image. The homogeneity value is defined using texture information value, and the ultrasound image is transformed into a homogeneity domain from the gray domain. If the homogeneity value is high, the region is homogenous and has less speckle noise. Otherwise, the region is nonhomogenous, and speckle noise occurs. The threshold value is employed to distinguish homogenous regions from regions with speckle noise obtained from a 2D homogeneity histogram according to the maximal entropy principle. A new directional filtering is convoluted to remove noise from pixels in a nonhomogenous region. The filtering processing iterates until the breast ultrasound image is homogenous enough. Experiments show the proposed method improves denoising and edge-preserving capability. We present a novel enhancement algorithm based on fuzzy logic to enhance the fine details of ultrasound image features, while avoiding noise amplification and over-enhancement. We take into account both the fuzzy nature of an ultrasound and feature regions on images, which are significant in diagnosis. The maximal entropy principle utilizes the gray-level information to map the image into fuzzy domain. Edge and textural information is extracted in fuzzy domain to describe the features of lesions. The contrast ratio is computed and modified by the local information. Finally, the defuzzification operation transforms the enhanced ultrasound images back to the spatial domain. Experimental results confirm a high enhancement performance including fine details of lesions, without over- or under-enhancement. Identifying object boundaries in ultrasound images is a difficult task. We present a novel automatic segmentation algorithm based on characteristics of breast tissue and eliminating particle swarm optimization (EPSO) clustering analysis, thus transforming the segmentation problem into clustering analysis. Mammary gland characteristics in ultrasound images are utilized, and a step-down threshold technique is employed to locate the mammary gland area. Experimental results demonstrate that the proposed approach increases clustering speed and segments the mass from tissue background with high accuracy.

Currently, mammography is the most common imaging technology used in breast screening. Low dose X-rays are passed through the breast to generate images called mammograms. One type of breast abnormality is a cluster of microcalcifications. Usually, in benign cases, microcalcifications result from the death of fat cells or are due to secretion by the lobules. However, in some cases, clusters of microcalcifications are indicative of early breast cancer, partly because of the secretions by cancer cells or the death of such cells. Due to the different attenuation characteristics of normal breast tissue and microcalcifications, the latter ideally appear as bright white spots and this allows detection and analysis for breast cancer classification. Microcalcification detection is one of the primary foci of screening and has led to the development of computer-aided detection (CAD) systems. However, a fundamental limitation of mammography is that it gives a 2D view of the tightly compressed 3D breast. The depths of entities within the breast are lost after this imaging process, even though the breast tissue is spread out as a result of the compression force applied to the breast. The superimposition of tissues can occlude cancers and this has led to the development of digital breast tomosynthesis (DBT). DBT is a three-dimensional imaging involving an X-ray tube moving in an arc around the breast, over a limited angular range, producing multiple images, which further undergo a reconstruction step to form a three-dimensional volume of breast. However, reconstruction remains the subject of research and small microcalcifications are "smeared" in depth by current algorithms, preventing detailed analysis of the geometry of a cluster. By using the geometry of the DBT acquisition system, we derive the "epipolar" trajectory of a microcalcification. As a first application of the epipolars, we develop a clustering algorithm after using the Hough transform to find corresponding points generated from a microcalcification. Noise points can also be isolated. In addition, we show how microcalcification projections can be detected adaptively. Epipolar analysis has also led to a novel detection algorithm for DBT using a Bayesian method, which estimates a maximum a posterior (MAP) labelling in each individual image and subsequently for all projections iteratively. Not only does this algorithm output the binary decision of whether a pixel is a microcalcification, it can predict the approximate depth of the microcalcification in the breast if it is. Based on the epipolar analysis, reconstruction of just a region of interest (ROI) e.g. microcalcification clusters is possible and it is more straightforward than any existing method using reconstruction slices. This potentially enables future classification of breast cancer when more clinical data becomes available.