ArticlePDF Available

Registration and matching of temporal mammograms for detecting abnormalities



Our aim is to ,establish accurate correspondences between ,temporal ,mammograms ,in order ,to improve,the detection of tumours. We propose,a method for detecting abnormalities after aligning a pair of mammograms,using thin-plate spline interpolation based on correspondingpoints on the breast boundary. We also suggest a method for the automatic detection of landmarks,within the breast tissue. We show how,using suitable boundary ,points yields reasonable results whilst finding ,internal landmarks ,is difficult and using points on the pectoral muscle unreliable.
Registration and matching of temporal mammograms for
detecting abnormalities
K. Marias
, J.M. Brady
, R.P. Highnam
, S. Parbhoo
, A.M. Seifalian
Robotics Research Group, Engineering Science, Oxford University, Parks Road, Oxford, OX1 3PJ,
Department of Surgery, Royal Free and University College Medical School, UCL, London NW3 2QG
Abstract. Our aim is to establish accurate correspondences between temporal mammograms in order to
improve the detection of tumours. We propose a method for detecting abnormalities after aligning a pair of
mammograms using thin-plate spline interpolation based on corresponding points on the breast boundary. We
also suggest a method for the automatic detection of landmarks within the breast tissue. We show how using
suitable boundary points yields reasonable results whilst finding internal landmarks is difficult and using
points on the pectoral muscle unreliable.
1 Introduction
The reliable diagnosis of abnormalities from a single mammogram is very difficult [1], and so it is increasingly
the case that a pair of mammograms is compared. This is particularly important in view of the fact that second
round screening results are frequently becoming available. If mammograms are available from an earlier date,
radiologists will compare them to the current ones in order to detect abnormalities on the basis of significant
differences. This process can be especially important for those women at high risk of developing breast cancer
since baseline mammograms are often taken. Our aim is to automate the comparison of temporal mammograms.
Comparison of a mammogram with a previous one is made more difficult by differences in compression of the
breast at the two time instances, and the likely differences in the imaging conditions. The effects of these
differences, which from now will be described as temporal changes, consist of a non-rigid transform in the
image plane due to difference in compression, and a non-rigid transformation in the intensity of the images due
to differences in the imaging conditions. In addition, normal expected changes during female adult life (e.g.
weight gain or loss, involution) as well as affects of HRT (hormone replacement therapy) and Tamoxifen,
usually introduce more changes in the architecture of a mammogram. The only reliable method for normalising
(and thus compensating for the variability in imaging conditions) digitised mammograms, is based on the h
representation of interesting (non-fatty) tissue, developed by Highnam and Brady [1]. On the other hand,
variability in breast compression remains a significant problem for registration of temporal mammograms.
However, Highnam and Brady [1] have shown that the non-rigid part of the transformation induced by
differential compression can be characterised as a divergence of the parenchymal tissue plus a change in some
curvilinear structures (vessels). Tumours, and other high-density tissues can be an exception to that smooth
motion under differential compression as they can move more and in a non-regular way.
In this paper we start by considering a mammogram registration technique based only upon boundary points to
undo the effects of different imaging conditions and compressions. We then consider how to improve that
matching by using internal points before presenting results in the form of difference images and a discussion.
2 Methods used
2.1 Registration, unwarping and difference analysis
The basic method that we use for image registration is thin-plate spline interpolation to align temporal
mammograms using a set of landmarks from the breast boundary. The utilisation of thin-plate spline
interpolation as a way of recovering deformations in medical images was introduced by Bookstein [2]. The
calculated interpolating function f
y) for the vertical or the horizontal direction is smooth and deforms the
image in such a way that the bending energy is minimised, while the landmarks are matched. The bending
energy is given by (1):
Author for correspondence,
The basis function U, is a fundamental solution of the biharmonic equation Δ
U = 0. This method is efficient at
recovering local deformations; but special care is needed in the selection of the landmarks. Bowyer and Sallam
[3], and Karssemeijer and te Brake [4] have presented work on registration of bilateral mammograms using thin-
plate splines for the automatic detection of abnormalities through visual examination of the difference images
after registration. Breast asymmetry is the main characteristic that can be assessed with that method. However,
both techniques use points evenly spaced along the breast boundary, along the chest wall, and include points
from the boundary between the breast and the pectoral muscle. In addition, the method for the automatic
detection of corresponding points presented in [3] cannot be robust, since in most cases it is impossible to
establish point-to-point correspondence in a bilateral pair of mammograms. We contend that special care is
needed in order to decide which points should be used in mammogram registration and this issue is addressed in
the discussion.
In 7 series of temporal mammograms that we have experimented with, we have observed that a good initial
alignment can be achieved using only five characteristic points along the breast boundary. These points are the
nipple, the two points at either end of the breast boundary near the chest wall, and the two points located
centrally along the breast boundary between the nipple and the two ends. The intensity histogram of a
mammogram is bimodal, and the breast boundary can be approximated by automatically selecting a threshold in
the valley of the histogram. Once we’ve calculated the interpolating function, we can produce unwarped images
by forcing every point (x, y) in a mammogram to take the intensity value of the point where the interpolating
function maps the (x, y) point of the previous mammogram. Bilinear interpolation can be used to calculate
intensities outside the pixel grid [3]. After image unwarping, we can produce difference images and search for
regions of large intensity differences. These regions can be either new growths (e.g. a cancer), changes due to
involution, or they can be due to local inaccuracy in registration. The registration can be improved if we are able
to automatically detect landmarks inside the breast tissue. In the next section, we propose a method for the
automatic detection of internal landmarks.
2.2 Proposed method for matching internal points
Matching a pair of mammograms of the same patient acquired at different times, is a process in which we look
for image primitives (points, regions or contours) that are present in both images. In other words, we use a
similarity measure to match important features across the two images. Kok-Wiles, Brady, and Highnam [5],
developed an algorithm for matching regions defined in temporal or bilateral mammograms. Their algorithm
uses heuristics to model the non-rigid transformation and matches iso-intensity contours corresponding to salient
regions of the image. This is done after removing the curvilinear structures of the breast (ducts and vessels) that
change significantly under the different compressions. The technique can be applied, with advantage, to h
images. On the other hand Vujovic and Brzakovic [6], select intersections of image structures such as ducts and
blood vessels as potential control points which subsequently they match between the temporal mammograms.
Although their algorithm is successful in detecting such points even in predominantly radio-opaque
mammograms, we have observed that the vessel and duct structure appearance is very susceptible to temporal
changes, because of the difference in compression [1, Chapter 9].
Our new approach to detecting internal landmarks is based on their geometrical and textural characteristics. Like
[5], we assume that even though tissue structures are deformed differently between temporal images there must
be points and regions (which may be part of the reason we perceive architectural similarity) with similar
geometrical and textural characteristics. For that reason we calculate 1-dimentional local maximum and
minimum points and use them as potential control points. A point is classified as maximum (minimum) if the
intensity of n previous pixels increases (decreases) monotonically (in the horizontal or vertical direction) and
decreases (increases) for the next n pixels. In that way we significantly reduce the number of points to be
matched to a very significant, in terms of information, subset of the 1-dimentional maximum and minimum
points. For each of these points in the first image (x
, y
), we search for candidate matches in a fixed window
around (x
, y
) in the second image and the closest point in feature space is assigned as the match of (x
, y
). The
features consist of texture measures based on co-occurrence matrix (e.g. entropy, homogeneity and contrast). We
also take into consideration the difference in the distance of a candidate pair and the mean distance of previous
matches, thus imposing a smoothness constraint.
3 Results
In figure 1 (a) and (b) we see an example of an interval cancer. Figure 1 (c) is the second image unwarped in the
co-ordinates of the first, while 1 (d) shows the difference image after registration. The difference image can be
used for the automatic detection and segmentation of the cancer using a one-tailed t-test on the pixel intensities
inside the pre-calculated breast boundary (figure 1 (e)). In that way we avoid segmenting the bright line of
misregistration along the boundary. Since the segmented cancer has been transformed to the co-ordinates of the
first mammogram we can estimate the region of future development of cancer in the first, “normal” mammogram
by superimposing the boundary of the transformed cancer region (figure 1 (f)). Figure 2, illustrates that
registration using the pectoral muscle can be severely in error, especially if landmarks from the inside of the
breast have not been included in the interpolation scheme. Since the pectoral muscle is currently being used in
mammogram registration ([3], [4]), we return to this issue in the discussion. Finally, in figures 3 (c) and (d), we
show calculated matches (maximum points) superimposed on the images shown in figure 3 (a) and (b), which
are “similar” parts of a temporal mammogram pair. The brighter points belong to the first image. We hope to
improve the results further by including internal points in the registration process, but we haven't yet included
them since the registration method we used so far does not allow any errors in the localisation of the landmarks.
(b) (c)
(e) (f)
Figure 1. (a): Before cancer developed, (b): A clear cancer has developed, (c): Mammogram (b) unwarped in the
co-ordinates of mammogram (a), (d): Difference image after registration, (e): Cancer segmented from (d), (f):
The transformed boundary of the cancer superimposed on the “normal” mammogram (a), indicates the region of
the future development of cancer.
4 Discussion and conclusions
Special care should be taken when selecting landmarks for the alignment of mammograms. Points in the line of
the chest wall (first column of the image) can be used with some certainty in cranio-caudal view mammograms,
but this might not be the case for medio-lateral mammograms where we usually observe different chest wall
portions in the first column of the image. Additionally, since the duct patterns of the breast converge to the
nipple they will not match if the nipple is not used as a landmark. This can affect the adequacy of the method for
asymmetry detection in bilateral mammogram registration. Nevertheless, the position of the nipple is difficult to
calculate automatically, so equally spaced points in the boundary can be used as an alternative, as in [3] and [4].
However, if such a method is to be used in clinical practice, the trade-off between automatic registration with the
nipple misaligned and semi-automatic with the nipple aligned, should be decided by the radiologist. Moreover,
as is illustrated in figure 2, trying to align the pectoral muscle without including landmarks from the inside of the
breast (like in [4]), leads to an increase in the error of registration. In fact, the difference image 2 (f) reveals that
although the pectoral muscles are aligned, the fibroadenoma present in both images is further away than in the
unregistered images. We have observed from differential compression mammography [1], that the pectoral
muscle seems to move independently of the breast tissue under different compression and this might be a
possible explanation. For that reason, further investigation is necessary to clarify if the pectoral muscle should be
taken into consideration in mammogram registration.
On the other hand, matching points inside the breast is difficult due to temporal changes and depends upon the
extent to which the architecture (or topology of the surface) is preserved. This method provides many good
matches, but insignificant and multiple matches are unavoidable. Thus the next step should be to locate more
reliably “topographic” features of the intensity surface for calculating internal landmarks more robustly and use
a registration method that takes into consideration the localisation uncertainty of internal landmarks.
In conclusion, previous work on matching temporal mammograms [5], has concentrated on defining regions
inside the mammograms as a basis for matching. That work ignores the breast boundary. Here, we demonstrate
the considerable redundancy that is available in matching breast images by showing that those good results can
be obtained using only the breast boundary. Future work will combine the two sources of information in order to
develop a robust matching technique. In addition, our work will be focused on combining normalisation (using
the h
representation) and registration of temporal mammograms. Since the imaging conditions can vary over
time, even if a temporal pair is very accurately aligned, the intrinsic differences in the intensity of corresponding
points remain, diminishing the value of the difference image. We therefore believe that the h
representation can
improve the quality of the resulting difference image.
(b) (c) (d) (e) (f)
Figure 2. (a),(b): A temporal pair of mammograms with a fibroadenoma present, (c): Difference image before
registration, (d): After registration using five points in the boundary the difference image indicates that we have
compensated for some (but not all) of the tissue deformations, (e): By manually adding two internal landmarks,
registration is improved, (f): If we use two points from the line of the pectoral muscle instead, although we align
the pectoral muscle, the mass is further displaced than in the unregistered difference image 2(c).
(b) (c) (d)
Figure 3. Matches in (c), (d) are superimposed on the original temporal images (a) and (b)
K. Marias would like to thank Cancerkin, Royal Free Hospital for supporting his PhD.
1. R.P Hignam and J.M. Brady. Mammographic image processing. Kluwer Academic Press, 1999
2. F. Bookstein, “Principal warps: Thin-plate splines and the decomposition of deformations”, IEE Transactions Pattern
Analysis and Machine Intelligence, 11, pp. 567-585, 1989.
3. M. Sallam and K. Bowyer, “Registration and difference analysis of corresponding mammogram images”, Medical
Image Analysis, 1, No. 1, pp 73-91, 1996.
4. N. Karssemeijer and G.T. Brake, “Combining single view features and asymmetry for detection of mass lesions”, In
Proceedings of Digital Mammography 98, pp. 95-102, Kluwer Academic Publisher, Nijmegen, 1998.
5. Siew-Li Kok-Wiles, M. Brady and R.Hignam, “Comparing mammogram pairs for the detection of lesions”, In
Proceedings of Digital Mammography 98, pp. 103-110, Kluwer Academic Publisher, Nijmegen 1998.
6. N. Vujovic and D. Brzakovic, “Establishing the correspondance between control points in pairs of mammographic
images”, IEE Transactions on Medical Imaging, 6, No. 10, pp. 1388-1398, October 1997.
... @BULLET Computation of the h int mammogram representation [2]. Due to different imaging conditions (e.g. ...
... scattering), result in poor intensity correspondence in temporal mammograms. The h int representation estimates the thickness of non-fatty tissue " above " each pixel, by using a physical model of the image formation [2]. This way, we have a quantitative representation of the breast which effectively normalises temporal mammogram pairs. ...
... However, it still needs to be clinically meaningful and in our future work we aim to validate that assumption. In an experiment to validate the intensity preserving alignment method presented here, we used differential compression data [2] (the patient is kept in the same position, and the breast is imaged at the same time, in two different compression levels). This could be considered as a ground truth, since (unlike temporal mammograms where the breast changes between acquisitions) the resulting h int images (after registration and intensity correction) should be exactly the same.Table 1 ...
Full-text available
In this paper, we explore the idea of quantifying local breast-tissue density changes. Breast tissue density has been correlated to breast cancer incidence in numerous studies which have shown a statistical relationship between glandular density and the occurrence of cancer. In particular, postmenopausal women who take HRT run an increased risk of developing cancer due to the "regeneration" of fibroglandular tissue that is often induced by the exogenous hormones. In this paper, we present a method that combines mammogram normalisation and volume-preserving registration, and which can be the starting point for temporal-local breast tissue quantification.
... ists of two separate registrations (figure 2) in order to compensate not just for breast-edge shape deformation under compression, but also the more complex internal deformation. The first " partial registration " utilises a curvature measure to correlate boundary points along the film edge of the mammograms with the edge of the volume projection (Marias et. al., 1999). This curvature measure depends on a good segmentation of the breast edge, which is achieved by an intensity-based search, mathematical morphology for smoothing and spline fitting to produce a smoothed edge profile (Marias et. al. – Tech. Rept., 1999). The effect of this partial registration is to deform the extremities of the X-ray im ...
... sure to correlate boundary points along the film edge of the mammograms with the edge of the volume projection (Marias et. al., 1999). This curvature measure depends on a good segmentation of the breast edge, which is achieved by an intensity-based search, mathematical morphology for smoothing and spline fitting to produce a smoothed edge profile (Marias et. al. – Tech. Rept., 1999). The effect of this partial registration is to deform the extremities of the X-ray image to the boundary shape of the MRI contrast projection. Detection of internal landmarks using wavelet scale-space Detection of edge points via curvature Partial registration using only breast edge curvature Breast edge detection (search & morphology) ...
Full-text available
X-ray mammography and contrast-enhanced Magnetic Resonance Imaging (MRI) provide complementary pathological information for breast cancer diagnosis and surgical/treatment planning. MRI provides good tissue specificity (whether or not the breast is compressed) via a pharmacokinetic model of contrast enhancement, whilst X-ray mammography has spatial resolution and image formation physics capable of providing information about curvilinear structures, tumour spiculation and microcalcifications. In this paper we present a method for data fusion between MRI and X-ray for the purposes of surgical planning. The objective is to register cranio-caudal (CC) and medio-lateral (ML - oblique) mammograms to equivalent geometric projections of a contrast-enhanced MRI (Gadolinium-DTPA) volume to enable feature correspondence to be made between modalities. However, rather than use a simple intensity-based projection, we exploit voxel gadolinium uptake (Hoffman et. al. 1995) to produce a pharmacokinetic projection of the tissues. The result is effectively an uncompressed "pseudo X-ray" that features functional, rather than structural information about breast tissues. An example of such a projection is
... The second issue, which is perhaps of more clinical interest, is registering the mammograms to the MRI projection. We have developed a technique based on curvature measures and wavelet landmark detection that utilises thin-plate splines to accurately register images. This technique has been reliably applied to register temporal mammograms (Marias et. al., 1999 and to perform 2D/3D data fusion between mammography and contrastenhanced MRI. We refer to for further detail of the process. The result of registration is a complex deformation of the mammograms to the spatial reference frame of the MRI volume (via the projection). More specifically, the registration process has the effect of "uncompre ...
... Further work will focus on clinical validation with a larger data set (approximately 35 cases) as well as phantoms and core-biopsy studies. We also hope to use our technique as a data fusion framework for validating the h_int representation developed by Highnam and Brady (1999). ...
... As a consequence, there are numerous methods proposed for image registration, the majority of which belong to one of the following categories: @BULLET Feature-based: A set of distinctive features (e.g. salient points([2], [3], [4],[5], [6]), curves, contours([7], [8], [9], [10], [11])) are selected from the Source Image and another one set from the Target Image corresponding to the first one. This can be done either manually or automatically . ...
Full-text available
Image Registration is the task of aligning a pair of images in order to bring them in a common reference frame. It is widely used in several image processing applications in order to align monomodal, multimodal and/or temporal image information. For example in medical imaging registration is often used to align either mono-modal temporal series of images or multi-modal data in order to maximize the diagnostic information and enable physicians to better establish diagnosis and plan the treatment. This paper first deals with the analysis of the well-known Maes' method for Image Registration explaining some of its drawbacks. As a remedy to these problems, a novel Image Registration method is proposed as an extension to Maes' one, which overcomes its drawbacks and improves robustness. Results presented in satellite and medical images indicate that the presented method is more robust comparing to well-known algorithms at the expense of increased computational time.
... In order to accurately quantify changes in HRT image sequences, we first normalise the intensities using the h int representation [2] (discussed in the next section) and then register the images to the same co-ordinate frame. To compensate for differences in compression, Marias et al. [3] demonstrated a registration method based on boundary points to approximate the changes in compression. A recently improved method [4] that takes into consideration internal landmarks calculated using multi-scale analysis is summarised in section 2.2. ...
Full-text available
In this paper we introduce ,a novel ,method ,for assessing mammographic ,(density) changes ,in sequences of patients on HRT. The temporal sequence of mammograms ,is represented using the h int representation of interesting tissue and subsequently registered. The % change inh int(non-fatty tissue) in the registered pair is a good approximation,of the fibroglandular change,in the composition,of the breast tissue. The expert's description of the changes in the HRT sequences of our,data set was used as the,gold standard for evaluating our quantitative measure.
Conference Paper
In this paper, we propose a statistical temporal change detec- tion scheme for early breast cancer detection . Temporal mammographic data have been found useful to detect changes in the breasts of women. Many temporal analysis approaches require temporal image registration. Variations in patient positioning, changes in the �eld of view and nat- ural changes in the breasts over time pose signi�cant challenges. Our proposed scheme, on the other hand, does not depend on image registra- tion. Instead, homogeneous local breast regions are found and tracked over time using a statistical region merging technique. Temporal changes found are then assessed by a rule-based classi�er for abnormality assess- ment. Using a small temporal dataset of 10 women (5 cancerous and 5 normal), the detection rate was found to be 100% with a 0:1 false pos- itive per case (that is, only one false positive was found in the entire dataset of 10 cases). This preliminary results show that the proposed temporal changes detection scheme have a great potential in providing clinical assistance in early breast cancer detection. The results, however, need to be further veri�ed with a larger dataset.
Mammogram registration is a critical step in automatic detection of breast cancer. Much research has been devoted to registering mammograms using either feature-matching or similarity measure. However, a few studies have been done on combining these two methods. In this research, a hybrid mammogram registration method for the early detection of breast cancer is developed by combining feature-based and intensity-based image registration techniques. Besides, internal and external features were used simultaneously during the registration to obtain a global spatial transformation. The experimental results indicates that the similarity between the two mammograms increases significantly after a proper registration using the proposed TPS-registration procedures.
Conference Paper
Full-text available
Increasing use is being made of contrast-enhanced Magnetic Resonance Imaging (Gd-DTPA) for breast cancer assessment since it provides 3D functional information via pharmacokinetic interaction between contrast agent and tumour vascularity, and because it is applicable to women of all ages. Contrast-enhanced MRI (CE-MRI) is complimentary to conventional X- ray mammography since it is a relatively low-resolution functional counterpart of a comparatively high-resolution 2D structural representation. However, despite the additional information provided by MRI, mammography is still an extremely important diagnostic imaging modality, particularly for several common conditions such as ductal carcinoma in-situ (DCIS) where it has been shown that there is a strong correlation between microcalcification clusters and malignancy (1). Pathological indicators such as calcifications and fine spiculations are not visible in CE-MRI and therefore there is clinical and diagnostic value to fusing the high-resolution structural information available from mammography with the functional data acquired from MRI imaging. This paper presents a novel data fusion technique whereby medio-lateral (ML) and cranio-caudal (CC) mammograms (2D data) are registered to 3D contrast- enhanced MRI volumes. We utilise a combination of pharmacokinetic modelling, projection geometry, wavelet-based landmark detection and thin- plate spline non-rigid registration to transform the coordinates of regions of interest (ROIs) from the 2D mammograms to the spatial reference frame of the contrast-enhanced MRI volume.
Conference Paper
Full-text available
A method to correspond linear structures in mammographic images is pre- sented. Our approach is based on automatically establishing correspondence between linear structures which appear in images using robust features such as orientation, width and curvature extracted from those structures. The re- sulting correspondence is used to track linear structures and regions in mam- mographic images taken at different times.
Computer support for early detection of breast cancer requires a proper mimicking of the way radiologists compare mammographic images; by comparing bilateral (images of the left and right breasts) and temporal images. In this paper, one method for bilateral registration and intensity normalization and two methods for difference analysis are described. The bilateral registration is based on anatomical features and assumptions of how the female breast is deformed under compression. The first method for differential analysis is based on the absolute difference between the registered images while the second method is based on statistical differences between properties of corresponding neighborhoods. The methods are tested on images from the MIAS database (on 100 images with 59 abnormalities distributed over four types) and evaluated by FROC-analysis. The performances of the two methods are similar but the statistical method gives better performance at a lower false positive rate and is better in particular for detecting asymmetrical developments.
The decomposition of deformations by principal warps is demonstrated. The method is extended to deal with curving edges between landmarks. This formulation is related to other applications of splines current in computer vision. How they might aid in the extraction of features for analysis, comparison, and diagnosis of biological and medical images is indicated.
Radiologists in breast cancer screening are trained to use comparisons of left and right mammograms to identify suspicious asymmetric densities. Asymmetry is not a very specific sign, as the majority of asymmetric densities are due to normal variation of the parenchymal pattern. However, when an asymmetric density has some of the characteristics that are typical for a malignant mass or appears in a region that should normally be empty, it may be suspicious. Previously, we have developed a method for detecting stellate and circumscribed masses in mammograms from single views [5], [11], based on a statistical analysis of line and gradient orientation patterns. In this study we investigated the use of a local measure of asymmetry as an additional feature, with the aim of improving overall detection performance on a large consecutive sample of screening cases.
An automated technique is proposed for identifying differences between corresponding mammogram images. The technique recovers an approximate deformation between a pair of mammograms based on identifying corresponding features across the two images. The registration process is completed using an unwarping technique for transforming one image into the coordinate system of the other. A difference image is generated using intensity-weighted subtraction in order to identify regions of large difference. Evaluation of the technique is performed using 124 bilateral image pairs which contain a total of 77 abnormalities of different types. The purpose of this paper is to measure the extent to which the mammogram registration technique is able to provide useful information for identifying abnormalities in mammograms.
Conference Paper
An imaging process creates an image that contains information about certain characteristics of the objects in the scene. Inevitably, the information is confounded by changes in the imaging conditions and scene background and it becomes difficult to judge simply by looking at the image whether the objects have changed or whether the imaging conditions or background has changed. In mammography, one is interested in the former not the latter. In order to overcome this problem for image analysis it is necessary to normalise the image and/or the feature measurements. Here, the authors show how mammographic images can be normalised in the form of a representation that makes explicit quantitative measurements of the breast itself and the authors argue that this is the correct representation to use for mammographic image processing. In particular, the authors note that their representation can be generated from conventional and directly-digital mammograms since it represents the breast, not the imaging process. The paper starts with the reasoning behind, and a description of, the h<sub>int</sub> representation which is a quantitative representation of the breast tissue. The next section demonstrates the power of the h<sub>int</sub> representation: it allows one to simulate any projective X-ray examination and to simulate the appearance of anatomical structures within the breast. The authors follow this with a comparison between the h<sub>int</sub> representation and a conventional representation from the point of view of invariance to imaging conditions and surrounding tissue
This paper describes part of a study aimed at developing a computer-based aid for mammogram screening that makes a detailed comparison between mammograms of the same patient acquired at different screenings and detects changes indicative of cancer. The focus is on determining control points in two mammograms; these points are used to put two mammograms into correspondence. The paper details the algorithm for identifying the potential control points and establishing the correspondence between the two sets of control points. The algorithm's performance was evaluated by three observers, one of whom is an experienced radiologist, and found to be adequate
We present an approach to detecting masses in temporal and bilateral (left-right) pairs of mammograms. Masses typically appear in a mammogram as well-defined, relatively bright regions. The converse is not true; by no means all relatively bright, circumscribed regions correspond to masses. Common to our approach to matching temporal and bilateral mammogram pairs is a representation of the nested structure of “salient” bright regions in a mammogram. The representation is described in Sections 2 and 3. In the case of temporal pairs, “salient” regions are extracted independently in two mammograms of the same breast, nominally the same view at two different times. Those that either appear in the later mammogram but not in the earlier one, or which have changed significantly between the two mammograms we seek to highlight. Similarly, in the case of matching bilateral pairs, salient regions that appear in only one of same-view left and right breast mammograms of the same patient, taken at approximately the same time we seek to highlight. A fuller version of this article appears as Chapter 13 of [4].
Combining single view features and asymmetry for detection of mass lesions
  • N Karssemeijer
  • G T Brake
N. Karssemeijer and G.T. Brake, "Combining single view features and asymmetry for detection of mass lesions", In Proceedings of Digital Mammography 98, pp. 95-102, Kluwer Academic Publisher, Nijmegen, 1998.
Comparing mammogram pairs for the detection of lesions
  • Li Siew
  • M Kok-Wiles
  • R Brady
  • Hignam
Siew-Li Kok-Wiles, M. Brady and R.Hignam, "Comparing mammogram pairs for the detection of lesions", In Proceedings of Digital Mammography 98, pp. 103-110, Kluwer Academic Publisher, Nijmegen 1998.