Segmentation of kidney from ultrasound B-mode images with texture-based classification

https://doi.org/10.1016/j.cmpb.2006.09.009Get rights and content

Abstract

The segmentation of anatomical structures from sonograms can help physicians evaluate organ morphology and realize quantitative measurement. It is an important but difficult issue in medical image analysis. In this paper, we propose a new method based on Laws’ microtexture energies and maximum a posteriori (MAP) estimation to construct a probabilistic deformable model for kidney segmentation. First, using texture image features and MAP estimation, we classify each image pixel as inside or outside the boundary. Then, we design a deformable model to locate the actual boundary and maintain the smooth nature of the organ. Using gradient information subject to a smoothness constraint, the optimal contour is obtained by the dynamic programming technique. Experiments on different datasets are described. We find this method to be an effective approach.

Introduction

Many applications in medical imaging rely on a segmentation process to define object contours. Tracing the contours in medical images allows the computation of clinical measures and important quantitative interpretation. Numerous methods have been developed to segment anatomical structures and delineate object contours in various kinds of biomedical images, including MR [1], [2], [3], [4], [5], CT [6], scintigram [7], DNA [8], biopsy [9], histopathological [10], X-ray [11], and ultrasound images. The non-invasive and real-time characteristics of ultrasound have enabled its wide use in hospitals, and even the Internet [12]. However, among all medical imaging modalities, segmentation from ultrasound images is one of the most challenging tasks, because ultrasound images are characterized by weak edge information, and it is difficult to locate an object's contour accurately within a fuzzy ultrasound image.

Object contour definition from ultrasound images can be accomplished in a number of ways. Manual contour definition is a popular but time-consuming process and suffers from a very low degree of reproducibility [13]. Some methods have proposed to use an image's gray level information to identify the object borders [14]. However, it is difficult to automatically detect the object by using gray level information only. Snake-based deformable models are better solutions for image segmentation and various applications have been found for medical image analysis [15]. A deformable model typically consists of a geometric representation of the target and a set of energy functions that control the evolution of a contour model. Because the image segmentation problem is formulated as an energy minimization problem in deformable models, the improvement often focuses on initialization, energy definitions, and optimization.

Texture features, which represent the variation (or pattern) in intensity of a particular region of an image, are capable of capturing soft-tissue characteristics appearing in images [16]; therefore, they have been proven to be useful in medical image segmentation [17], [18] and analysis [19], [20], [21], [22], [23]. Texture features provide important and useful clues for the identification of organs and tissue, but with additional computational costs, especially when multiple features are needed. As mentioned, weak edge information is a serious problem in ultrasound image segmentation. This is also the reason why snake-based deformable models do not work effectively because of the lack of a strong intensity gradient. Nevertheless, edges are usually the interfaces between different tissue types in ultrasound images, so texture features should be a more reliable factor in detecting edges. Thus, the proposed algorithm combines a deformable model and texture classification for kidney segmentation, and this approach does help to alleviate the weak edge problem in our experiments.

Ultrasound imaging has been widely applied to various kinds of examinations and treatments. In this research, the focus is to segment the kidney contour from a B-mode ultrasound image which is used to assist the evaluation and diagnosis of acute pyelonephritis. For children, acute pyelonephritis is a serious disease that may result in irreversible renal scarring [32], and it can also occur in adults, especially adult women [33]. If left untreated or unresolved over a period of months or years, acute pyelonephritis will lead to scarring and the possible loss of kidney function. The sequelae of pyelonephritic scarring include but are not limited to hyposthenuria, proteinuria, and hypertension. For the diagnosis of renal diseases, kidney contour information is useful for physicians to evaluate the diseased kidney and is also an important feature used in the data fusion from different image modalities. In this study, we design a deformable model with texture-based classification to locate kidney contours. We use texture features and MAP estimation to classify pixels. A deformable model is used to locate the object based on this classification. This model is then optimized by dynamic programming along with a smoothness constraint. Because kidney ultrasound images are usually contaminated with noises and sometimes with weak edges that make contour detection a difficult task, our idea is to combine the deformable model with texture classification to make the segmentation algorithm of kidney ultrasound more reliable. In fact, most of the deformable models rely only on minimizing a cost function defined by a weighted sum of conventional energy terms, whereas the proposed deformable model adopts Laws’ texture energy measures representing kidney texture in an ultrasound image and an ellipse model approximating the kidney shape in the model initialization.

This paper is organized as follows. The overview of the proposed model is described in Section 2. Section 3 describes the model initialization. Section 4 details the approaches for pixel classification. The procedure for optimal contour finding is presented in Section 5. The experimental results and relevant discussion are given in Section 6. Section 7 contains our conclusions and plans for future development.

Section snippets

Overview

In order to overcome the difficulties of segmentation in ultrasound images, we design a generic model capable of capturing the anatomical characteristics of sonograms. The proposed model consists of two stages as described in Fig. 1. In the first stage, pixel classification is carried out by MAP estimation and texture features. The MAP estimation classifies image pixels into two classes, inside or outside the organ. In the second stage, the contour is obtained by a dynamic programming approach

Model initialization

The aim of model initialization is to provide a general approximation of the boundary location. The shape approximation selected here is an ellipse, since the shape of a kidney in a two-dimensional image is close to an ellipse. It is simple enough to be generated by using only four parameters x0, y0, a, and b by:(xx0)2a2+(yy0)2b2=1where x0 and y0 represent the ellipse's center, and a, b are the half lengths of the long and short axes, respectively.

For better performance on matching the

Feature extraction

A set of micromasks proposed by Laws [24] is used to extract a set of feature values. We selected these texture descriptors because they are separable and suitable for fast computation and provide effective measures of statistical textural characteristics. The Laws approach consists of two steps. First, the image f(i, j) is convolved with a set of micromasks h(i, j) whose outputs are called microtexture features x(i, j). The microtexture features are a two-dimensional convolution of the image

Optimization of deformable contour

After the class parameter of each pixel in the search region is determined, the edge points between different classes in the search line are easily determined. They reduce the search domain from all elements on the entire search line to a few edge points between different classes. This greatly reduces the search domain when locating the optimal contour. After pixel classification, the organ's boundary should pass the edge points between different classes. Suppose that there are nei edge points

Experimental results and discussion

In order to test the proposed segmentation method, test data, taken from National Cheng Kung University, Tainan, Taiwan, consisting of 52 randomly selected kidney B-mode ultrasound images, were acquired, without special consideration of the image quality. To evaluate the performance of the segmentation method, we need the ground truth of the test data. Manual segmentation is a popular way to build the ground truth. It can also be found that many papers compare their results with manual

Conclusion

Image segmentation in ultrasound is very important in many medical applications, such as implementing a computer image analysis system for clinical usage. Due to low image quality in ultrasound images, the segmentation task is generally non-trivial. This paper introduced a novel image segmentation method based on a deformable model with pixel classification. It uses texture characteristics associated with the MAP estimation. The boundaries of kidneys are successfully segmented from ultrasound

Acknowledgements

The authors would like to thank the guest editor and the anonymous reviewers for helpful comments, valuable suggestions, and efforts in handling this paper. It is gratefully acknowledged that this work was supported by a grant from the National Science Council, Taiwan, under contract NSC-91-2213-E-006-059.

References (35)

  • H. Zeng et al.

    Internet (WWW) based system of ultrasonic image processing tools for remote image analysis

    Comput. Methods Programs Biomed.

    (2003)
  • J. Eilbert et al.

    The variation in user drawn outlines on digital images: effects on quantitative autoradiography

    Comput. Med. Imaging Graph.

    (1990)
  • T. McInerney et al.

    Deformable models in medical image analysis: a survey

    Med. Image Anal.

    (1996)
  • Y.T. Wun et al.

    Ultrasound characterization by stable statistical patterns

    Comput. Methods Programs Biomed.

    (1998)
  • Y. Xiang et al.

    Texture-based approaches for identifying neuro-anatomical structures and electrode tracks

    Comput. Methods Programs Biomed.

    (2004)
  • D.E. Maroulis et al.

    CoLD: a versatile detection system for colorectal lesions in endoscopy video-frames

    Comput. Methods Programs Biomed.

    (2003)
  • J. Estévez et al.

    Cytological image analysis with a genetic fuzzy finite state machine

    Comput. Methods Programs Biomed.

    (2005)
  • Cited by (32)

    • A novel convolutional neural network for kidney ultrasound images segmentation

      2022, Computer Methods and Programs in Biomedicine
      Citation Excerpt :

      In addition, manual segmentation is a very time-consuming task. In order to alleviate this challenge, a number of semiautomatic segmentation methods based on images, variable volume models, and shape models have been proposed [3–7]. Zheng et al. [3] developed a graph cuts based method to segment KUS images by integrating original image intensity information and texture feature maps extracted using Gabor filters.

    • Automated localization and segmentation techniques for B-mode ultrasound images: A review

      2018, Computers in Biology and Medicine
      Citation Excerpt :

      Much focus is also given to texture analysis, which can help overcome some of these challenges as well. Older techniques often require some kind of user interaction, by manually placing markers to delineate in some way the initial segmenting curve or model [46,59], whereas more recent research has focused on completely automatic techniques on 3D ultrasound images, often integrating ad-hoc or patient-specific functions [57,58,117], as personalized medicine is becoming a main focus in medical imaging. To note are the completely automatic kidney segmentation techniques from 3D US images developed by Cerrolaza et al. [117], based on fuzzy appearance models and patient-specific alpha shapes that allows to control an ad-hoc stopping function created from the adipose tissue bands within the kidney, and by Marsousi et al. [58], who developed a technique that offers a new shape model representation that combines anatomical knowledge with prior knowledge of training shapes.

    • Registration of renal SPECT and 2.5D US images

      2011, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      Our method uses manual initialization, which consists in aligning the axes of an ellipsoid with the kidney position in a single 2D US slice (Fig. 5). After this initialization, an automatic 2D segmentation is performed on each 2D image of the set that conforms the 2.5D US image, applying a deformable contour [44,45] method and a set of Gabor filters [46–49] to capture multi-scale image features. The aim of our method is to find enough points in the US image to carry out the registration, rather than finding the whole kidney geometry.

    View all citing articles on Scopus
    View full text