Elsevier

Medical Image Analysis

Volume 6, Issue 3, September 2002, Pages 267-273
Medical Image Analysis

A CAD system for the 3D location of lesions in mammograms

https://doi.org/10.1016/S1361-8415(02)00084-1Get rights and content

Abstract

A CAD system for estimating the 3D (three-dimensional) positions of lesions found in two mammographic views is described. The system is an extension of our previous method [Comput. Vis. Image Understand. 83 (2001) 38] which finds corresponding 2D positions in different mammographic views. The method calculates curved epipolar lines by developing a simulation of breast deformation into stereo camera geometry. Using such curved epipolar lines, not only can we determine point correspondences, but can estimate the 3D location of a lesion. In this paper, we first explain the underlying principles and system organisation. The correctness of the 3D positions calculated by the system is examined using a set of breast lesions, which appear both in mammograms and in MRI data. The experimental results demonstrate the clinical promise of the CAD system.

Introduction

Mammography (breast X-ray) is currently by far the best trade-off between specificity/sensitivity and cost for the detection of breast cancer in its early stages. As a result, screening programmes have been established in a number of countries, including the UK, Netherlands, Sweden, and Australia. Currently, screening programmes are being established in France, Germany, Italy, and Japan. There has been a considerable amount of previous work on CAD (computer-aided diagnosis) systems for mammo-graphy (Doi et al., 1995). Most aim at detecting lesions (including tumors) in images. Hardly any yield 3D (three-dimensional) information, for example the 3D position and volume of lesions, information which is important for the ensuing diagnosis and treatment.

Obtaining 3D information about breast lesions from mammograms has not been accorded a great deal of attention, because breast compression (primarily to reduce X-ray dosage), which almost always varies markedly between the cranio-caudal (CC) and mediolateral oblique (MLO) views, involves a complicated relationship between the 2D positions of a point in the two images and its actual 3D position in the uncompressed breast. Although other modalities, most notably MRI, nuclear medicine and ultrasound, can be used to obtain the required 3D information, some of the most important early indicators of cancer, e.g. calcification, can only be observed in mammograms. It turns out that the 3D distribution of such early signs is clinically significant Yam et al., 2001, Veldkamp et al., 2000). Nevertheless, few clinical studies consider how a lesion appears in an X-ray image based on the projective principle Roebuck, 1990, Novak, 1989).

Acquiring two views of the breast, medio-lateral oblique (MLO) and cranio-caudal (CC), greatly improves sensitivity and specificity. If a lesion is seen in both images, then, theoretically, its 3D position may be determined based on the principles of stereo vision. However, as we noted above, breast compression in the CC and MLO differ quite markedly. As a result, the epipolar geometry, that is the determination of the straight line in one of the images that corresponds (i.e. the locus of candidate matches) to a point in the other image, is deformed into a curve. Hence, the correspondence problem for lesions is not at all intuitive and becomes a difficult task. We proposed the first method for the estimation of curved epipolar lines by developing a simulation of breast deformation into stereo camera geometry (Kita et al., 2001). Using such curved epipolar lines, not only can we determine correspondences, but can estimate the 3D location of a lesion within the uncompressed breast. Using this information, together with a quantitative measure obtained from calibrated image brightness, Yam et al. (2001) showed how to match each microcalcification in a cluster between the two views and how to reconstruct the cluster in 3D. However, the correctness and accuracy of the 3D location obtained from the epipolar curves have, to date, not been investigated precisely.

The 3D position of a lesion in a mammogram, as provided by our method, presents the radiologist with novel, clinically significant information that is useful both for integrative diagnosis with other modalities such as MRI and ultrasound and for the planning of less invasive operations. For this reason, we have constructed a pilot CAD system based on the method. In this paper, we describe the system and analyze the errors in the 3D locations of lesions. In the following section, we first explain the underlying principles and the system organisation. The results produced by the system are examined using breast lesions which appear both in mammograms and in MR images. Finally, we discuss the current capabilities of the system and outline future work.

Section snippets

Principle

When a mammogram is performed, the breast is compressed between the film-screen cassette and the compression plate in the direction of the X-ray source: “head to toe” for the CC view and “over the shoulder diagonally to the hip” for the MLO view. Fig. 1 shows an overview of the system. Further detail is given in (Kita et al., 2001b). First, suppose that a point in one image is pointed at by a radiologist. The method calculates the epipolar curve, that is the locus of possible corresponding

Results

It is not always easy, even for radiologists, to determine accurate correspondences between the CC and MLO images. However, radiologists can determine some correspondences with confidence, apparently based largely on the similarity of the intensity patterns in the two images. We have collected nine lesions for which the correspondences between the CC and MLO images were known, and in such a way that the lesion is also observed in an MR volume. We applied the method described in this paper to

Conclusion

In this paper, we have described the current state of development of a CAD system for estimating the 3D positions of breast lesions found in two mammographic views. This is the first CAD system which enables the radiologist to obtain 3D information from a conventional pair of CC–MLO mammograms. Since some lesions are observed only in mammograms, this has considerable clinical significance.

From the experimental results, particularly the comparison with the 3D information provided by MRI images,

Acknowledgements

This research was partially supported by the Specific International Joint Research programme of AIST. The first author thanks Dr. Kazuo Tanie, Dr. Shigeoki Hirai and the members of the Computer Vision group at AIST.

References (9)

  • K. Doi et al.

    Recent progress in development of computer-aided diagnostic (CAD) schemes

    Med. Imag. Tech.

    (1995)
  • R.P. Highnam et al.

    Mammographic Image Processing

    (1999)
  • Y. Kita et al.

    A CAD system for 3D locating of lesions in mammogram

  • Y. Kita et al.

    Correspondence between different view breast X rays using curved epipolar lines

    Comput. Vis. Image Understand.

    (2001)
There are more references available in the full text version of this article.

Cited by (26)

  • Breast lesion co-localisation between X-ray and MR images using finite element modelling

    2013, Medical Image Analysis
    Citation Excerpt :

    The large deformations of the breast tissues between the two imaging modalities can make it difficult to co-localise information in the breast. Efforts to simulate mammographic compression of the breast tissues from the prone gravity loaded shape of the breast include scaling, rotation, and translation (Marti et al., 2004, 2006), geometric transforms (Kita et al., 2002; Behrenbruch et al., 2003, 2004) or volume-preserving affine transformation of the three dimensional breast image (Mertzanidou et al., 2012). Though these methods allow for fast calculations, they cannot reliably reconstruct the localised deformations of the breast tissues from the prone to compressed states.

  • Support vector machine for breast MR image classification

    2012, Computers and Mathematics with Applications
    Citation Excerpt :

    It produces a sequence of multiple spectral images of tissues with a variety of contrasts using three magnetic resonance parameters: spin–lattice (T1), spin–spin (T2), and dual echo–echo proton density (PD) [1]. Much research has been reported in recent years [2–11] and many approaches to biomedical image processing of the breasts have been proposed. In this paper, we present a “Support Vector Machine” (SVM) application to the multi-spectral breast MR image classification.

  • Multiresolution mammogram analysis in multilevel decomposition

    2007, Pattern Recognition Letters
    Citation Excerpt :

    They achieve 90% mean true positive (TP) detection rate at the price of 0.5 false positive (FP) per image. A CAD system designed by Kita et al. (2002) for estimating the 3D positions of lesions found in two mammographic views is described. The method calculates curved epipolar lines by developing a simulation of breast deformation into stereo camera geometry.

View all citing articles on Scopus

An earlier version of this paper was published as (Kita et al., 2001a).

View full text