Dr. Tourassi has been developing a robust computational technique based on transductive algorithms that enables a CAD system to assess the expected reliability of its response for a specific patient, and inform the radiologist accordingly. The proposed technique monitors the system’s accuracy in a dynamically selected sample of known, relevant cases, similar to the one in question. The measured accuracy is used as a surrogate measure of the system’s expected reliability on the unknown case. By providing a robust computational technique for quality assessment monitoring of a CAD system on a case-by-case basis, she aims to facilitate better communication between the CAD system and the user, thus more effectively utilizing decision support systems in Radiology. Figure 1 illustrates the proposed reliability analysis framework, integrated with a typical CAD system.

With a series of pilot studies using in-house CADe and CADx systems, we have demonstrated that the proposed reliability measure carries additional information of clinical value. We have also demonstrated empirically that the reliability analysis framework is applicable to diverse decision algorithms and can be adapted to both prediction and classification tasks.

Representative Publications

• P.A. Habas, N.H. Eltonsy, A.S. Elmaghraby, J. Zurada, G.D. Tourassi, “Reliability analysis of CAD decisions,” Medical Physics 34: 763-772 (2007).
• P.A. Habas, J.M. Zurada, G.D. Tourassi, “Case-Specific Reliability Assessment for Improved False Positive Reduction with an Information-Theoretic CAD System”, International Workshop on Digital Mammography, Tucson AZ, July 25-28, 2008.
• M.A. Mazurowski, J.M. Zurada, G.D. Tourassi, “Reliability assessment of ensemble classifiers: application in mammography”, International Workshop on Digital Mammography, Tucson AZ, July 25-28, 2008.