Model observers for medical imaging. In the early days of image processing, image quality was judged by classical metrics such as signal-to-noise ratio (SNR) or mean-square error (MSE). However, it is has become widely accepted that these metrics do not necessarily reflect the degree to which the images help to achieve a desired goal, such as accurate performance of a diagnostic task in medical imaging. Therefore, such metrics are being replaced by so-called task-based evaluations. Ideally, if a human expert will use the images to perform a diagnostic task, then human experts should be used to measure image quality in this way. However, it is not always practical, particularly in early stages of algorithm of hardware design, to ask experts to look at large numbers of images. Thus, computer algorithms that model the expert—so-called model observers—are now widely used in the optimization and testing of medical imaging methods.
In my work, I have demonstrated that existing model observers have major limitations, and I have proposed that they can be derived more accurately by using machine learning techniques to create model observers that emulate humans.
In this way, I am creating a suite of model observers, that will shed light on a comprehensive set of clinical tasks, offering greater accuracy than existing methods. This research is significant because it will yield an evaluation methodology that could potentially be used very widely by the research community, underpinning the development of a variety of kinds of imaging hardware and software. This project is supported by an NIH R01 grant which I was fortunate to be awarded on the first try. Siemens Medical Solutions and Philips Healthcare have expressed interest in employing these model observer techniques in their imaging system evaluations.
Analyzer-based phase contrast X-ray imaging. X-ray phase-contrast imaging permits visualization of soft-tissue structures that are not detectable by use of conventional X-ray radiographic methods, and typically delivers a much smaller radiation dose to the subject. Because of this, it holds great potential for a wide range of human, small-animal, and microscopic bioimaging applications. Analyzer-based phase-contrast imaging utilizes a semiconductor crystal, called an analyzer, to selectively measure certain diffracted components of an X-ray wavefield illuminating the object to be to imaged. Such methods can reveal diagnostic information that is not measured by conventional absorption-based methods.
To further this research, I founded the Advanced X-ray Imaging Laboratory (AXIL) at IIT, where we are developing a prototype tabletop imaging system based on the analyzer concept. Because the project involves development of experimental instrumentation, it has taken several years of effort; however, the system is nearing its final stages of completion, and we are beginning to obtain preliminary image results. This work builds upon earlier work in which I investigated methods of image reconstruction, and the physics of image formation, for this type of imaging. These early analytical steps were corroborated by experimental studies performed at Brookhaven National Laboratory.