 ANN Project
Assessment of coronary artery disease typically involves interpreting
cardiovascular nuclear medicine images, which measure myocardial perfusion
(blood flow) distribution. The assessment typically involves the interpretation
of two types of images, obtained at stress (during patient exercise) and at
rest (several hours after the stress study is performed). A recent clinical
development in the diagnostic process may obviate this second test, however. It
has been observed that myocardial thickening (the rate of thickening of the
myocardial mass during the cardiac cycle) can also be used as a possible
indicator of myocardial viability. In particular, clinical studies suggest that
myocardial thickening information, used in conjunction with stress perfusion
information, can serve as a measure of the redistribution of perfusion in
viable (but infarcted) myocardial tissue. This would obviate the need for
performing the second image acquisition (at rest), since only myocardial
thickening and stress perfusion information would be required, and both of
these types of information can be obtained simultaneously during one
acquisition. Importantly, by performing a single clinical test to measure
stress perfusion and thickening information, significant improvements can be
realized in terms of patient discomfort, risk, and time. The goal of this
project is to exploit this concept by "predicting" the at-rest perfusion
information from the thickening and stress perfusion images.
With this in mind, a neural network approach is being explored to analyze and
process thickening and stress perfusion information to predict associated
perfusion redistribution information. A multilayer, backpropagation neural
network is trained to predict the redistribution information from two other
types of images: stress perfusion and myocardial thickening using SPECT
imaging. The significance of this approach is two-fold: (i) the predicted
reversibility information obviates the additional acquisition of delayed images
(with the patient at rest), and (ii) the neural network approach represents a
novel way with which to analyze and predict images from other images.
The ANN-derived image is also being used as input to a knowledge-based system
(PERFEX) that further interprets both the stress and
predicted image information and provides an overall diagnostic interpretation.
The results obtained experimentally in predicting the occurrence of
reversibility using the ANN-based method described above, based on tests
conducted with a 211 defect area data set, are as follows:
- Sensitivity: 74%
- Specificity: 55%
- Overall Accuracy: 72%
These preliminary results show the viability of this approach. At present,
different and more complex data configurations and ANN topologies are being
explored. Current research thrusts center on improving the predictive
performance in terms of accuracy, reliability, and the ability to make
reversibility predictions with greater quantitative granularity.
Project Members:
References:
- "Analyzing and Predicting Images Through a Neural Network
Approach,"
L. de Braal, N. Ezquerra, E. Schwartz, C.D. Cooke, and E.
Garcia. Proceedings of the Visualization in Biomedical Computing 1996
(VBC 96) Conference, Hamburg Germany, September 22-25 1996.
- "Connectionist Methods in Medicine,"
N. Ezquerra, invited
presentation at the International Congress on Knowledge Engineering,
Seville, Spain, October 1992.
- "A Neural Networks Approach to Medical Image Interpretation,"
A.
Pazos, N. Ezquerra, F. Martin, and V. Maojo. Proc. World Congress on
Medical Informatics (MEDINFO `92), Geneva, Switzerland, September 1992.
|
|
| | |
