 Research
 Open Access
 Published:
Computeraided diagnosis of breast microcalcifications based on dualtree complex wavelet transform
BioMedical Engineering OnLinevolume 11, Article number: 96 (2012)
Abstract
Background
Digital mammography is the most reliable imaging modality for breast carcinoma diagnosis and breast microcalcifications is regarded as one of the most important signs on imaging diagnosis. In this paper, a computeraided diagnosis (CAD) system is presented for breast microcalcifications based on dualtree complex wavelet transform (DTCWT) to facilitate radiologists like double reading.
Methods
Firstly, 25 abnormal ROIs were extracted according to the center and diameter of the lesions manually and 25 normal ROIs were selected randomly. Then microcalcifications were segmented by combining space and frequency domain techniques. We extracted three texture features based on wavelet (Haar, DB4, DTCWT) transform. Totally 14 descriptors were introduced to define the characteristics of the suspicious microcalcifications. Principal Component Analysis (PCA) was used to transform these descriptors to a compact and efficient vector expression. Support Vector Machine (SVM) classifier was used to classify potential microcalcifications. Finally, we used the receiver operating characteristic (ROC) curve and freeresponse operating characteristic (FROC) curve to evaluate the performance of the CAD system.
Results
The results of SVM classifications based on different wavelets shows DTCWT has a better performance. Compared with other results, DTCWT method achieved an accuracy of 96% and 100% for the classification of normal and abnormal ROIs, and the classification of benign and malignant microcalcifications respectively. In FROC analysis, our CAD system for clinical dataset detection achieved a sensitivity of 83.5% at a false positive per image of 1.85.
Conclusions
Compared with general wavelets, DTCWT could describe the features more effectively, and our CAD system had a competitive performance.
Background
Breast cancer is one of the most common cancers among female diseases all over the world. As the causes are unknown, early diagnosis and treatment is particularly important in reducing the mortality rate. Currently, the most effective method for early detection of breast cancer is mammography [1, 2], and microcalcifications are important sign of earlystage breast cancer, which are tiny deposits of calcium that appear on the mammograms as bright spots. Because of subtle findings, radiologist distraction or complex architecture the number of false negative mammograms are higher during evaluating screening mammograms. Consequently, about 1030% cases are missed during the routine check [1]. With the development of computer technology, a computeraided diagnosis system comes into being, which can be useful for the radiologists as a second reader. Balleyguier et al. [3] showed that the use of CAD is more useful for the junior radiologist with an improvement in sensitivity from 61.9% to 84.6% compared to a slight improvement from 76.9% to 84.6% for the experienced radiologist.
Generally, a typical CAD system consists of four stages including image preprocessing, extraction of ROIs, detection of microcalcifications, feature extraction and classification.
Detection of microcalcifications
There are many methods of microcalcifications detection in mammograms, mainly including traditional image processing methods, filtering, threshold algorithm, neural network, SVM, etc. Boccignone G et al. [4] proposed a method combining multiresolution analysis based on wavelet transform and threshold segmentation based on Renyi entropy. Nakayama R et al. [5] decomposed the mammograms using orthogonal 2D wavelet transform to obtain Hessian matrix of every pixel, and then microcalcifications were detected by computing the Eigen values of matrixes. In [6] the breast area was first segmented using morphological filtering and threshold method, then the difference image was obtained by subtracting the noisesuppressed image from the enhanced image of the breast area. Finally, microcalcifications were segmented by classification of difference images based on the neural network classifier. Melloul M et al. [7] described a threshold segmentation algorithm using entropy to select the threshold automatically and freely. All algorithms mentioned above had their advantages and drawbacks respectively.
Feature extraction and classification
According to Breast Imaging Reporting and Data System of the American Radiology College [8], the characterizations of microcalcifications include size, shape and distribution. Currently, the descriptions of the extraction area are mainly gray features, shape features and texture features. Gray features reflect the density of breast tissue and contrast between the lesion and surrounding tissue. The common used features are mean, variance, contrast, etc. Kinoshita SK et al. [9] calculated the histogram statistical characters of breast area, including mean, variance, skewness, kurtosis and entropy. Reference [9] used particle size character to describe the structures distribution of the different size, and the Leyden domain characterization to describe the distribution of linear structure. Yu S et al. [10] used mixing features consisting of wavelet features and gray level statistical features as inputs to a multilayer neural network. The feature space describing the mammograms is often large and complex. Therefore, feature selection is an essential work. The common methods include principal component analysis, linear decision analysis, logistic regression, backward selection, onedimensional analysis and genetic algorithm. There are also many classification methods, including linear discriminate analysis, artificial neural network, Bayesian methods, rulebased detection methods, decision tree, etc. [11]. Besides, many studies used the mixed classification [12].
DTCWT
DTCWT, proposed by Kingsbury [13] is a recent enhancement to the Discrete Wavelet Transform (DWT). Compared to DWT, it is nearly shiftinvariant, good multidirection analysis and limited data redundancy which make it appropriate for feature extraction. High directional selectivity is useful for texture analysis.
DTCWT employs two real DWTs; the first gives the real part of the transform while the second gives the imaginary part [14]. Figure 1 shows the framework scheme of DTCWT for 1D signal. In the figure, X stands for the input signal.h _{0} (n) and h _{1} (n) represent the lowpass and highpass filter pair for the upper filter bank, g _{0} (n) and g _{1} (n) represents the lowpass and highpass filter pair for the lower filter bank, respectively. 2D DTCWT is the extension of 1D DTCWT. It is essentially performs 1D DTCWT on rows and columns of the image in parallel. After 2D DTCWT, consisting of four parallel 2D DWT, we can get six directional subbands (±15^{0}, ±45^{0}, ±75^{0}) and the first and latter two parallel transforms produce the real and imaginary part of the six subbands, respectively.
The idea behind DTCWT is inspired from Fourier Transform (FT) which does not suffer from shift variance. This property of FT is based on complexvalued oscillating signals which form a Hilbert pair out of phase components.
Kingsbury proposed to construct a complexvalued wavelet basis which also forms a Hilbert pair.
Therefore, DTCWT can decompose a signal into real and imaginary components as FT. these components are obtained separately using different filter banks. The proposed wavelets in [15] were used in this paper.
Methods and results
Dataset and ROI selection
The mammograms used in our experiments were selected from the Mammography Image Analysis Society (MIAS) and a clinical dataset of mammograms from the Capital Medical University affiliated hospital. The study was approved by the Ethics Committees of Capital Medical University. ROIs of clinical dataset of mammograms selected and decided by four experienced senior radiologists. MIAS is open access to researchers. This dataset contains 322 images, in which there are 207 normal cases and 115 abnormal cases. Each mammogram is of 1024×1024 pixels, with a spatial resolution of 50 μm/pixel. Figure 2 and Figure 3 show a typical mammogram in MLO view from MIAS and clinical dataset respectively. The coordinates of abnormality center and the approximate radius are provided. Knowing the location and the approximate size of abnormality allows us to manually extract ROIs with proper dimension.
In our experiment, we used 50 ROIs with a size of 128×128 pixels, in which there are 25 normal ROIs and 25 ROIs containing microcalcifications. Figure 4 shows eight ROIs containing microcalcifications. After we obtained the ROIs, we stretched their gray scales to [0, 255] according to the following expression:
Where G _{0}(x, y) is the original grey value of the point(x, y), G _{1}(x, y) is the grey value of the point(x, y) in the ROI after grey normalization.
Detection of microcalcifications
In the mammograms, microcalcifications are usually brighter and smaller than the surrounding normal tissue. From the points of frequency domain, the microcalcifications mainly consist of highfrequency component, while the background mainly consists of the low frequency component. In this section, we used a method combining difference image technique and wavelet transform. The steps are as follows:

1.
Difference image technique
Since the Laplace filtering can sharp a figure and enhance the highfrequency components, while smooth filtering can blur a figure and suppress the highfrequency components, the highfrequency image are constructed by the following scheme. f_{1} represents figure filtered by Laplace operator, f_{2} represents figure filtered by 3×3 smooth filtering operator f_{3} was obtained Subtract f_{2} from f_{1} and image. Then detect the edge of f_{3} using Kirsch operator and do binary segmentation.

2.
Wavelet transform
By the simulation and analysis, we think the highpass subband of first level consists of highfrequency components and the highpass subband of fourth level consists of some lowfrequency components. Thus, the original image was decomposed by Daubechies wavelet at level four and the coefficients of the first layer and the fourth layer were set to zero. Then the wavelet reconstruction was done. Finally threshold segmentation was done.

3.
After the first step, we got a binary image containing microcalcifications and noise. Also, after the second step, we got a binary image containing microcalcifications and background. Then, logical ‘And’ operation was taken on the results of the previous two steps. The partial results were shown in Figure 5.
Feature extraction
Feature extraction and selection are important steps in CAD. DTCWT has a better directional selection than general DWT, which describes the directional features of texture preferably. So DTCWT was used to get the wavelet features.
In order to reduce the calculation burden and complexity, we used PCA to choose a few key features from the potential microcalcifications features. PCA can reduce the high dimensional correlated features into low dimensional features. In PCA, the symmetric covariance matrix or symmetric correlation matrix are calculated, and the eigenvalues and eigenvectors of these matrix are calculated. By PCA, a few irrelevant principal components are removed so that can retain the original information as complete as possible.
In our work, we choose fourteen features, including 2nd moment, 4th moment, 9 texture features and 3 wavelet coefficients. We used different wavelets in extracting the features. Then we transformed the 14 features to 4 linearly independent features. We have gotten a final feature called comprehensive score according to the contributions of the former 4 features. The final features of 50 ROIs are shown in Figure 6. The figures shows ROIs of the two categories have differences, and DTCWT indeed has a better performance. The detailed descriptions of features used were as follows:

1.
If the gray scale, gray average and histogram of an image are K, μ and h (κ), κ=0, 1, 2, …, K1 respectively, the norder moment of the mean of histogram is defined as:
$${m}_{n}=\frac{1}{N}{\displaystyle \sum _{k=0}^{K1}{\left(k\mu \right)}^{n}}h\left(k\right)$$(3)
The secondorder moment, also known as variance, describes the visual roughness of the image. The thirdorder and fourthorder moment are called inclination and kurtosis, respectively, reflecting the asymmetry and the uniformity of the histogram. This paper used the secondorder and fourthorder moment.

2.
Texture features were calculated using the gray level cooccurrence matrixes (GLCM) [16]. The results are 9 texture parameters. Suppose GLCM are defined as Cij. The angular second moment, inertia, inverse difference moment, entropy, correlation, sum average, difference average, sum entropy and difference entropy are computed as follows:
$$\begin{array}{l}{{\displaystyle \text{T}}}_{1}={\displaystyle \sum _{i=0}^{K1}{\displaystyle \sum _{j=0}^{K1}{{\displaystyle \text{C}}}_{\mathit{ij}}^{2}}}\\ {{\displaystyle T}}_{2}={\displaystyle \sum _{i=0}^{K1}{\displaystyle \sum _{j=0}^{K1}{\left(ij\right)}^{2}}}{{\displaystyle C}}_{\mathit{ij}}\\ {{\displaystyle T}}_{3}={\displaystyle \sum _{i=0}^{K1}{\displaystyle \sum _{j=0}^{K1}\frac{1}{1+{\left(ij\right)}^{2}}}}{{\displaystyle C}}_{\mathit{ij}}\\ {{\displaystyle T}}_{4}={\displaystyle \sum _{i=0}^{K1}{\displaystyle \sum _{j=0}^{K1}{{\displaystyle C}}_{\mathit{ij}}}}log{{\displaystyle C}}_{\mathit{ij}}\\ {{\displaystyle T}}_{5}=\frac{{\displaystyle \sum _{i=0}^{K1}{\displaystyle \sum _{j=0}^{K1}\left(\mathit{ij}\right){{\displaystyle C}}_{\mathit{ij}}}}{\mu}_{x}{\mu}_{y}}{{\sigma}_{x}{\sigma}_{y}}\\ {{\displaystyle T}}_{6}={\displaystyle \sum _{k=0}^{2K2}k{{\displaystyle C}}_{x+y}\left(k\right)}\\ {{\displaystyle T}}_{7}={\displaystyle \sum _{k=0}^{K1}k{{\displaystyle C}}_{xy}}\left(k\right)\\ {{\displaystyle T}}_{8}={\displaystyle \sum _{k=0}^{2K2}{{\displaystyle C}}_{x+y}\left(k\right)log\left\{{{\displaystyle C}}_{x+y}\left(k\right)\right\}}\\ {{\displaystyle T}}_{9}={\displaystyle \sum _{k=0}^{K1}{{\displaystyle C}}_{xy}\left(k\right)log\left\{{{\displaystyle C}}_{xy}\left(k\right)\right\}}\end{array}$$(4) 
3.
We directly decomposed the ROIs using 2D DTCWT and 12 subbands to get the wavelet coefficients. Then use the wavelet coefficients as the characteristics to classify the lesions.

4.
The area is estimated according to the size of the image foreground. The area is measured by the number of pixels roughly. The different weights are added to the different pixels to compensate for the discrete pixels description of the continuous image. The Euler number (E) is a topology descriptor of a region, which describes the connectivity of the region. For a given area, the number of holes (H) and the number of connectivity crew(C) in the region are commonly used in the topological properties. We get E = CH. In our paper we computed the area and Euler number of the microcalcifications.
Classification based on SVM
In order to get a quantitative result we used the SVM classifier. SVM was first introduced by Vapnik and it offers several advantages such as better performance in higher dimension space [17]. The basic idea of SVM is to find a hyperplane which can best separate the input feature vectors of two classes while maximizing the distance from either class to the hyperplane. In this paper, we use nonlinear SVM with the quadratic and cubic Polynomial Kemel functions.
The classification process included two steps. In the first step, samples corresponding to benign and malignant cases were labeled into a single class named abnormal class, while the samples not containing microcalcifications were denoted by normal cases. In the second step, the task was to discriminate between benign and malignant samples. These two steps were both a binary classification problem.
In order to have an adequately representative training set, we adopted a leaveoneout crossvalidation in the experiments. In the first step, we accomplished fifty different runs, where fortynine images were used for training and the remaining one for testing. Similarly, in the second step we accomplished twentyfour different runs, where twentythree images were used for training and the rest one for testing. The correct recognition rates (accuracy) of different wavelets are shown in Table 1, while the accuracy is defined as the ratio of correctly classified samples versus all samples. Table 2 shows some classification results of other references using different wavelets. Compared with other results, DTCWT method achieved an accuracy of 96% and 100% for the classification of normal and abnormal and the classification of benign and malignant, respectively. The results also show that the generalizability of this approach is quite well.
Performance evaluation of CAD
The ROC curve and the FROC analysis are common methods in evaluating CAD performance. In the ROC method, the observer classifies each image as normal or abnormal. The ROC curve is defined as a twodimensional plot of True Positive Fraction (TPF) and False Positive Fraction (FPF). The ROC curve of MIAS is shown in Figure 7. Similarly, the FROC is used to assess the ability to correctly detect the abnormalities. The FROC curve is obtained by plotting the sensitivity, synonymous with TPF, versus the number of false positives per image (FP/image). As shown in the Figure 8, a sensitivity value of 83% is obtained at a rate of 1.85 FP per image.
Discussion
Computeraided detection (CADe) systems address the problem that radiologists often miss signs of cancers that are retrospectively visible in mammograms. Furthermore, computeraided diagnosis (CADx) systems have been proposed that assist the radiologist in the classification of mammographic lesions as benign or malignant [22]. Key CAD techniques developed recently for breast cancer, including detection of calcifications, detection of masses, detection of architectural distortion, detection of bilateral asymmetry, image enhancement, and image retrieval [23]. Deserno TM et al. [24] recently presented an implementation of a SVMbased CBIR (contentbased image retrieval) system for CADx in screening mammography.
This paper proposed a new CAD method of microcalcifications based on DTCWT. The new diagnosis algorithm to detect and classify microcalcifications is verified by MIAS mammograms and clinical dataset. By combining special difference image technique and wavelet transform, the suspicious microcalcifications can be effectively segmented from mammograms. It shows DTCWT is efficient in describing image features. By PCA, merely four principal components out of fourteen original features are obtained to describe the microcalcifications. SVM is applied to classify the microcalcifications. Compared with other results, DTCWT method achieved an accuracy of 96% and 100% to the classification of normal and abnormal and the classification of benign and malignant, respectively. The experiments showed that our attempt is efficient. To provide more reliable evaluation results on the performance of the proposed scheme, a larger database should be used in future work.
Future the scholars will remain focusing on how to improve the performance of CAD. However, there are two issues that limit the current development of CAD. These are the inability to optimize a scheme for clinical impact—current methods only optimize CAD in the absence of a radiologist—and the lack of a figure of merit that quantifies the performance efficiency of CAD [25]. Such a figure of merit could be used to determine how much better performance CAD could obtain and which component of the several techniques employed is the weakest. The future work should be focused on how to solve these issues which may need more efforts. Furthermore, CAD must not be responsible for omitting the step of the complete evaluation of mammograms by the radiologist. A CAD system cannot and should not replace the radiologist as either or final interpretation [26].
Conclusions
In conclusion,compared with other wavelets (Haar, DB4), DTCWT could describe the features more effectively. The paper showed our CAD system had a competitive performance. The use of a CAD system helps the radiologist as a second reviewer to evaluate screening mammograms.
Authors' information
About the Author—Wushuai Jian Currently he is a Ph.D candidate at the College of Biomedical Engineering, Capital Medical University, Beijing, P.R. China. His research interests are medical image processing and phasecontrast imaging.
About the Author—Xueyan Sun received the Master of Science degree from the College of Biomedical Engineering, Capital Medical University, Beijing, P.R. China in 2012. Her research interests include: medical image processing, pattern recognition.
About the Author—Shuqian Luo Currently, he is a Full Professor, College of Biomedical Engineering, Capital Medical University, Beijing, China, director of Medical Imaging Lab, IEEE Senior Member, project Leader: MultiModality Medical Image Registration, Brain Tissue Segmentation and Classification, 3D Digitalized Human Brain Atlas, Chinese Digital Human. He is editor of many Chinese journal and principle investigator of many projects, including the National High Technique Research and Development Plan (863 Plan), Project of National Natural Science Foundation. He won Henan Science and Technology Progress Second Place Award, Project of MultiFunctional ECG Analyzer, 1993, and Asian Ten major CT (Computed Tomography) Science Award, Project of MultiModality Medical Image Registration, 1999, and Beijing Traditional Chinese Medicine First Place Award, Project of Meridian Adjustment Diagnosis and Therapy System, 2001.3. Prof. Luo has published 150 papers and 6 books.
Abbreviations
 CAD:

Computeraided diagnosis
 DTCWT:

Dualtree complex wavelets transform
 ROI:

Regions of interest
 PCA:

Principal component analysis
 SVM:

Support vector machine
 ROC:

Receiver operating characteristic
 FROC:

Freeresponse operating characteristic.
References
 1.
Cheng HD, Cai X, Chen X: Computeraided detection and classification of microcalcifications in mammograms: a survey. Pattern Recognition 2003, 36(12):2967–2991. 10.1016/S00313203(03)001924
 2.
Cheng HD, Shi XJ, Min R: Approaches for automated detection and classification of masses in mammograms. Pattern Recognition 2006, 39(4):646–668. 10.1016/j.patcog.2005.07.006
 3.
Balleyguier C, Kinkel K, Fermanian J: Computedaided detection (CAD) in mammography: does it help the junior or the senior radiologist? Eur J Radiol 2005, 54(1):90–96. 10.1016/j.ejrad.2004.11.021
 4.
Boccignone G, Chianese A, Picariello A: Computer aided detection of microcalcifications in digital mammograms. Comput Biol Med 2000, 30(5):267–286. 10.1016/S00104825(00)000147
 5.
Nakayama R, Uchiyama Y, Yamamoto K: Computeraided diagnosis scheme using a filter bank for detection of microcalcifications clusters in mammograms. IEEE Trans Biomed Eng 2006, 53(2):273–283. 10.1109/TBME.2005.862536
 6.
Papadopoulos A, Fotiadis DI, Likas A: An automatic microcalcifications detection system based on a hybrid neural network classifier. Artif Intell Med 2002, 25(2):149–167. 10.1016/S09333657(02)000131
 7.
Melloul M, Joskowicz L: Segmentation of microcalcifications in Xray mammograms using entropy thresholding. Comput Assist Radiol Surg 2002, 1–6.
 8.
American College of Radiology: ACR BIRADS Mammography: Breast Imaging Reporting and Data System, Breast Imaging Atlas. Preston: ACR; 2003.
 9.
Kinoshita SK, de AzevedoMarques PM, Pereira RR: Contentbased retrieval of mammograms using visual features related to breast density patterns. J Digit Imaging 2007, 20(2):172–190. 10.1007/s1027800790040
 10.
Yu S, Guan L: A CAD system for the automatic detection of clustered microcalcifications in digitized mammogram films. IEEE Trans Med Imaging 2000, 19(2):115–126. 10.1109/42.836371
 11.
Giger M, Huo Z, Kupinski M: ComputerAided Diagnosis in Mammography. Handbook of Medical Imaging. Washington, DC: SPIE; 2000:915–986.
 12.
Li L, Zheng Y, Zhang L, Clark RA: Falsepositive reduction in CAD mass detection using a competitive classification strategy. Med Phys 2001, 28(2):250–258. 10.1118/1.1344203
 13.
Kingsbury NG: The dualtree complex wavelet transform: a new efficient tool for image restoration and enhancement. Proc European Signal Processing Conf; 1998:319–322.
 14.
Selesnick IW, Baraniuk RG, Kingsbury NG: The dualtree complex wavelet transform. IEEE Signal Proc Mag 2005, 22(6):123–151.
 15.
Shi HL, Hu B, Zhang JQ: A novel scheme for the design of approximate Hilbert transform pairs of orthonormal wavelet bases. IEEE Trans Signal Process 2008, 56(6):2289–2297.
 16.
Shuqian L, Guohong Z: Medical Image Processing and Analysis. 1st edition. Beijing: Science; 2003:33–35.
 17.
Cortes C, Vapnik V: Supportvector networks. Mach Learn. 1995, 20(3):273–297.
 18.
Verma B, Zakos J: A computeraided diagnosis system for digital mammograms based on fuzzyneural and feature extraction techniques. IEEE Trans Inf Technol Biomed 2001, 5(1):46–54. 10.1109/4233.908389
 19.
Panchal R, Verma B: Neural classification of mass abnormalities with different types of features in digital mammography. Int. J Comp Intel Appl 2006, 6(1):61–75. 10.1142/S1469026806001757
 20.
Lee S, Lo C, Wang C: A computeraided design mammography screening system for detection and classification of microcalcifications. Int J Med Inform 2000, 60(1):29–57. 10.1016/S13865056(00)000678
 21.
Ren JC, Wang D, Jiang JM: Effective recognition of MCCs in mammograms using an improved neutral classifier. Eng Appl Artif Intel 2011, 24(4):638–645. 10.1016/j.engappai.2011.02.011
 22.
Elter M, Horsch A: CADx of mammographic masses and clustered micro calcifications: a review. Med Phys 2009, 36(6):2052–68. 10.1118/1.3121511
 23.
Tang J, Rangayyan RM, Xu J, Naqa IE, Yang Y: Computeraided detection and diagnosis of breast cancer with mammography: recent advances. IEEE Trans Inf Technol Biomed 2009, 13(2):236–251.
 24.
Deserno TM, Soiron M, de Oliveira JEE: Computeraided diagnostics of screening mammography using contentbased image retrieval. Proc SPIE 2012, 8315: 271–279.
 25.
Nishikawa RM, Pesce LL: Fundamental limitations in developing computeraided detection for mammography. Nucl Instrum Meth A 2010, 648(Suppl 1):251–254.
 26.
Dromain C, Boyer B, Ferre R: Computedaided diagnosis (CAD) in the detection of breast cancer. Eur J Radiol in press
Acknowledgments
This work was supported by the National Natural Science Foundation of China (Grant Numbers. 60532090 and 30770593) and was partly supported by the 7th Framework Programme of the European Community (Grant Agreement Number: PIRSESGA2009269124).
Author information
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
WJ conceived of the study, and participated in its design and coordination and helped to draft the manuscript. XS worked on the algorithm design and implementation.SL contributed to discussion and suggestions throughout this topic, including the manuscript writing. All authors read and approved the final manuscript.
Authors’ original submitted files for images
Below are the links to the authors’ original submitted files for images.
Rights and permissions
About this article
Received
Accepted
Published
DOI
Keywords
 Microcalcifications
 Computeraided diagnosis
 Dualtree complex wavelet transform