Assessment of image features for vessel wall segmentation in Intravascular Ultrasound images

L. LO VERSIO; M. DEL FRESNO; I. LARRABIDE

Congreso; Computer Assisted Radiology and Surgery; 2015

Purpose: Intravascular ultrasound (IVUS) provides axial gray-scale images, allowing the assessment of vessel wall and surrounding tissues. In the past, various studies have described automatic segmentation of luminal boundary and media-adventitia interface considering different image features. The aim of the present study is to evaluate image features found in the literature for segmenting IVUS images. Methods: Several image filters, textural descriptors, noise and spatial measures were took into account. The likelihood of pixels belonging to a specific region of the artery is a continuous value obtained by using Support Vector Machines (SVMs). To retrieve relevant features, sequential feature selections were performed guided by the area under the Precision-Recall curves (AUC-PR). Linear correlation between features themselves and labels are used to explain the resulting subsets. Results: Subsets of relevant image features for segmenting lumen and arterial wall are obtained and SVMs trained with these features accurately identified the different regions. The experimental results were tested against reference segmentation from available datasets of IVUS images, reaching values of AUC-PR up to 0.97 and of F- score close to $0.92. Conclusion: Noise-reduction filters were chosen first in the feature selection process denoting their relevance when identifying vessel wall. Haralick´s textural features and spatial descriptors provide additional information that improves the classifier accuracy. Finally, Laws´ textural features had less relevance in the selection process.