BJR
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
First published online July 9, 2007
British Journal of Radiology (2007) 80, 648-656
© 2007 British Institute of Radiology
doi: 10.1259/bjr/30415751
This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF)
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Karahaliou, A
Right arrow Articles by Costaridou, L
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Karahaliou, A
Right arrow Articles by Costaridou, L

Full paper

Texture analysis of tissue surrounding microcalcifications on mammograms for breast cancer diagnosis

A Karahaliou, MSC 1 S Skiadopoulos, PHD 1 I Boniatis, MSC 1 P Sakellaropoulos, PHD 1 E Likaki, MD 2 G Panayiotakis, PHD 1 and L Costaridou, PHD 1

Departments of 1 Medical Physics and 2 Department of Radiology, School of Medicine, University of Patras, 265 00 Patras, Greece

Correspondence: Dr Lena Costaridou, Department of Medical Physics, School of Medicine, University of Patras, Patras 26 500, Greece. E-mail: costarid{at}upatras.gr


   Abstract
Top
 Abstract
Introduction
 Methods and materials
 Results
 Discussion
 References
 
Diagnosis of microcalcifications (MCs) is challenged by the presence of dense breast parenchyma, resulting in low specificity values and thus in unnecessary biopsies. The current study investigates whether texture properties of the tissue surrounding MCs can contribute to breast cancer diagnosis. A case sample of 100 biopsy-proved MC clusters (46 benign, 54 malignant) from 85 dense mammographic images, included in the Digital Database for Screening Mammography, was analysed. Regions of interest (ROIs) containing the MCs were pre-processed using a wavelet-based contrast enhancement method, followed by local thresholding to segment MCs; the segmented MCs were excluded from original image ROIs, and the remaining area (surrounding tissue) was subjected to texture analysis. Four categories of textural features (first order statistics, co-occurrence matrices features, run length matrices features and Laws' texture energy measures) were extracted from the surrounding tissue. The ability of each feature category in discriminating malignant from benign tissue was investigated using a k-nearest neighbour (kNN) classifier. An additional classification scheme was performed by combining classification outputs of three textural feature categories (the most discriminating ones) with a majority voting rule. Receiver operating characteristic (ROC) analysis was conducted for classifier performance evaluation of the individual textural feature categories and of the combined classification scheme. The best performance was achieved by the combined classification scheme yielding an area under the ROC curve (Az) of 0.96 (sensitivity 94.4%, specificity 80.0%). Texture analysis of tissue surrounding MCs shows promising results in computer-aided diagnosis of breast cancer and may contribute to the reduction of unnecessary biopsies.


   Introduction
Top
 Abstract
 Introduction
Methods and materials
 Results
 Discussion
 References
 
Mammography is currently the most effective imaging modality for breast cancer screening. However, 10–30% of breast cancers may be missed at mammography [1], while its specificity in discriminating between malignant and benign lesions is low, resulting in an increased number of benign (unnecessary) biopsies [2]. Particularly in the case of dense breast parenchyma, the diagnostic accuracy of both screen–film [3–5] and digital mammography [6] is limited.

Microcalcification (MC) clusters are considered a strong indicator of malignancy, and they appear in 30–50% of the mammographically diagnosed cases [7]. Computer-aided detection (CADe) systems for MCs have reported high performance, while the automated interpretation of MCs (computer-aided diagnosis, CADx) remains a challenging task [8, 9]. Difficulty in MC cluster interpretation is mainly due to their fuzzy nature and low contrast (difficulty in distinguishing MCs from their surroundings) [10].

A lot of research has focused on the development of algorithms for the automated classification of MCs. These algorithms are based either on morphology and distribution features of MCs extracted by radiologists [11–13] or on computer-extracted image features [14–27]. Two main categories of computer-extracted image features are used. The first category accounts for morphology/shape features of individual MCs or of MC clusters [14–25], while the second category corresponds to texture features extracted from regions of interest (ROIs) containing the MCs [16, 26–28]. While a review of the proposed CADx schemes can be found elsewhere [8, 10, 29], in the following paragraph, representative studies in terms of features used are provided.

Shen et al [14] developed shape features (compactness, moments and Fourier descriptors) of individual MCs, achieving 100% overall accuracy in the classification of 143 individual MCs. Jiang et al [15] used eight features of MC clusters in a neural network classifier, and achieved an area under the receiver operating characteristic (ROC) curve Az of 0.92 in a data set of 53 patients. Veldkamp et al [18] used cluster distribution, shape and location features for classification of MCs. A patient-based classification was performed by combining information from both views (mediolateral oblique (MLO) and craniocaudal (CC)), achieving an Az value of 0.83. Kallergi [23] used morphological features of individual MCs and MC clusters; when age was incorporated in his classification scheme, a high performance was achieved (Az of 0.98). Chan et al [16] developed morphological features of MCs as well as texture features (co-occurrence matrices based) extracted from ROIs containing the MCs; the combined morphological and texture features achieved an Az of 0.89, which increased to 0.93 when averaging discriminant scores from all views of the same cluster. Dhawan et al [26], following a texture analysis approach, used co-occurrence matrices and wavelet features extracted from ROIs containing the MCs and obtained an Az of 0.86 for the classification of 191 "difficult to diagnose" cases. Soltanian-Zadeh et al [27] compared the performance of four feature sets (co-occurrence matrices based, shape, wavelet and multiwavelet features); the multiwavelet features outperformed the other three feature sets, achieving an Az of 0.89.

The performance of the above CADx schemes is differentiated with respect to the features investigated, the classifiers used and the data sets analysed. The success of the morphological features-based schemes strongly depends on the robustness of the MC segmentation algorithms [8, 30, 31]. Specifically, in the case of dense breast parenchyma abutting the MCs, classification is a challenging task due to difficulty in the segmentation process.

The texture analysis approach seems to overcome this limitation as no segmentation stage is required. The rationale of using texture features is based on capturing changes in the texture of the tissue surrounding MCs. Most texture-based classification studies include MCs in the regions to be analysed further; however, this rationale is expected to introduce bias as the MC, a tiny deposit of calcium in breast tissue, can be neither malignant nor benign. The tissue surrounding or underlying the MC can be characterized as malignant or benign. This tissue is also the one subjected to pathoanatomical and immunochemistry analysis to derive a benign or malignant outcome.

To the authors' knowledge, there is only one study focused on texture analysis of the tissue surrounding MCs for breast cancer diagnosis [32]. This study used a data set of 54 digitally acquired images during stereotactic biopsy. The extracted textural features were based on co-occurrence matrices and fractal geometry, and classification was performed with linear and logistic discriminant analysis. They achieved a sensitivity of 89% with a specificity of 83%, validating the hypothesis that tissue surrounding MCs can be used for breast cancer diagnosis.

The current study investigates whether texture properties of the tissue surrounding MCs, as depicted on screening mammograms, can be used for breast cancer diagnosis, thus aiding radiologists in decisions concerning follow up and biopsy. The discriminatory power of four textural feature categories is investigated using a k-nearest neighbour (kNN) classifier. An additional classification scheme is performed by combining classification outputs of the three most discriminating feature categories. Classification performance is evaluated by means of ROC analysis.


   Methods and materials
Top
 Abstract
 Introduction
 Methods and materials
Results
 Discussion
 References
 
The steps of the proposed method are briefly described as follows. ROIs containing the MC clusters are processed using a wavelet-based contrast enhancement method, and local thresholding is performed in order to segment individual MCs. The segmented MCs are excluded from the original image ROIs, and the remaining tissue area (surrounding tissue) is used for texture analysis and classification.

A medical visualization tool, developed in our department [33, 34], was used for the implementation of enhancement and segmentation procedures, while feature extraction and classification algorithms were implemented in Matlab (The MathWorks Inc., Natick, MA).

Case sample
The case sample consists of 85 mammographic images originating from the Digital Database for Screening Mammography (DDSM) [35], digitized with the LUMISYS 200 scanner at 12 bits pixel depth and 50 µm spatial resolution. The selected mammograms contain 100 MC clusters in total (46 benign and 54 malignant, according to database ground truth tables) and correspond to heterogeneously dense and extremely dense breast parenchyma (density 3 and 4 according to the American College of Radiology (ACR) BIRADS lexicon [36]). The DDSM database provides a malignancy assessment for each MC cluster, also according to the ACR BIRADS lexicon. The assessment ratings are encoded into numerical values from 1 to 5 in increasing order of their likelihood of malignancy: 1, negative; 2, benign; 3, probably benign; 4, suspicious abnormality; and 5, highly suggestive of malignancy. Figure 1Go shows the distribution of the case sample with respect to malignancy rating.


Figure 1
View larger version (12K):
[in this window]
[in a new window]

  		Figure 1. Distribution of the case sample with respect to malignancy rating provided in the DDSM: 1, negative; 2, benign; 3, probably benign; 4, suspicious abnormality; and 5, highly suggestive of malignancy.

 
Enhancement
Images were preprocessed using a wavelet-based spatially adaptive method for mammographic contrast enhancement [37, 38]. This method was selected because it has shown high performance in enhancing MCs compared with other methods proposed for mammographic enhancement [39]. The method is based on local modification of multiscale gradient magnitude values provided by the redundant dyadic discrete wavelet transform. Specifically, a denoising process is performed first taking into account local signal in breast area and noise standard deviation estimated in the mammogram background. Contrast enhancement is accomplished by applying a local linear mapping operator on denoised wavelet gradient magnitude values; coefficient mapping is controlled by a local gain limit parameter. The processed image is derived by reconstruction from the modified wavelet coefficients. Pre-processing was performed on rectangular 600 x 600 pixel ROIs containing the MC cluster instead of the whole image to reduce calculation time. Figure 2Go presents a ROI of 600 x 600 pixels containing the MC cluster in original image (a) and the corresponding processed ROI (b).


Figure 2
View larger version (170K):
[in this window]
[in a new window]

  		Figure 2. (a) 600 x 600 pixel region of interest (ROI) containing a microcalcification (MC) cluster in the original mammogram (DDSM: volume cancer_09, case B_3406, RIGHT_CC). (b) Processed ROI with delineated MC cluster area. (c) Surrounding tissue ROI (ST-ROI). (d) Magnified 128 x 128 pixel subregion of ST-ROI.

 
Definition of tissue surrounding microcalcifications
An experienced radiologist defined a ROI manually, delineating the MC cluster area on each enhanced 600 x 600 pixel ROI (Figure 2bGo), and selected a local threshold to segment individual MCs. The segmented MCs were excluded from the original image 600 x 600 pixel ROI, providing the surrounding tissue ROI (ST-ROI), shown in Figure 2cGo.

In some cases, the segmentation procedure resulted in overestimation of MCs, as well as the inclusion of isolated pixels of high grey level value, corresponding to normal dense tissue. These isolated pixels were removed from the labelled MCs applying a size criterion. The use of a more robust segmentation technique was not deemed necessary for the aim of this study, as morphology analysis of individual MCs was not performed.

Texture analysis
Texture analysis was performed in a 128 x 128 pixel subregion of each ST-ROI (Figure 2dGo), positioned to contain the cluster at its centre. For clusters larger than a single ROI, multiple ROIs (up to three ROIs with less than 30% overlap) were used to cover the entire cluster area. The texture feature values extracted from multiple ROIs, covering a large cluster area, were averaged.

In the case sample analysed, the average percentage of pixels, corresponding to MCs, excluded from each ST-ROI subregion (128 x 128 pixels) was 2.5%.

In this study, four categories of textural features were extracted: first order statistics (FOS); grey level co-occurrence matrices (GLCMs) features; grey level run length matrices (GLRLMs) features; and Laws texture energy measures (LTEMs). Prior to feature extraction, each 128 x 128 pixel subregion was stretched to a normalized grey level range of 0–255.

First order statistics
FOS provide different statistical properties (statistical moments) of the intensity histogram of an image [40]. They depend only on individual pixel values and not on the interaction or co-occurrence of neighbouring pixel values. In this study, four first order textural features were calculated: mean, standard deviation, kurtosis and skewness.

Grey level co-occurrence matrices features
The GLCM is a well-established robust statistical tool for extracting second order texture information from images [41, 42]. The GLCM characterizes the spatial distribution of grey levels in an image. Specifically, an element in the GLCM, Pd,{theta}(i,j), represents the probability of occurrence of the pair of grey levels (i,j) separated by a distance d at direction {theta}. In this study, four GLCMs were computed, corresponding to four different directions ({theta} = 0°, 45°, 90° and 135°) and one distance (d = 1 pixel). 13 features were derived from each GLCM: angular second moment, entropy, contrast, local homogeneity, correlation, shade, prominence, variance, sum average, sum entropy, difference entropy, sum variance and difference variance. The mean and range of each feature over the four GLCMs were calculated, comprising a total of 26 GLCM features.

Grey level run length matrices features
The GLRLM provides information about the coarseness of image texture in specified directions [43]. A grey level run is a set of consecutive, collinear pixels (i.e. a pixel structure) in a given direction that have the same grey level value. The length of a run is the number of pixels in a run. Features extracted from GLRLM evaluate the distribution of small (short runs) or large (long runs) organized structures within the image. In this study, four GLRLMs were computed, corresponding to four different directions (0°, 45°, 90° and 135°). Five features were derived from each GLRLM: short runs emphasis (SRE), long runs emphasis (LRE), grey level non-uniformity (GLNU), run length non-uniformity (RLNU) and run percentage (RPERC). The mean and range of each feature over the four GLRLMs were calculated, comprising a total of 10 GLRLM features.

Laws' texture energy measures
Textural features were extracted based on the method proposed by Laws [44]. According to this approach, textural features are extracted from images that had previously been filtered by each of the 25 Laws' masks or kernels. Five one-dimensional operators (L5 = [1 4 6 4 1], E5 = [–1 –2 0 2 1], S5 = [–1 0 2 0 –1], R5 = [1 –4 6 –4 1] and W5 = [–1 2 0 –2 1]) are used for generation of the 25 Laws' masks. Specifically, each mask is generated by convolving a vertical one-dimensional operator with a horizontal one-dimensional operator. The filtered images are characterized as texture energy images (TE images). Averaging the TE images corresponding to symmetrical kernels (such as R5L5 and L5R5), and taking into account that 20 out of 25 kernels are symmetric, 15 TR images were produced (R stands for "rotational invariance"). From each of the 15 TR images, five first order statistics (mean, standard deviation, range, skewness and kurtosis) were computed (i.e. five LTEM subcategories, each one containing 15 features), giving 75 LTEMs in total.

The extracted textural features of the four aforementioned feature categories were normalized to zero mean and unit standard deviation [45] and subsequently used for classification.

The typical computational time required to extract texture features from one ST-ROI subregion was 0.07 s for FOS, 4.65 s for GLCMs, 3.50 s for GLRLMs and 2.95 s for LTEMs, using a Pentium IV processor running at 3 GHz.

Classification of tissue surrounding microcalcifications
A k-nearest neighbour (kNN) classifier was employed for the classification of the tissue surrounding MCs, based on the extracted textural features. kNN makes a class assignment based on the classes of the k training samples nearest to the unknown sample. In this study, the inverse distance-weighted voting was used [46]. In this approach, the contribution of each of the k neighbours is weighted according to its distance from the unknown sample, giving greater weight to closer neighbours. Specifically, the vote of the kth neighbour is defined as:


Formula 001

where dk is the Euclidean distance of the kth neighbour from the unknown sample. The votes of each class are summed, and the unknown sample is assigned to the class with the highest sum of votes. Specifically, the Decision function for classification is given by:


Formula 002

where m is the number of neighbours belonging to class M (malignant), b is the number of neighbours belonging to class B (benign), and m + b = k. In this study, k ranged from 1 up to 7 neighbours with step 1. If Decision is greater than zero, the unknown sample is assigned to class M; otherwise, the unknown sample is assigned to class B.

The discriminating ability of each textural feature category was investigated using all the individual features of each category as inputs to the classifier. For each textural feature category, a best feature set was selected with respect to overall accuracy achieved, employing an exhaustive search procedure [45]. Specifically, combinations of two to six features were investigated, and the combination of the minimum number of features that provided the highest overall accuracy was selected. In the case of LTEMs, the exhaustive search procedure was initially performed for each LTEM subcategory (mean, standard deviation, range, skewness and kurtosis) and, then, among the selected features from the five subcategories. The training and testing of the classifier, for each textural feature category (and each LTEM subcategory), was performed using the leave-one-out methodology [45].

To enhance the classification success rate, an additional classification scheme was performed by combining the classification outputs of the most discriminating feature sets, with a majority voting rule [47]. In this approach, the unknown sample is assigned to the class of the majority of the classification outputs.

The performance of the classifier for each textural feature set and the combined classification scheme was evaluated by means of ROC analysis [48].

ROC analysis
To obtain a ROC curve for classification based on individual textural feature sets, malignancy thresholds (confidence threshold values) have to be set; above the malignancy threshold, a sample is considered malignant and, below the threshold, it is considered benign. The Decision value given in Equation 2 provides a measure of malignancy for each sample; positive values reflect a high likelihood of malignancy, whereas negative values reflect a low likelihood of malignancy (benignity). Thus, we partitioned the range of Decision values over the whole case sample (maximum Decision value (most positive) minus minimum Decision value (most negative)) in 10 values to obtain 10 malignancy thresholds. In this way, 10 raw data points of a ROC curve were derived. When the threshold is set to the maximum Decision value, none of the samples is malignant, causing both sensitivity (vertical axis in ROC representation) and 1–specificity (horizontal axis in ROC representation) to be 0. When the threshold is set to the minimum Decision value, all samples are malignant, causing both sensitivity and 1–specificity to be 1.

To obtain a ROC curve for the combined classification scheme, we defined as the malignancy threshold (confidence threshold value) for each sample the number of malignant classification outputs provided by the three textural feature sets, ranging from 0 up to 3. We changed the threshold from –1 up to 3 with step 1, to derive five raw data points of the ROC curve, according to the procedure introduced by Soltanian-Zadeh et al [27].

In order to provide a baseline reference for the performance of the proposed surrounding tissue texture-based classification approach, a ROC curve was also generated for the DDSM assessment of malignancy. For this purpose, the malignancy ratings were used as malignancy thresholds (confidence threshold values), and five raw data points of the ROC curve were generated.

The ROCKIT program (Metz CE, University of Chicago, IL) was used for the generation of ROC curves. Specifically, a conventional binormal ROC curve was individually fitted to the raw data points of each textural feature set, the combined classification scheme and the DDSM assessment, with a maximum likelihood procedure. Then, the area under the estimated ROC curve (Az), the standard error (SE) as well as the asymmetric 95% confidence interval (CI) were calculated [48, 49]. Differences in Az values were analysed statistically using the area test (z-score). Derived two-tailed values of p<0.05 indicate statistically significant differences between classification schemes.


   Results
Top
 Abstract
 Introduction
 Methods and materials
 Results
Discussion
 References
 
Table 1Go provides the best feature set for each textural feature category, selected with respect to overall classification accuracy achieved by means of an exhaustive search. The corresponding sensitivity and specificity values are also provided.


View this table:
[in this window]
[in a new window]

  		Table 1. Best feature set for each textural feature category and corresponding values of neighbours(k), sensitivity, specificity and overall classification accuracy

 
Mean and skewness were included in the best feature set of the FOS. In the data set analysed, malignant cases had higher mean grey level values compared with the benign ones, which could be justified by the fact that the occurrence of breast cancers is greater in areas of mammographically dense tissue [50]. Both malignant and benign cases had negative skewness values (histogram of grey level values skewed to the left), with malignant cases presenting higher absolute values. The features included in the best feature set of the GLCM features are mean of difference entropy, range of local homogeneity and range of difference variance. In the data set analysed, malignant cases had lower corresponding values than the benign cases. It is not easy to justify this performance, or to identify which specific qualitative image characteristic is represented by each of the selected GLCM features [41]. However, these results indicate that the combination of some GLCM features can more efficiently discriminate malignant from benign tissue than others. Among the GLRLM features, the combination of mean of SRE and mean of LRE provided the higher overall accuracy in classifying malignant and benign tissue. Among LTEMs, the combination of five features yielded the higher overall classification accuracy and provided the best feature set (Table 1Go).

Figure 3Go shows the ROC curves for the best textural feature sets, presented in Table 1Go, while Table 2Go provides the corresponding Az, SE and CI values. The best feature set of the LTEM category demonstrated the highest performance 0.90 ± 0.03 (Az ± SE). The GLCM best feature set provided a sufficient classification performance (0.86 ± 0.04). The FOS best feature set provided a classification performance of 0.78 ± 0.05. The GLRLM best feature set demonstrated the poorest performance (0.46 ± 0.06), corresponding to random classification.


Figure 3
View larger version (19K):
[in this window]
[in a new window]

  		Figure 3. Receiver operating characteristic(ROC) curve for the best feature set of each textural feature category studied (first order statistics (FOS), grey level co-occurrence matrices (GLCMs), grey level run length matrices (GLRLMs) and Laws texture energy measures (LTEMs)).

 

View this table:
[in this window]
[in a new window]

  		Table 2. Performance evaluation of the k-nearest neighbour (kNN) classifier for best textural feature sets, by means of the area under the receiver operating characteristic (ROC) curve (Az) and corresponding standard error (SE) and assymetric 95% confidence interval (CI) values

 
Figure 4Go presents the ROC curves corresponding to the classification scheme that combined the outputs of the three best textural feature sets (FOS, GLCM features and LTEMs), and the DDSM assessment of malignancy. The raw data points used as inputs to the ROCKIT program are also depicted. The combined classification scheme achieved a performance of 0.96 ± 0.02, with a CI of (0.91, 0.99). As observed in the corresponding ROC curve, a sensitivity of 94.4% can be achieved with a specificity of 80.0%. Concerning the DDSM assessment of malignancy, a performance of 0.84 ± 0.05 was achieved, with a CI of (0.72, 0.92). As observed in the corresponding ROC curve, a sensitivity of 94.4% can be achieved with low specificity (20.0%).


Figure 4
View larger version (17K):
[in this window]
[in a new window]

  		Figure 4. Receiver operating characteristic(ROC) curves corresponding to the classification scheme that combined the outputs of the three best textural feature sets (first order statistics (FOS), grey level co-occurrence matrices (GLCMs) and Laws texture energy measures (LTEMs)) and the DDSM assessment of malignancy. ROC curves' raw data points corresponding to the combined scheme ({blacksquare}) and the DDSM assessment ({blacktriangleup}) are also depicted.

 
Table 3Go summarizes the results of the performance comparison among all classification schemes considered (individual textural features sets, the combined classification scheme and the DDSM assessment of malignancy).


View this table:
[in this window]
[in a new window]

  		Table 3. Two-tailed p-values and corresponding differences in Az values (statistically significant or not) among all classification schemes studied (the outperforming classification scheme is provided in parenthesis)

 
Performance comparison among individual textural features sets indicated that the LTEMs performed best and were statistically significant compared with the FOS and GLRLM features (p<0.05). The LTEMs performed better than the GLCM features, although not statistically significantly (p>0.05). For the GLCM features, although performing better than FOS, the difference in corresponding Az values was not statistically significant (p>0.05). The GLRLM features yielded the poorest performance with a statistically significant difference (p<0.0001) in Az value compared with all the classification schemes considered. The combined texture-based classification scheme statistically significantly outperformed (p<0.05) the remaining classification schemes studied. Both GLCM features and LTEMs performed better than DDSM assessment, although not statistically significantly (p>0.05). DDSM assessment had a higher performance than FOS, although not statistically significantly (p>0.05).


   Discussion
Top
 Abstract
 Introduction
 Methods and materials
 Results
 Discussion
References
 
In this study, a texture analysis approach for breast cancer diagnosis is presented. The method is based on the analysis of the tissue surrounding the MC cluster for the prediction of malignancy. This hypothesis has been motivated by the fact that a MC cluster can be neither malignant nor benign. This characterization corresponds to the tissue surrounding and underlying the MC cluster, which is subjected to pathoanatomical analysis. A similar study has been reported previously by Thiele et al [32], analysing the tissue surrounding MCs as depicted on digitally acquired images during stereotactic biopsy procedure.

In the present study, we have analysed the surrounding tissue as depicted on screening mammograms, in order to develop a tool that could aid radiologists in their decisions concerning biopsy and follow-up. Four categories of textural features were investigated, with the LTEMs demonstrating the highest classification performance. An additional classification scheme was performed that combined the classification outputs of the three most discriminating textural feature sets with a majority voting rule. This combined classification scheme achieved the highest Az value (0.96), significantly outperforming classification based on DDSM assessment and classification based on individual textural feature sets. GLCM features provided a sufficient performance, in accordance with reported CADx studies employing similar features extracted from ROIs containing [16, 26–28] or excluding [32] the MCs. While LTEM and GLCM features performed better than the DDSM assessment, comparison did not reveal statistically significant differences (p>0.05). However, the sparsely distributed raw data points used for generating the DDSM ROC curve (Figure 4Go) may have introduced bias (overestimation of Az value).

The feasibility of the proposed texture-based classification scheme was demonstrated on mammograms corresponding to heterogeneously dense and extremely dense breast parenchyma, rendering classification a difficult task. The difficulty of the data set analysed is further reflected by the fact that 80% of the benign cases (37/46) have been assigned a rating of 4 (suspicious abnormality). It is a well-known clinical fact that the presence of dense breast parenchyma degrades the diagnostic performance of radiologists, on account of increased inter- and intraobserver variability in the interpretation of lesions in both screen–film [51] and digital mammography [6], and the diagnostic performance of CAD systems [8, 10].

While a comparison with other texture-based classification studies is not possible due to different classification algorithms, textural features and data sets (MC subtlety, density categories and number of cases) used, the proposed method has shown promising results. The achieved performance suggests that texture analysis of the tissue surrounding the MCs, as depicted on screening mammograms, may contribute to computer-aided diagnosis of breast cancer by reducing the number of benign (unnecessary) biopsies, while maintaining high sensitivity.

Completion of the proposed method should include the investigation of additional classification schemes and textural features, as well as validation over a larger data set. Reinforcement of the hypothesis of the surrounding tissue texture analysis will be accomplished by investigating the correlation between computer-extracted textural features and pathoanatomical findings.


   Acknowledgments
 
This work is supported by the European Social Fund (ESF), Operational Program for Educational and Vocational Training II (EPEAEK II), and particularly the Program PYTHAGORAS I (B.365.011).

Received for publication September 28, 2006. Revision received December 9, 2006. Accepted for publication January 2, 2007.


   References
Top
 Abstract
 Introduction
 Methods and materials
 Results
 Discussion
 References
 

  1. Majid AS, de Paredes ES, Doherty RD, Sharma NR, Salvador X. Missed breast carcinoma: pitfalls and pearls. Radiographics 2003;23:881–95.[Abstract/Free Full Text]
  2. Kopans DB. The positive predictive value of mammography. AJR Am J Roentgenol 1992;158:521–6.[Free Full Text]
  3. Kolb TM, Lichy J, Newhouse JH. Comparison of the performance of screening mammography, physical examination, and breast US and evaluation of factors that influence them: an analysis of 27,825 patient evaluations. Radiology 2002;225:165–75.[Abstract/Free Full Text]
  4. Bird RE, Wallace TW, Yankaskas BC. Analysis of cancers missed at screening mammography. Radiology 1992;184:613–17.[Abstract/Free Full Text]
  5. Van Gils CH, Otten JD, Verbeek AL, Hendriks JH, Holland R. Effect of mammographic breast density on breast cancer screening performance: a study in Nijmegen, the Netherlands. J Epidemiol Commun Health 1998;52:267–71.[Abstract]
  6. Cole EB, Pisano ED, Kistner EO, Muller KE, Brown ME, Feig SA, et al. Diagnostic accuracy of digital mammography in patients with dense breasts who underwent problem-solving mammography: effects of image processing and lesion type. Radiology 2003;226:153–60.[Abstract/Free Full Text]
  7. American Cancer Society. Cancer facts and figures 1998. Atlanta, GA: American Cancer Society, Inc., 1998
  8. Sampat PM, Markey MK, Bovik AC. Computer-aided detection and diagnosis in mammography. In: Bovik AC, editor. Handbook of image and video processing. New York: Academic Press, 2005:1195–217
  9. Markey MK, Lo JY, Floyd CE, Jr. Differences between computer-aided diagnosis of breast masses and that of calcifications. Radiology 2002;223:489–93.[Abstract/Free Full Text]
  10. Cheng HD, Cai X, Chen X, Hu L, Lou X. Computer-aided detection and classification of microcalcifications in mammograms: a survey. Pattern Recognit 2003;36:2967–91.[CrossRef]
  11. Wu Y, Giger ML, Doi K, Vyborny C J, Schmidt RA, Metz CE. Artificial neural networks in mammography: application to decision making in the diagnosis of breast cancer. Radiology 1993;187:81–7.[Abstract/Free Full Text]
  12. Baker JA, Kornguth PJ, Lo JY, Floyd CE. Artificial neural network: improving the quality of breast biopsy recommendations. Radiology 1996;198:131–5.[Abstract/Free Full Text]
  13. Lo JY, Baker JA, Kornguth PJ, Iglehart JD, Floyd CE. Predicting breast cancer invasion with artificial neural networks on the basis of mammographic features. Radiology 1997;203:159–63.[Abstract/Free Full Text]
  14. Shen L, Rangayyan RM, Desautels JEL. Application of shape analysis to mammographic calcifications. IEEE Trans Med Imaging 1994;13:263–74.[Medline]
  15. Jiang Y, Nishikawa RM, Wolverton DE, Metz CE, Giger ML, Schmidt RA, et al. Malignant and benign clustered microcalcifications: automated feature analysis and classification. Radiology 1996;198:671–8.[Abstract/Free Full Text]
  16. Chan HP, Sahiner B, Lam KL, Petrick N, Helvie MA. Goodsitt M, et al. Computerized analysis of mammographic microcalcifications in morphological and texture feature spaces. Med Phys 1998;25:2007–19.[CrossRef][Medline]
  17. Tsujii O, Freedman MT, Mun SK. Classification of microcalcifications in digital mammograms using trend-oriented radial basis function neural network. Pattern Recognit 1999;32:891–903.[CrossRef]
  18. Veldkamp WJH, Karssemeijer N, Otten JDM, Hendriks JHCL. Automated classification of clustered microcalcifications into malignant and benign types. Med Phys 2000;27:2600–8.[CrossRef][Medline]
  19. Markopoulos C, Kouskos E, Koufopoulos K, Kyriakou V, Gogas J. Use of artificial neural networks (computer analysis) in the diagnosis of microcalcifications on mammography. Eur J Radiol 2001;39:60–5.[CrossRef][Medline]
  20. Verma B, Zakos J. A computer-aided diagnosis system for digital mammograms based on fuzzy-neural and feature extraction techniques. IEEE Trans Inform Technol Biomed 2001;5:46–54.[CrossRef]
  21. De Santo M, Molinara M, Tortorella F, Vento M. Automatic classification of clustered microcalcifications by a multiple expert system. Pattern Recognit 2003;36:1467–77.[CrossRef]
  22. Lee SK, Chung P, Chang CI, Lo CS, Lee T, Hsu GC, et al. Classification of clustered microcalcifications using a shape cognition neural network. Neural Netw 2003;16:121–32.[CrossRef][Medline]
  23. Kallergi M. Computer-aided diagnosis of mammographic microcalcification clusters. Med Phys 2004;31:314–26.[CrossRef][Medline]
  24. Wei L, Yang Y, Nishikawa RM, Jiang Y. A study on several machine-learning methods for classification of malignant and benign clustered microcalcifications. IEEE Trans Med Imaging 2005;24:371–80.[CrossRef][Medline]
  25. Papadopoulos A, Fotiadis DI, Likas A. Characterization of clustered microcalcifications in digitized mammograms using neural networks and support vector machines. Artif Intell Med 2005;34:141–50.[CrossRef][Medline]
  26. Dhawan AP, Chitre Y, Kaiser-Bonasso C. Analysis of mammographic microcalcifications using gray-level image structure features. IEEE Trans Med Imaging 1996;15:246–59.[Medline]
  27. Soltanian-Zadeh H, Rafee-Rad F, Pourabdollah-Nejad D. Comparison of multiwavelet, wavelet, Haralick, and shape features for microcalcification classification in mammograms. Pattern Recognit 2004;37:1973–86.[CrossRef]
  28. Kramer D, Aghdasi F. Texture analysis techniques for the classification of microcalcifications in digitized mammograms. In: Proceedings of the 5th IEEE AFRICON Conference Electrotechnical Service for Africa, Cape Town, 1999
  29. Sakka E, Prentza A, Koutsouris D. Classification algorithms for microcalcifications in mammograms (review). Oncol Rep 2006;15(Spec No):1049–56.[Medline]
  30. Veldkamp WJH, Karssemeijer N. Influence of segmentation on classification of microcalcifications in digital mammography. In: Proceedings of the 18th Annual International Conference of the IEEE Engineering in Medicine and Biology Society; 1996 October 31–November 3. Amsterdam, Netherlands: Institute of Electrical and Electronic Engineers, Inc., 1996
  31. Paquerault S, Yarusso LM, Papaioannou J, Jiang Y, Nishikawa RM. Radial gradient-based segmentation of mammographic microcalcifications: observer evaluation and effect on CAD performance. Med Phys 2004;31:2648–57.[CrossRef][Medline]
  32. Thiele DL, Kimme-Smith C, Johnson TD, McCombs M, Bassett LW. Using tissue texture surrounding calcification clusters to predict benign vs malignant outcomes. Med Phys 1996;23:549–55.[CrossRef][Medline]
  33. Sakellaropoulos P, Costaridou L, Panayiotakis G. An image visualization tool in mammography. Med Inform 1999;24:53–73.
  34. Sakellaropoulos P, Costaridou L, Panayiotakis G. Using component technologies for web based wavelet enhanced mammographic image visualization. Med Inform 2000;25:171–81.
  35. Heath M, Bowyer K, Kopans D, Moore R, Kegelmeyer P. The Digital Database for Screening Mammography. In: Proceedings of the 5th International Workshop on Digital Mammography, IWDM; 2000 June 11–14; Toronto, Canada. Wisconsin: Medical Physics Publishing, 2000
  36. American College of Radiology. American College of Radiology breast imaging reporting and data system atlas (BI-RADS Atlas). Reston, VA: American College of Radiology (ACR), 2003
  37. Sakellaropoulos P, Costaridou L, Panayiotakis G. A wavelet-based spatially adaptive method for mammographic contrast enhancement. Phys Med Biol 2003;48:787–803.[CrossRef][Medline]
  38. Costaridou L, Sakellaropoulos P, Skiadopoulos S, Panayiotakis G. Locally adaptive wavelet contrast enhancement. In: Costaridou L, editor. Medical image analysis methods. Boca Raton, FL: Taylor & Francis Group LCC, CRC Press, 2005:225–70
  39. Costaridou L, Skiadopoulos S, Karahaliou A, Sakellaropoulos P, Panayiotakis G. On the lesion specific enhancement hypothesis in mammography. In: Proceedings of 14th International Conference of Medical Physics, ICMP; 2005 September 14–17; Nuremberg, Germany. Berlin, Germany: Fachverlag Schiele & Schon GmbH, 2005
  40. Gonzalez RC, Woods RE, editors. Digital image processing. Upper Saddle River, NJ: Prentice-Hall, Inc., 2002
  41. Haralick RM, Shanmugam K, Dinstein I. Textural features for image classification. IEEE Trans Syst Man Cybern 1973;SMC-3:610–21.[CrossRef]
  42. Walker RF, Jackway P, Longstaff ID. Improving co-occurrence matrix feature discrimination. In: Proceedings of the 3rd Conference on Digital Image Computing: Techniques and Applications (DICTA '95); 1995 December 6–8; Brisbane, Australia, 1995
  43. Galloway M. Texture analysis using gray level run lengths. Comput Graph Image Proc 1975;4:172–9.
  44. Laws KI. Texture energy measures. In: Proceedings of DARPA Image Understanding Workshop; Los Angeles, 1979
  45. Theodoridis S, Koutroumbas K, editors. Pattern recognition. Amsterdam: Elsevier Academic Press, 2003
  46. Dudani SA. The distance weighted nearest neighbour rule. IEEE Trans Systems Man Cybern 1976;SMC-6:325–7.
  47. Christodoulou CI, Pattichis CS, Pantziaris M, Nicolaides A. Texture-based classification of atherosclerotic carotid plaques. IEEE Trans Med Imaging 2003;22:902–12.[CrossRef][Medline]
  48. Erkel AR, Pattynama PM. Receiver operating characteristic (ROC) analysis: basic principles and applications in radiology. Eur J Radiol 1998;27:88–94.[CrossRef][Medline]
  49. Metz CE. Quantification of failure to demonstrate statistical significance. The usefulness of confidence intervals. Invest Radiol 1993;28:59–63.[CrossRef][Medline]
  50. Ursin G, Hovanessian-Larsen L, Parisky YR, Pike MC, Wu AH. Greatly increased occurrence of breast cancers in areas of mammographically dense tissue. Breast Cancer Res 2005;7:R605–8.[CrossRef][Medline]
  51. Kerlikowske K, Grady D, Barclay J, Frankel SD, Ominsky SH, Sickles EA, et al. Variability and accuracy in mammographic interpretation using the American College of Radiology Breast Imaging Reporting and Data System. J Natl Cancer Inst 1998;90:1801–8.[Abstract/Free Full Text]



This article has been cited by other articles:


Home page
 Br. J. Radiol.Home page
BJR review of the year -- 2007
Br. J. Radiol., April 1, 2008; 81(964): 265 - 269.
[Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF)
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Karahaliou, A
Right arrow Articles by Costaridou, L
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Karahaliou, A
Right arrow Articles by Costaridou, L

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
BJR DMFR IMAGING  ALL BIR JOURNALS