An image understanding system is proposed to analyze cardiac dysfunction found by combining information from multiple, noninvasive imaging modality studies. The system is a model for intelligent multimodality image understanding in general, but is applied to a very specific problem: detection and rating of left ventricular (LV) aneurysms. Key features of the system are: 1) its ability to handle and quantify uncertain or partial image- derived information in a concise way using probabilistic evidential reasoning, 2) its ability to fuse pertinent relative information from independent diagnostic images of the same patient to achieve an algorithm-assembled, consensus, quantitative opinion of cardiac shape and motion, and finally 3) to arrive at a decision- level set of numbers that quantify and localize left ventricular aneurysm formation for each patient. The availability of data such as described in 3) will enable more precise prognostic or diagnostic risk classification for patients, making therapy alternatives more rational. Because of the subjective probabilistic reasoning strategy (based on the principle of maximum entropy) the final quantitative results will carry not only the system's assessment of a particular patients' heart, but also the degree of confidence that the automated analysis system has in the result it presents. This confidence is increased when similar LV motion and shape is perceived by the multiple imaging modalities. The proposed system design is influenced by current artificial intelligence and image understanding technologies, but remains firmly grounded in more classical mathematical and statistical methodology. The system will be initially tested using radionuclide angiographic (RNA) and two-dimensional echocardiographic (2DE) semi- quantified (manual tracing of LV borders) and then fully quantified (by the system's own border-finding algorithm) studies and compared with qualitative readings, quantified contrast ventriculograms, and then with surgeon's and pathologist's measurements and reports. Preliminary studies show that 1) RNA left lateral view qualitative studies correlate well with pathology, 2) that complementary 2DE views help resolve RNA problems with finding the anterior LV wall and 3) quantification of LV motion and shape better defines the spectrum between normal and grossly abnormal. Subsequent incorporation of magnetic resonance images (which have improved LV boundary definition) into the evaluation system will test its ability to adapt to and profit from new types of data.
