TY - JOUR
T1 - Three-dimensional mapping of acute ischemic regions using artificial neural networks and tagged MRI
AU - Azhari, Haim
AU - Oliker, Shai
AU - Rogers, Walter J.
AU - Weiss, James L.
AU - Shapiro, Edward P.
N1 - Funding Information:
Manuscript received March 24, 1994; revised January 29, 1996. This work was supported by the US-Israel Educational Fund (Fulbright) and later by the NIH Fogarty under Grant no. 5-F05-TW04415-02 ICP(5); by the NHLBI Ischemic Heart Disease SCOR under Grants HL-17655-16, R01-HL-43722 (J.L.W.), R01-HL-4622302 (E.P.S.); and the Chesapeake EducatiodResearch Trust. *H. Azhari is with the Department of Biomedical Engineering, Technion, Israel Institute of Technology, Haifa 32000, Israel (e-mail: ham@ biomed.technion.ac.il). S. Oliker is with the Department of Biomedical Engineering, Technion, Israel Institute of Technology, Haifa 32000, Israel. W. J. Rogers, J. L. Weiss, and E. P. Shapiro are with the Division of Cardiology, Johns Hopkins School of Medicine, Baltimore, MD 21218 USA. Publisher Item Identifier S 0018-9294(96)03989-4.
PY - 1996/6
Y1 - 1996/6
N2 - Many methods for mapping ischemic myocardial regions by functional analysis have been suggested. However, the complicated relationship between myocardial function and perfusion, and the inherent limitations of the imaging techniques used, have led to a generally low mapping accuracy. We show herein, that highly accurate mapping can be obtained by combining tagged magnetic resonance imaging (MRI), three-dimensional (3-D) analysis, and artificial neural networks. Nine canine hearts with acute ischemia were studied using multiplanar tagged MRI. Twenty-four myocardial cuboids were tagged in each heart and reconstructed in 3-D at end diastole (ED) and end systole (ES). The cuboids were arranged in three slices approximately 1 cm thick and covered most of the left ventricle (LV). Transmural thickening and endocardial area strain were calculated for each cuboid. Applying a post- mortem (PM) analysis, the percent ischemia in each cuboid was estimated using monastral blue dye; the PM analysis served as a 'gold standard.' An artificial neural network (ANN), designed to estimate the percent ischemia in each cuboid from the functional indexes, was then created. The ANN 'learned' the function-ischemia relationship in 192 cuboids taken from eight of the hearts and was asked to estimate the percent ischemia in the 24 cuboids of the ninth heart. The process was repeated nine times, each time using a different heart as test case. The average accuracy of mapping, i.e., the accuracy with which the ANN has mapped the normal and ischemic cuboids using the functional parameters, was 87.5% ± 7.8 (s.d.). This accuracy was superior to the accuracy obtained by optimal thresholding of the same thickening (80.1%) and endocardial strain (76.9%) data.
AB - Many methods for mapping ischemic myocardial regions by functional analysis have been suggested. However, the complicated relationship between myocardial function and perfusion, and the inherent limitations of the imaging techniques used, have led to a generally low mapping accuracy. We show herein, that highly accurate mapping can be obtained by combining tagged magnetic resonance imaging (MRI), three-dimensional (3-D) analysis, and artificial neural networks. Nine canine hearts with acute ischemia were studied using multiplanar tagged MRI. Twenty-four myocardial cuboids were tagged in each heart and reconstructed in 3-D at end diastole (ED) and end systole (ES). The cuboids were arranged in three slices approximately 1 cm thick and covered most of the left ventricle (LV). Transmural thickening and endocardial area strain were calculated for each cuboid. Applying a post- mortem (PM) analysis, the percent ischemia in each cuboid was estimated using monastral blue dye; the PM analysis served as a 'gold standard.' An artificial neural network (ANN), designed to estimate the percent ischemia in each cuboid from the functional indexes, was then created. The ANN 'learned' the function-ischemia relationship in 192 cuboids taken from eight of the hearts and was asked to estimate the percent ischemia in the 24 cuboids of the ninth heart. The process was repeated nine times, each time using a different heart as test case. The average accuracy of mapping, i.e., the accuracy with which the ANN has mapped the normal and ischemic cuboids using the functional parameters, was 87.5% ± 7.8 (s.d.). This accuracy was superior to the accuracy obtained by optimal thresholding of the same thickening (80.1%) and endocardial strain (76.9%) data.
UR - http://www.scopus.com/inward/record.url?scp=0030175413&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=0030175413&partnerID=8YFLogxK
U2 - 10.1109/10.495281
DO - 10.1109/10.495281
M3 - Article
C2 - 8987266
AN - SCOPUS:0030175413
SN - 0018-9294
VL - 43
SP - 619
EP - 626
JO - IEEE Transactions on Biomedical Engineering
JF - IEEE Transactions on Biomedical Engineering
IS - 6
ER -