Research Project: Computer-Aided Diagnosis
Most CAD devices are approved or cleared as a fixed ad-hoc algorithm available to the clinician as second readers (i.e., a device used as a second opinion only after the clinician has completed a full independent clinical review of the patient data). A limited number of CAD devices have also been cleared for use as concurrent readers (i.e., the device cues can be viewed at any time during the full clinical review of the patient data). We anticipate significant continued changes in the intended use of CAD devices with the algorithms becoming more optimized, potentially adaptive, and with a larger fraction shifting to concurrent or even first reader mode (i.e., the clinical review consists of only reviewing the CAD prompts and avoiding the need for a full clinical review). This research project focuses on methodologies for combining, optimizing and creating adaptive CAD algorithms, and obtaining a better understanding of the influence the reading mode has on CAD utility. We will consider a diverse range of imaging technologies including x-ray breast imaging, computed tomography (CT) imaging of the colon and lung, and microscopy imaging of immunohistochemically-stained pathology specimens in our project.
As part of our project in CAD, we are investigating how the difference of the standalone performance of a CAD system for “easy to detect” and “difficult to detect” lesion subgroups affects the improvement of readers using the CAD system. The figure shows two CAD marks on thoracic CT scans: The CAD mark on the left detects a difficult lung nodule that was missed by all radiologists without CAD in a reader study involving six experienced radiologists. The CAD mark on the right detects a relatively easy lung nodule that was detected by all six readers without CAD.