Automatically designed machine vision system for the localization of CCA transverse section in ultrasound images

Abstract

The common carotid artery (CCA) is a source of important information that doctors can use to evaluate the patients’ health. The most often measured parameters are arterial stiffness, lumen diameter, wall thickness, and other parameters where variation with time is usually measured. Unfortunately, the manual measurement of dynamic parameters of the CCA is time consuming, and therefore, for practical reasons, the only alternative is automatic approach. The initial localization of artery is important and must precede the main measurement. This article describes a novel method for the localization of CCA in the transverse section of a B-mode ultrasound image. The novel method was designed automatically by using the grammar-guided genetic programming (GGGP). The GGGP searches for the best possible combination of simple image processing tasks (independent building blocks). The best possible solution is represented with the highest detection precision. The method is tested on a validation database of CCA images that was specially created for this purpose and released for use by other scientists. The resulting success of the proposed solution was 82.7%, which exceeded the current state of the art by 4% while the computation time requirements were acceptable. The paper also describes an automatic method that was used in designing the proposed solution. This automatic method provides a universal approach to designing complex solutions with the support of evolutionary algorithms.

Introduction

The biomedical image analysis is nowadays used in many areas of medicine. It may include the diagnosis of various diseases, object detection, teaching medical students, or just the management and sorting of images for later processing. The main contribution of this paper is a proposal and subsequent evaluation of automatic construction of a machine vision system for automatically solving the task of detecting a common carotid artery (CCA) in ultrasound video sequences. Using ultrasound for an analysis of the CCA is currently a relatively hot topic and offers certain benefits when compared with other approaches. The solution found by a modified semi-automatic system exceeds the current state-of-the-art localization of CCA by almost 4% and, at the same time, the solution is found automatically by machine. Another contribution of this work is the publication of a video database that was created and used in experimental detection of arteries.

Nowadays, the methods for object detection in medical images, such as methods for artery localization, are mainly designed by an expert. The expert performs the concatenation of partial image processing methods to achieve the desired result based on their acquired experience and knowledge. The design of a system is different for different tasks, it is time consuming and the resulting success is largely dependent on the experts’ experience, the amount of tested combinations, and the settings of partial image processing methods. Methods designed by expert are often based on geometrical analysis of the object being localized. The main disadvantage of such methods is that the expert has to implement a complex system which is sensitive to variations in source data. During the sonographic examination, the operator can set many parameters such as frequency, focus, zoom, time gain compensation apart from the probe orientation and pressure on the tissue being examined. Together with human body variability, these factors lead to appreciable variations in the image data being analysed, which makes it difficult to find a robust method for solving a given problem.

This article presents a new way of designing a method for artery localization which increases the accuracy of detection and automatize the design process. The proposed method takes advantage of genetic programming to automatically finding the best solution using a training data set which includes a wide range of above-mentioned variations of the image object being localized. Genetic programming, which is used in our implementation, is supplemented with expert knowledge of the problem domain, in this case the expert knowledge of US image processing. The expert therefore only specifies the necessary domain knowledge to solve the problem and the complete design is seen to by the evolutionary process. The domain knowledge suitable for the artery localization was inspired by the basic methods described [1], [2].

The main purpose of this paper is the design of the automatic, robust and reliable artery localization method. The importance of localization method is obvious because it is an essential first step preceding non-invasive monitoring of static and dynamic parameters of the arteries [3], [4] such as arterial stiffness, arterial wall thickness [5], [6], vessel diameter or intima-media thickness [7]. Automatic and precise measurement of these parameters is highly requested by doctors. Recently used localization procedures are either not automatic or often fail, therefore they are not suitable for automatic measurement of above-mentioned parameters. An additional effort from doctors is required during performing such measurement. However, the proposed method is automatic and precise enough for direct utilization in many recent systems [8], [9], [10]. The precision of subsequent measurements performed on localized arteries depends on accuracy of localization method and that is why the high accuracy is so important. In proposed solution, the highest possible accuracy is ensured by the utilization of computer and genetic programming in the process of designing. Measurement of above-mentioned parameters is very important due to its correlation with patients risk of cardiovascular events (CVE) [7], [11], [12]. Recently, the cardiovascular diseases have been responsible for about 30% of all deaths and this number has been increasing due to the lifestyle in industrialized countries.

The problem of object recognition in US images is a very challenging task. In this case, using an automatic method based on modified genetic programming, a technique for artery localization was invented which surpassed the success of the artery localization method designed by a human.

The rest of the paper is organized as follows. The second section summarizes the current state of the art and compares it with the approach proposed in this article. The third section introduces training a validation sets. The fourth section refers to the shortcomings of current approaches and describes the proposed solution of the construction of machine vision systems. Applications of the proposed method, an evaluation of the results and a review of these results are discussed in sections five and six. The last section summarizes the results obtained in this paper and outlines opportunities for further work.