Computer-aided detection and characterization of interstitial lung disease (UNN.6126)

The detection and characterization of interstitial lung disease is a difficult task in radiology. Interstitial disease (ID) can be detected in chest radiographs, by far the most common examination in radiology, both in a clinical setting and in mass screening, notably for tuberculosis. The global tuberculosis epidemic leads to 3 million deaths each year. The alternative for chest radiographs is a computed tomography (CT) examination. In this project we intend to develop three systems for computer-aided diagnosis (CAD) of ID. The first system will be for tuberculosis screening. The second system will analyze clinical chest radiographs as input. The third system will examine isotropic CT scans with sub-millimeter resolution. Modern multi-detector row scanners produce such high resolution data, and these scans are so large that computer assistance in reading these data is absolutely necessary and currently one of the most active research areas in radiology.
Little has been published on CAD systems for these applications. Previous work has used only basic classifiers, relatively small training databases and limited numbers of features. This project is a collaboration that combines the expertise of the Image Sciences Institute in Utrecht in medical image analysis with the machine learning expertise of the Delft Pattern Recognition Group. The desired project goal is a set of systems that can work with multiple advanced classifiers, large databases and a large set of features.
Our approach is based on the notion that ID manifests itself in an immense variety of different patterns and that the number of examples of normal lung tissue in any representative database is far larger than the number of abnormal cases. Therefore we want to apply one-class classifiers, which reject any object that does not resemble the target class. Research questions addressed will be how to perform feature selection in one-class classifiers, and how to construct the best classifiers for this large application. Multi-class classifiers can also be useful if sufficient numbers of abnormal cases are available. We will combine one- and multi-class classifiers, which is a new research area. We hope that this combination strategy can be used to deal with missing feature values. For feature extraction, statistical segmentation methods and texture analysis methods based on multi-scale local histograms will be developed. The generality of these methods makes them applicable in other CAD products and pattern recognition applications as well.

Utilisation
CAD is now moving from the research lab into clinical practice and is one of the main topics in radiology today. It is especially interesting for applications with a large volume of exams or data per exam. Hence the large industrial interest for CAD products in chest radiography, tuberculosis screening and the analysis of high resolution chest CT scans. The world's first CAD product for chest radiographs obtained approval in the USA last year. Currently many companies are working on the integration of CAD in chest workstations. With CAD, radiologists can reduce the number of missed lesions and thus improve health care. CAD systems for the detection of interstitial lung disease are not yet on the market, but once reliable technology becomes available, there is an enormous interest from industry to develop products.
In this project we collaborate with three industrial partners, Nucletron, R2 Technology, and Philips Medical Systems, who represent the world top in X-ray acquisition technology and software development for medical image analysis and CAD. Nucletron offers digital chest units and is strong in the mass chest screening. R2 Technology is a world leader in CAD for mammography and currently working on CAD for chest CT and X-ray. Philips Medical Systems is one of the world leaders in multi-slice CT technology and medical imaging in general.

Er zijn 3 bedrijven bij dit project betrokken.