Lesion-specific coronary artery calcium quantification for predicting cardiac event with multiple instance support vector machines.

Conventional whole-heart CAC quantification has been demonstrated to be insufficient in predicting coronary events, especially in accurately predicting near-term coronary events in high-risk adults. In this paper, we propose a lesion-specific CAC quantification framework to improve CAC's near-term predictive value in intermediate to high-risk populations with a novel multiple instance support vector machines (MISVM) approach. Our method works on data sets acquired with clinical imaging protocols on conventional CT scanners without modifying the CT hardware or updating the imaging protocol. The calcific lesions are quantified by geometric information, density, and some clinical measurements. A MISVM model is built to predict cardiac events, and moreover, to give a better insight of the characterization of vulnerable or culprit lesions in CAC. Experimental results on 31 patients showed significant improvement of the predictive value with the ROC analysis, the net reclassification improvement evaluation, and the leave-one-out validation against the conventional methods.