Diagnostic Imaging: Finding the Critical Features

¹ Office of Science & Technology, Center for Devices & Radiological Health, Food & Drug Administration, Rockville, Maryland 20857

The clinical interpreter of a diagnostic image has the task of reducing the information displayed— in the order of a megabyte — to a set of classifications that convey answers to specialized diagnostic questions — in the order of just a few bits. A high-dimensional feature space is thereby reduced to a low-dimensional feature space. This paper reviews the features available in x-ray imaging, magnetic resonance imaging, and diagnostic ultrasound that derive from the geometry and the physics of the imaging process. An example is drawn from diagnostic ultrasound to demonstrate an objective assessment of a small feature space and the utility of adding back additional features. The assessment is based on the methodology of the receiver-operating characteristic (ROC) curve. Motivation is provided for use of this fundamental tool from statistical decision theory, together with definitions of the quantities involved in scoring performance.