Identification of the breast edge using areas enclosed by iso-intensity contours
Introduction
According to the Cancer Association of South Africa, breast cancer is the most common form of cancer among women worldwide and second most common among women in South Africa (http://www.cansa.org.za). Successful treatment relies on early detection and the most common method of diagnosing cancer in its early stages is X-ray mammography [1], [2]. A significant problem with mammography is the variability of diagnoses among radiologists and computer-aided diagnosis (CAD) methods (that consistently highlight regions on mammograms that might warrant further examination) are being developed to assist the radiologist.
One of the first steps in CAD is the segmentation of the image into background and breast. This has the advantage of simplifying further processing of the image (by eliminating the background) and also provides a reference for the alignment of views when two views are being compared. Knowledge of the breast edge also sometimes helps with the identification of large masses that have distorted the outline of the breast.
Although many methods have been developed to detect the breast edge in mammograms (details in Section 2), very few researchers use borders drawn by radiologists to evaluate the automated fits and even fewer quantitatively evaluate results. In this paper, a method using areas enclosed by iso-intensity contours is presented as an improvement to the basic thresholding algorithm. The effect of various pre-processing methods on the accuracy of automated borders is investigated. The algorithms developed are tested on 25 mammograms, for which automated borders are quantitatively compared to manual borders drawn by three radiologists. Results are also compared to those obtained from a standard global thresholding method.
The paper is structured as follows: an overview of some methods that have previously been used to find the breast border is given in Section 2. The iso-intensity border detection method is described in Section 8. The results are presented and discussed in Section 10. The paper is summarised in Section 11.
Section snippets
Overview of breast border detection methods
Many methods have been used to detect the breast border in mammograms, including thresholding, tracking, artificial neural networks and modelling (of background and breast border). These methods are briefly described.
Thresholding
For mammograms, thresholding usually involves selecting a single grey-level from an analysis of the grey-level histogram, to segment the mammogram into background and breast tissue. All pixels with grey-levels less than the threshold are marked as background and the rest as breast. Thresholding uses only the grey-level histogram and no spatial information is considered. Therefore, the major shortcoming of thresholding is that there is often an overlap between the grey-levels of objects in the
Tracking
Tracking the breast border involves implementing a tracking algorithm that marks a pixel as a border pixel if it satisfies certain criteria.
Yin et al. [6] used a four-connectivity tracking algorithm to identify the border. The results were not evaluated.
Bick et al. [7] identified unexposed and direct-exposure regions in the mammogram and generated a border surrounding the valid breast border by combining grey-level histogram analysis with morphologic filtering. A closed, eight-connected border
Artificial neural networks
Suckling et al. [13] used multiple, linked self-organising neural networks to segment the breast into four components: background, pectoral muscle, fibro glandular tissue and adipose tissue. This method had the advantage of simultaneously identifying the background and pectoral muscle, but no evaluation of the background segmentation results were given.
Models of the background
Chandrasekhar and Attikiouel [14] used the Weierstrass approximation theorem as a basis for fitting a surface to the background [15]. The method was tested on 58 images and results were evaluated by visual comparison with the original images using pseudo-colour. The algorithm [16] was further tested on 28 images from the MIAS database, where all images gave clear breast-background segmentation. Results were again not evaluated by a radiologist. The fully automated method [17] was tested on the
Models of the breast border
Morton et al. [18] and Goodsitt et al. [19] reported on a breast border detection algorithm using a two-pass, model-guided edge-tracking algorithm which, when compared to manually traced out borders, yielded an average root-mean-square difference of 1.4 mm. The algorithm was tested on more than 1000 mammograms and the border was accurately found in about 95% of the images. It was not stated whether radiologists traced out the manual borders.
Ojala et al. [20] used grey-level histogram
Iso-intensity breast edge detection
Very few researchers have used borders drawn by radiologists to quantitatively evaluate the results of the automated borders. Of the two research articles [19], [20] encountered where quantitative evaluation results were quoted, it was not specified whether or not radiologists drew the reference borders. This is important if the breast edge in a mammogram is not clear, as is the case of most of the images used in this study.
Basis of the method
The breast and background form the two largest contiguous regions on a mammogram, with the background dominating at the low grey-levels. Highnam et al. [22](Fig. 1) provides a good description of the breast edge based on how the breast is compressed during mammography. The portion of the breast between the compression plates is of equal thickness, but some of the breast bulges out towards the edges. This bulge is mostly composed of fat, except near the nipple, and does not form a straight
Application of algorithm to simulated mammograms
The algorithm was first applied to two simulated mammograms. The results of applying the algorithm to the simulated image with noise can be seen in Fig. 7.
The results of applying the breast border detection algorithm to the simulated images is shown in Table 3.
The breast edge along the line equivalent to line D in Fig. 1 has been successfully detected. The results demonstrate that pre-processing with the top-hat kernel is generally more accurate than pre-processing with the Lorentzian kernel.
Summary
A novel, simple method of finding the breast edge using areas enclosed by iso-intensity contours was presented that improves on traditional thresholding methods for segmentation, by incorporating spatial information into the segmentation. The method does not rely on models of the breast or background and borders. Results were evaluated by comparison to breast borders drawn by three radiologists in their normal working environment. The effect of various pre-processing methods on the accuracy of
Acknowledgments
The authors would like to thank the staff of Addington Hospital for assistance with mammograms and University of KwaZulu-Natal, South African Department of Labour and the South African Medical Research Council (Grant: E161/02) for funding.
Jayanethie Padayachee graduated with a BSc (Hons) from the University of Natal (South Africa) in 1995 and with an MSc from the University of Cape Town in 1997. Employed at the iThemba Laboratory for Accelerator Based Sciences from 1996 to 2002 as a materials scientist, working on applying maximum entropy methods to the analysis of data from ion-beam analysis techniques. She left iThemba in 2002 to pursue a full-time PhD in computer-aided diagnosis in mammography at the University of
References (23)
- et al.
The control of breast cancer through mammography screening
Radiol Clin North Am
(1987) - et al.
Automated segmentation of digitized mammograms
Acad Radiol
(1995) - et al.
Automatic detection of breast border and nipple in digital mammograms
Comput Methods Programs Biomed
(1996) - et al.
Accurate segmentation of the breast region from digitized mammograms
Comput Med Imag Graph
(2001) - et al.
Screening for breast cancer
J Am Med Assoc
(2005) Automated classification of parenchymal patterns in mammograms
Phys Med Biol
(1998)- Blot L, Zwiggelaar R. Background texture extraction for the classification of mammographic parenchymal patterns. In:...
- Masek M, Attikiouzel Y, deSilva C. Combining data from different algorithms to segment the skin-air interface in...
- et al.
Computerized detection of masses in digital mammograms: analysis of bilateral subtraction images
Med Phys
(1991) - et al.
Computer-assisted diagnosis: the classification of mammographic breast parenchymal patterns
Phys Med Biol
(1995)
Cited by (0)
Jayanethie Padayachee graduated with a BSc (Hons) from the University of Natal (South Africa) in 1995 and with an MSc from the University of Cape Town in 1997. Employed at the iThemba Laboratory for Accelerator Based Sciences from 1996 to 2002 as a materials scientist, working on applying maximum entropy methods to the analysis of data from ion-beam analysis techniques. She left iThemba in 2002 to pursue a full-time PhD in computer-aided diagnosis in mammography at the University of KwaZulu-Natal, which she completed in 2006.
Michael J. Alport graduated with a PhD from the University of Iowa, USA in 1981. For the next 10 years, he specialised in plasma waves and instabilities and the industrial uses of ion implantation. He spent a number of sabbatical visits at the University of Maryland and West Virginia University working on the US space program.
More recently, Prof. Alport switched research topics somewhat to study turbulence in water waves and the application of imaging and neural network techniques to solve industrial problems. He has written over 21 papers, presented papers at 75 South African and International conferences and supervised 35 MSc and PhD graduate students.
In 2004, Prof Alport was given special leave from the university to form a spin-off company, Advanced Imaging Technologies (Pty) Ltd., where his current projects have involved the development of a seismic analyser and an X-ray imaging system to monitor the internal structure of fabric and steel cord conveyor belts.
William Rae qualified as a registered medical physicist in late 1987 and practiced in a range of fields including radiotherapy physics, nuclear medicine physics and diagnostic imaging physics. Currently, employed at Universitas Hospital, Bloemfontein, and responsible for MRI physics. Current research interests include mammography image analysis and diffusion MRI.