Skip to main content

Georgia D. Tourassi

Adjunct Professor in the Department of Radiology
2424 Erwin Road, Suite 302, Duke Advanced Imaging Laboratories (Dai, Durham, NC 27705
2424 Erwin Road, Suite 302, Duke Adv Imaging Laboratories (Dai-Labs), Durham, NC 27705


My research focuses on the development and novel application of computer vision and computational intelligence algorithms for medical image analysis content-based image retrieval, medical decision making, and effective clinical integration of computer aids. Specifically, there are five ongoing studies:

1.Information-Theoretic CAD System in Mammography: Computer-assisted diagnosis is an active field of research with several commercial CAD products available for the detection of breast cancer in screening mammograms. The available products are used as “black-boxes” to provide radiologists with a second opinion regarding the presence of potential abnormalities in the breast images. However, the currently used “black-box” CAD paradigm is rather limited. A CAD system that is more interactive and capable of justifying the opinions it provides may help radiologists’ cognitive process more effectively. Moreover, as clinical image libraries grow rapidly in Radiology, contemporary CAD systems should be able to capitalize on accumulating image data without requiring painstaking retraining or recalibration. We have been developing an interactive, knowledge-based CAD system that relies on content-based image retrieval and information theoretic principles. The system is designed to provide evidence-based decision support regarding the presence of potential abnormalities in a query mammogram by comparing the unknown query case with known cases stored in a knowledge database. The main advantage of the system is its ability to capitalize on an adaptive knowledge database where new mammographic cases can be continuously deposited without disrupting the system’s operation. Thus far, our laboratory studies have shown competitive detection performance, ability to transfer knowledge across image databases, and multiplatform adaptability (i.e., robust performance in screen-film mammograms, digital mammograms, and breast tomosynthesis data).

2.Building and Mining Knowledge Databases of Imaging Data in Radiology: Although knowledge-based systems are adaptive and flexible, their clinical application in Radiology is often restricted due to the computational demands of maintaining and querying a continuously growing databank of radiologic images. To address this concern, we have been exploring indexing schemes for improving the speed of analysis without compromising the diagnostic performance of the knowledge-based system. The indexing schemes are investigated in two different capacities (i) as the basis of search mechanisms to sift fast through the knowledge database, and (ii) as the basis of a selection mechanisms to build a concise knowledge database that is still effective but easier to maintain. Initial results with an entropy-based indexing scheme for our mammography CAD system are extremely encouraging suggesting a 75% reduction in computational demands without any compromise in diagnostic performance.

3.Reliability Analysis of CAD Technology: The development of a CAD system involves careful optimization so that the diagnostic performance of the system is maximized for the target patient population. When the system is deployed for clinical use, the radiologist is informed about the system’s expected diagnostic yield. However, the diagnostic yield may vary substantially from case to case due to the variable complexity of each case. Thus, the radiologists are left unguided as to how to integrate the CAD opinion in their decision making process on a per case basis. We have been developing a robust computational technique that enables a CAD system to affirm its user when it offers a highly reliable opinion and alert him/her when a questionable opinion is offered. Thus, the technique is a mechanism for risk stratification for CAD technology. The proposed technique monitors the system’s accuracy in a dynamically selected sample of known cases, similar to the one in question. If the accuracy measured on the selected sample is significantly lower or higher than what expected on average, the CAD system will inform its user accordingly. The long-term goal of this research is to improve the human-computer communication. We are working towards this goal by developing an effective and generalizable mechanism for patient-specific customization of CAD technology to improve the efficacy of computerized decision aids in Radiology.

4.Advanced Computational Intelligence Techniques for CAD Optimization: We are currently pursuing advanced computational intelligence methods such as genetic algorithms and particle swarm optimization for multi-objective optimization of our CAD systems using clinically relevant objective functions.

5.Neutron Imaging: This represents a new form of biomedical imaging in which 3D tomographic images are formed of individual stable isotopes in the subject by illuminating the body with neutrons and collecting the characteristic emitted gamma rays in a tomographic geometry. This research is conducted in collaboration with Triangle Universities Nuclear Laboratory.

Current Appointments & Affiliations

Adjunct Professor in the Department of Radiology · 2011 - Present Radiology, Clinical Science Departments

Education, Training & Certifications

Duke University · 1993 Ph.D.