Learning Manifolds

Computer aided diagnosis is often confronted with processing and analyzing high dimensional data. One alternative to deal with such data is dimensionality reduction. This chapter focuses on manifold learning methods to create low dimensional data representations adapted to a given application. From pairwise non-linear relations between neighboring data-points, manifold learning algorithms first approximate the low dimensional manifold where data lives with a graph; then, they find a non-linear map to embed this graph into a low dimensional space. Since the explicit pairwise relations and the neighborhood system can be designed according to the application, manifold learning methods are very flexible and allow easy incorporation of domain knowledge. The authors describe different assumptions and design elements that are crucial to building successful low dimensional data representations with manifold learning for a variety of applications. In particular, they discuss examples for visualization, clustering, classification, registration, and human-motion modeling.

Manifold Learning for Medical Image Registration, Segmentation, and Classification

The term manifold learning encompasses a class of machine learning techniques that convert data from a high to lower dimensional representation while respecting the intrinsic geometry of the data. The intuition underlying the use of manifold learning in the context of image analysis is that, while each image may be viewed as a single point in a very high-dimensional space, a set of such points for a population of images may be well represented by a sub-manifold of the space that is likely to be non-linear and of a significantly lower dimension. Recently, manifold learning techniques have begun to be applied to the field of medical image analysis. This chapter will review the most popular manifold learning techniques such as Multi-Dimensional Scaling (MDS), Isomap, Local linear embedding, and Laplacian eigenmaps. It will also demonstrate how these techniques can be used for image registration, segmentation, and biomarker discovery from medical images.

Computer-Aided Detection of Polyps in CT Colonography by Means of Feature Selection and Massive-Training Support Vector Regression

One of the major challenges in current Computer-Aided Detection (CADe) of polyps in CT Colonography (CTC) is to improve the specificity without sacrificing the sensitivity. If a large number of False Positive (FP) detections of polyps are produced by the scheme, radiologists might lose their confidence in the use of CADe. In this chapter, the authors used a nonlinear regression model operating on image voxels and a nonlinear classification model with extracted image features based on Support Vector Machines (SVMs). They investigated the feasibility of a Support Vector Regression (SVR) in the massive-training framework, and the authors developed a Massive-Training SVR (MTSVR) in order to reduce the long training time associated with the Massive-Training Artificial Neural Network (MTANN) for reduction of FPs in CADe of polyps in CTC. In addition, the authors proposed a feature selection method directly coupled with an SVM classifier to maximize the CADe system performance. They compared the proposed feature selection method with the conventional stepwise feature selection based on Wilks’ lambda with a linear discriminant analysis classifier. The FP reduction system based on the proposed feature selection method was able to achieve a 96.0% by-polyp sensitivity with an FP rate of 4.1 per patient. The performance is better than that of the stepwise feature selection based on Wilks’ lambda (which yielded the same sensitivity with 18.0 FPs/patient). To test the performance of the proposed MTSVR, the authors compared it with the original MTANN in the distinction between actual polyps and various types of FPs in terms of the training time reduction and FP reduction performance. The authors’ CTC database consisted of 240 CTC datasets obtained from 120 patients in the supine and prone positions. With MTSVR, they reduced the training time by a factor of 190, while achieving a performance (by-polyp sensitivity of 94.7% with 2.5 FPs/patient) comparable to that of the original MTANN (which has the same sensitivity with 2.6 FPs/patient).

Relevance Feedback as New Tool for Computer-Aided Diagnosis in Image Databases

With the ever-growing volume of images used in medicine, the capability to retrieve relevant images from large databases is becoming increasingly important. Despite the recent progress made in the field, its applications in Computer-Aided Diagnosis (CAD) thus far have been limited by the ability to determine the intrinsic mapping between high-level user perception and the underlying low-level image features. Relevance Feedback (RFB) is a post-query process to refine the search by using positive and/or negative indications from the user about the relevance of retrieved images, which has been applied successfully in traditional text-retrieval systems for improving the results of a retrieval strategy. In this chapter, the authors review some recent advances in RFB technology, and discuss its expanding role in content-based image retrieval from medical archives. They provide working examples, based on their experience, for developing machine-learning methods for RFB in mammography and highlight the potential opportunities in this field for CAD applications and clinical decision-making.

Computer-Aided Detection and Diagnosis for 3D X-Ray Based Breast Imaging

The last 15 years has seen the advent of a variety of powerful 3D x-ray based breast imaging modalities such as digital breast tomosynthesis, digital breast computed tomography, and stereo mammography. These modalities promise to herald a new and exciting future for early detection and diagnosis of breast cancer. In this chapter, the authors review some of the recent developments in 3D x-ray based breast imaging. They also review some of the initial work in the area of computer-aided detection and diagnosis for 3D x-ray based breast imaging. The chapter concludes by discussing future research directions in 3D computer-aided detection.

Neuroimage Classification for Early Diagnosis of Alzheimer’s Disease

Diagnostic criteria for neurological and psychiatric disorders are typically based on clinical and psychometric assessment, which might not be effective for early detection of the disease onset. For brain disorders such as Alzheimer’s Disease (AD), neuroimaging can potentially play an important role in the development of imaging-based biomarkers. Following voxel-wise univariate neuroimage analysis methods, machine learning and pattern recognition based neuroimage analysis techniques have been increasingly adopted in neuroimaging studies of neurological and psychiatric disorders, aiming to provide tools that classify individuals, based on their neuroimaging scans, rather than detect statistical group difference. The machine learning based methods, optimally combining information of multiple measures derived from images, have demonstrated promising performance in diagnosis of AD and early prediction of conversion of Mild Cognitive Impairment (MCI) individuals. This chapter introduces the general framework of such techniques with a focus on structural MRI analyses and their applications to studies of AD.

Computer-Aided Image Analysis and Detection of Prostate Cancer

Immunohistochemistry (IHC) is an adjunct tool for clinical histologic diagnosis of diseases. A common IHC technique for prostate cancer diagnosis is a triple-antibody cocktail with Alpha-Methylacyl-CoA Racemase (AMACR), p63, and High-Molecular-Weight Cytokeratin (HMWCK), which stains certain types of cells into two distinct colors. The authors have developed an automated computer technique that detects prostate cancer in prostate tissue sections processed with the triple-antibody cocktail. Test and validation of the authors’ technique on digital images obtained from conventional microscopes (region of interest images) showed that the computer technique can recognize prostatic adenocarcinoma with both high sensitivity and high specificity. The authors also used this computer technique to analyze whole-slide images of prostate biopsy and the initial results are promising. With further development and refinement, this computer technique could become a useful tool for pathologists to detect prostate cancer foci in histologic sections of tissue processed with the triple-antibody cocktail.

Clinical Machine Learning in Action

This chapter first discusses the database problems in CAD development comprehensively. Then, it introduces the authors’ integrated platform, called the Clinical Infrastructure for Radiologic Computation of United Solutions (CIRCUS), for in-hospital research, development, use, and evaluation of clinical image processing. Based on the authors’ clinical experience and the data collected through the CIRCUS system, they present research results on the improvement of CAD performance as well as simulated studies for additional learning. Finally, the authors’ future plans, including radiologist-CAD collaboration beyond machine learning, are also discussed.

Advanced Fuzzy Methods for Mammography Image Analysis

Breast cancer is the second most common type of cancer worldwide and one of the most common causes of cancer deaths. Worryingly, breast cancer incidence rates have increased over recent years. Computer Aided Diagnosis (CADx) systems are designed to help radiologists identify cancerous signs earlier, and hence to reduce the death rate. These systems involve at least two main stages: feature extraction to derive useful information from the images, and diagnosis which is typically handled as a machine learning/pattern classification problem. For breast cancer diagnosis, x-ray mammography is the main modality of diagnosis. The inherent fuzziness in the nature of mammography images makes fuzzy set theory a useful technique for handling these images. It is used as a well-suited tool to extract meaningful information from inexact data and generate appropriate solutions. In this chapter, the authors present a fast overview of some fuzzy-based methods for computer-aided detection and computer aided diagnosis of breast cancer using mammography images. Their focus is on fuzzy logic-based methods developed for mammogram enhancement, microcalcification (MC) detection, and detection and classification of masses.

Digital Image Processing and Machine Learning Techniques for the Detection of Architectural Distortion in Prior Mammograms

Architectural distortion is a subtle but important early sign of breast cancer. The purpose of this study is to develop methods for the detection of sites of architectural distortion in prior mammograms of interval-cancer cases. Screening mammograms obtained prior to the detection of cancer could contain subtle signs of early stages of breast cancer, in particular architectural distortion. The methods for the detection of architectural distortion are based upon Gabor filters, phase portrait analysis, a novel method for the analysis of the angular spread of power, fractal analysis via Fractal Dimension (FD), structural analysis of texture via Laws’ texture energy measures derived from geometrically transformed regions of interest (ROIs), and statistical analysis of texture using Haralick’s 14 texture features. The application of Gabor filters and linear phase portrait modeling was used to detect initial candidates of sites of architectural distortion; 4,224 ROIs, including 301 true-positive ROIs related to architectural distortion, were automatically obtained from 106 prior mammograms of 56 interval-cancer cases and from 52 mammograms of 13 normal cases. For each ROI, the FD, three measures of angular spread of power, 10 Laws’ measures, and 14 Haralick’s features were computed. The areas under the receiver operating characteristic curves obtained using the features selected by stepwise logistic regression and the leave-one-ROI-out method are 0.76 with the Bayesian classifier, 0.75 with Fisher linear discriminate analysis, and 0.78 with a single-layer feed forward neural network. Free-response receiver operating characteristics indicated sensitivities of 0.80 and 0.90 at 5.8 and 8.1 false positives per image, respectively, with the Bayesian classifier and the leave-one-image-out method. The methods have shown good potential in detecting architectural distortion in mammograms of interval-cancer cases.