Boundary Detection in Cardiovascular Ultrasonic Images Based on Multiscale Dynamic Programming
Non-invasive ultrasonic imaging is widely used in cardiovascular studies as well as clinical diagnostics. This is due to its non-invasiveness, low cost and easy operation. However, ultrasonic images are noisy and present artifacts, speckles, and echo dropouts. The interpretation and measuring of these images are usually carried out manually by clinical experts with the assistance of computerized interactive analysis systems. One of the main tasks of the operator is to trace the boundaries of objects such as vessel walls and heart chambers. Although research has proven that manually tracings correlate reasonably well with true anatomy, intra- and inter-observer variability is high because of the inherent subjectiveness. Besides this, the manual procedure is time consuming and laborious. Hence automated analysis techniques including object boundary detection in medical ultrasonic images are highly desirable.
The present Thesis deals with problem of automated and quantitative ultrasonic measurement of the human superficial arteries as well as heart chambers. In particular, for the artery images, a new boundary detection approach is reported. By applying a multiscale dynamic programming algorithm, approximate positions of the artery wall are first estimated in a coarse scale image. Then, under the guidance of this estimate, the exact boundary positions are detected in a fine scale image. For both coarse and fine scale images, dynamic programming is applied for finding a global and optimal solution to the problem of minimizing a cost function. The cost function is a weighted sum of terms, in the form of fuzzy expressions, representing multiple image features and geometrical characteristics of the boundaries. Prior to detection, the weights of the cost function are adjusted by a training procedure using human expert tracings. The method makes it possible that the human intervention, when needed, also function in the framework of optimality in minimizing the cost function. The resulting detection algorithm is robust, reduces the amount of human interventions and consequently reduces inter- and intra-operator variability. A thorough evaluation of the method as applied to clinically acquired artery images was performed. The artery measurement results showed a high correlation between automated and manual measurements (r=0.98-1.00). Inter-observer variability decreased by 82% and 50% as compared to the manual system for common carotid artery lumen diameter and wall-thickness, respectively. The overall analysis time was reduced by two thirds.
A specific problem in relation to automated boundary detection in ultrasonic imagery concerns noise suppression. The Thesis presents a novel technique for edge-preserving noise reduction referred to as IsoIntensity Directional Smoothing (IISD). Pre-processing the artery images using IIDS did not improve boundary detection. Although originally intended for ultrasonic imagery, it can be anticipated that IIDS performs better for other types of images, specifically those with Gaussian type noise and low signal-to-noise ratio. This was demonstrated by a quantitative comparison between IIDS and some well-known filters belonging to the same category.
multiscale dynamic programming