Detection and classification of lobular and DCIS (small cell) microcalcifications in digital mammograms

https://doi.org/10.1016/S0167-8655(00)00083-0Get rights and content

Abstract

Microcalcifications are detected by fitting a model to every location in the mammogram. Model parameters yielding the best fit are used as features for detection and classification. The fraction of true positive (tp) detection is 60% with 1.23 false detections per cm2. The rate of correct classification is 69%.

Introduction

Breast cancer is a major cause of death for women. Many countries recommend regular screening for women of certain age groups (usually 50–70 yr of age) and those belonging to other high risk groups. Although these programs are effective in reducing the mortality rate (Andersson, 1998, Thurfjell and Lindgren, 1994), 10–30% of cancers which could have been detected are missed (Bird et al., 1992, Duncan et al., 1998, Holland et al., 1982) (the wide range of these figures stems from the difficulty in defining what is meant by a missed cancer). In addition, a high percentage of patients called back at screening turn out not to have cancer (about 2/3 at BreastScreenSA which oversees screening in SA, Australia). In an effort to improve the accuracy of screening programs, many groups have investigated the possibility of using computers to assist radiologists in reading screening mammograms (Bottema and Slavotinek, 1999).

Clustered microcalcifications are one sign of cancer which has been the target of many studies on computer assisted screening. A linear filter, called the boxrim filter, was introduced in 1987 by a group at the University of Chicago (Chan et al., 1987) to form the basis of a detection scheme which was extended in subsequent papers (Chan et al., 1988, Ema et al., 1995, Nishikawa et al., 1993). Similar approaches using Gaussian filters (Zheng et al., 1995) and splines (Maitournam et al., 1998) instead of the boxrim filter have been described as well as methods using wavelets (Qian et al., 1995, Strickland and Hahn, 1996, Zhang et al., 1998), fractals (Lefebvre and Benali, 1995, Li et al., 1997), and neural networks (Chan et al., 1995, Zhang et al., 1996, Zheng et al., 1996).

The detection of calcification is important, but since a large percentage of normal mammograms also show some calcification, detection by itself is only of limited value. Classifying calcification as being associated with a benign or malignant process requires extracting features of diagnostic importance from the image. Examples of features reported in the literature for this purpose include the number of calcifications in the cluster, mean area of calcifications, variation in computed density, variation in volume, shape irregularity, circularity, statistical properties of the intensity surface, properties of co-occurrence matrices, statistical properties of the distribution of calcifications within the cluster, and fractal dimension of the surrounding tissue (Dhawan et al., 1998, Jiang et al., 1996, Thiele et al., 1996, Wu et al., 1995). Many of these features (but not all) are ones that radiologists use in assessing mammograms.

Although several studies report high detection and classification rates (Jiang et al., 1996, Wu et al., 1995, Zhang et al., 1998), there are some prevailing problems.

  • 1.

    Most calcifications are easy to detect both visually and via a computer algorithm. If a detection scheme finds 90% of all clusters there is a danger that it has found roughly the same 90% that the radiologist discovers without assistance. Using a random collection of mammograms showing calcification for training and testing algorithms probably biases the routine toward finding clusters which offer little additional diagnostic value.

  • 2.

    Detection algorithms reported in the literature target all classes of calcifications at once, despite the fact that there are huge varieties in shape, size, and contrast.

  • 3.

    Although the emphasis has rightly been focused on detecting clusters of calcifications, once a cluster is detected, classification depends on measuring features from a diverse collection of individual calcifications within the cluster. So while a few prominent calcifications might suffice to detect the cluster, improved classification may well depend on detecting and analyzing subtle examples.

  • 4.

    Many of the features listed in the previous paragraph for use in classification are not meaningful if the calcification is represented by only a few pixels.

Our group seeks to mitigate these problems by developing different detecting schemes for different classes of microcalcifications. In this way, detection is (hopefully) improved by being more focused and classification is boosted by information supplied by the various schemes. In this paper, attention is restricted to the detection of small, more or less spherical calcifications. This is done by constructing a model for such calcifications based on the projection of a ball of constant density. By varying the radius and density of the model, a best fit is found. The values of radius and density which provide the best fit are used to decide if the candidate is a true calcification or not.

Section snippets

Calcification model

The measured intensity of an X-ray beam may be modeled asI(x)=I0e−∫Lf(z)dz,where I0 is the source intensity, f(z) the X-ray attenuation at a point z in space, x the projection of z onto the image plane, and L is the path of the X-ray beam from the source to the detector (film). Scattering is ignored and X-ray beams are assumed to be parallel. Our model consists of a ball of constant X-ray attenuation, μ, embedded in tissue. We assume that the contribution to the image intensity surface of the

Results

Table 1 summarizes the detection rates for the algorithm described here and for a previous version (Bottema and Slavotinek, 1998). For training and testing images, the percentage of true positive (tp) detections is given and the false positive (fp) rate is given in terms of the number of false positive detections per cm2. The algorithm was also applied to a number of “normal” images, meaning cases where no indications of cancer were found at screening and no cancer developed within 3 yr.

The

Discussion

For reasons stated in the introduction, attention was restricted to clusters of calcifications which are difficult to detect. In particular, only calcifications of radius less than 0.2 mm were considered. In comparing the performance with our previous work, a moderate reduction in the detection rate was well compensated for by a substantial reduction in the fp detection rate.

The performance of the detecting algorithm is difficult to assess independently of the performance of a full scheme for

References (25)

  • K.A. Duncan et al.

    Incident round cancers: what lessons can we learn?

    Clin. Radiol.

    (1998)
  • B. Zheng et al.

    Computer-aided detection of clustered microcalcifications in digitized mammograms

    Acad. Radiol.

    (1995)
  • Andersson, I., 1998. Breast cancer screening results. In: Karssemeijer, N., Thijssen, M., Hendriks, J., van Erning, L....
  • R.E. Bird et al.

    Analysis of cancers missed at screening mammography

    Radiology

    (1992)
  • Bottema, M.J., Slavotinek, J.P., 1998. Detection of subtle microcalcification in digital mammograms. In: Karssemeijer,...
  • Bottema, M.J., Slavotinek, J.P., 1999. Computer aided screening mammography. In: Pham, B., Braun, M., Maeder, A.J.,...
  • H.-P. Chan et al.

    Image feature analysis and computer-aided diagnosis in digital radiography. I. Automated detection of microcalcifications in mammography

    Med. Phys.

    (1987)
  • H.-P. Chan et al.

    Computer-aided detection of microcalcifications in mammograms methodology and preliminary clinical study

    Invest. Radiol.

    (1988)
  • H.-P. Chan et al.

    Computer-aided detection of mammographic microcalcifications: Pattern recognition with an artificial neural network

    Med. Phys.

    (1995)
  • A.P. Dhawan et al.

    Analysis of mammographic microcalcifications using gray-level image structure features

    IEEE Trans. Med. Imaging

    (1996)
  • T. Ema et al.

    Image feature analysis and computer-aided diagnosis in mammography: reduction of false-positive clustered microcalcifications using local edge-gradient analysis

    Med. Phys.

    (1995)
  • R. Holland et al.

    So-called interval cancer of the breast: pathologic and radiologic analysis of sixty-four cases

    Cancer

    (1982)
  • Cited by (0)

    View full text