| Computer algorithms that analyze medical image data to detect abnormalities such as tumors have recently become commercially available and are already widely used: Computer-aided diagnosis (CAD) is one of the most rapidly expanding areas in medical imaging. Current systems do, however, not yet operate at the level of human experts. I believe this is caused by the fact that current systems are {\em static}: after an initial training stage they do not change anymore. This is comparable to a human who blindly performs a task according to some initial instructions but does not learn anything by doing, which is the essence of the training of a radiologist. The goal of this project is to develop tools to design {\em dynamic} CAD systems. These systems are radically different in that they learn from experience. They autonomously identify difficult and unfamiliar cases and seek expert feedback when needed, while processing vast amounts of data that are nowadays available. With this expert feedback, they continually adjust their internal algorithms. To make this possible, a rich set of features to describe and compare images is needed. One subproject focuses on better, more general and more invariant features and explores how to measure the similarity between candidate lesions. A second subproject will develop methods to automatically select cases for which the system seeks feedback. A framework will be constructed to identify common errors of the system and to train classifiers that correct these errors. It will also be investigated what forms of expert feedback are most efficient and what image information is used by experts to detect abnormalities. |
