Intravascular ultrasound: tridimensional quantification and analysis / Análise e quantificação tridimensional em imagens de ultra-som intravascular

AUTOR(ES)
DATA DE PUBLICAÇÃO

2009

RESUMO

INTRODUCTION: Several efforts for in vivo atherosclerotic plaque estimation has been realized worldwide. Current methods for related application are based on spectral analysis of backscattered ultrasound signals before intravascular image formation (IVUS), once there is not available techniques to perform composition analysis only from IVUS images. On this study, image processing and pattern recognition techniques has been used to classify plaques according to Intravascular Ultrasound Virtual Histology (IVUS-VH) classification. METHODS: A sample set of eight (08) coronary arteries from five (05) different patients, resulting on 96492 regions of interest for plaque analysis. For automatic segmentation of lumen-intima interface and intima-adventitia interface, active contours technique has been used. Conventional and restrictive anisotropic diffusion filters have been used for estimating the potential energy of active contours. Haralicks co-occurrence matrix and invariant moments based in Hu moments were used as features for pattern classification. Three different windows has been used on feature extraction procedure: 5x5,7x7 and 9x9 pixels. In order to analyze the influence of grey-scale quantization levels into Haralicks co-occurrence matrix, we tested four different types numbers of grey-scale levels: 32, 64,128 and 256. KNearest Neighbors algorithm as the classification rule, and we tested both Mahalanobis and Euclidian Metrics in this analysis. RESULTS: Segmentation of luminal-intimal layers average accuracy for was 72.30% and its median was 81.20%. Segmentation of intimal-adventitia layers average accuracy for was 80,86% and its median was 91,36%, for 495 manually segmented frames. Plaque classification according to IVUS-VH four classes scheme has shown not feasible with the proposed model, once average classification error-rate foe every component: 2.35%(DC), 20.51%(NC), 92.21%(FF) e 0.04%(FT). Our approach has not the ability to differentiate fibrotic tissue from fibro-fatty tissue. However, in the three classes scheme: DC, NC and (FF+FT), we achieved some encouraging results: 2.35% (DC), 20.51%(NC) e 0.04%(FT+FF), leading to an average error rate of 5,15%. With these results, a thin-cap fibroatheroma (TCFA) lesions detection algorithm for conventional IVUS images should be developed using our approach. Once TCFA plaques has been assigned to have an association with plaque rupture and sudden death, those results should be used for in vivo IVUS based TCFA detection. In an experiment related to this work, we have shown that there is an artifactual relationship between NC and DC components into VH images

ASSUNTO(S)

doença da artéria coronariana recognition automated pattern processamento de imagem assistida por computador ultrasond of intervention ultra-sonografia de intervenção reconhecimento automatizado de padrão assisted image processing by computer coronary artery disease

Documentos Relacionados