In this paper we propose an image-based approach for in-vivo assessment of IVUS images. The method discriminates plaque components into four classes: calcium, necrotic core, fibrous and fibro-fatty. We employ the IVUS frames characterized by virtual histology (VH) for tissue labeling. As a result, we avoid the demerits of visual assessments of observers while at the same time the longitudinal resolution of VH is increased. To describe the textural properties of the tissue classes five different features are extracted from IVUS images. The features are computed by using multiple window sizes so that their values are adapted to the varying heterogeneity of the local patterns. In the next stage, we apply an effective feature selection algorithm on the combined feature space of original features, yielding a small subset of discriminating and non-redundant features. The retained features are used for tissue classification via an SVM classifier. The method is validated against the available VH reference data. The experimental results show that the proposed approach achieves an average accuracy of 81%. This result is obtained by a reduced subset comprising 34 features of the appropriate type and scale of extraction.
展开▼
