A Selection and Reduction Approach for the Optimization of Ultrasound Carotid Artery Images Segmentation

The segmentation of the carotid artery wall is an important aid to sonographers when measuring intima-media thickness (IMT). Automated and completely user-independent segmentation techniques are gaining increasing importance, because they avoid the bias coming from human interactions. However, automated techniques still underperform semi-automated IMT measurement methods. Automated techniques cannot reproduce human expertise in selecting the optimal point where IMT should be measured. Hence, superior intelligence must be embedded into automated techniques in order to overcome the performance limitations. A possible solution is to extract more information from the image, which could be obtained by an accurate analysis of the image at pixel level. In this study, we applied a feature selection and reduction approach to ultrasound carotid images, and measured 141 features for each image pixel and supposed that a pixel could belong to one of three classes: artery lumen, intima or media layer, or the adventitia layer. Among several approaches that are available for dimensional reduction, we chose to test three based on the Rough-Set Theory (RST): the QuickReduct Algorithm (QRA), the Entropy-Based Algorithm (EBR) and the Improved QuickReduct Algorithm (IQRA). QRA achieved the best performance and correctly classified 97.5 % of the pixels on a reduced testing image dataset and about 91.5 % for a large validation dataset. On average, QRA reduced the complexity of the system from 141 to 8 or 9 features. This result could represent a pilot study for developing an intelligent pre-classifier to improve the image segmentation performance of automated techniques in carotid ultrasound imaging.