Improving Automatic Coronary Stenosis Classification using a Hybrid Metaheuristic with Diversity Control

In the present work, an improved Hybrid Metaheuristic with a Diversity Control is proposed. The method is focused on finding an optimal feature subset, which is able to improve the performance in the automatic classification of coronary stenosis cases. Compared with traditional evolutionary computing approaches, which considers only the best individuals of a population, the proposed strategy considers the worst individuals under certain conditions. By applying the diversity control, the search space which involves a high-dimensional complexity, is explored widely. In consequence, the feature selection frequencies trends to be uniform, decreasing the probability of premature convergent results and local-optima solutions. The experiments involved the formation of a dataset, consisting of $608$ instances and $473$ features, from an image database of X-ray coronary angiographies. Using the proposed strategy, it was achieved a classification rate of $0.92$ and $0.85$ in terms of the Accuracy and the Jaccard Coefficient metrics, respectively, using only $16$ of the $473$ initial features. In addition, the average time required for the classification of a single instance was $0.0003$ seconds. Based on the achieved results, the found feature subset is adequate to be used in clinical practice in support decision information systems.

The present ground-truth database of positive and negative images of coronary stenosis is available to the scientific community for research and comparison purposes.

Authors: