Marina Milosevic , Dragan Jankovic and Aleksandar Peulic

Comparative analysis of breast cancer detection in mammograms and thermograms

De Gruyter | Published online: October 28, 2014

Abstract

In this paper, we present a system based on feature extraction techniques for detecting abnormal patterns in digital mammograms and thermograms. A comparative study of texture-analysis methods is performed for three image groups: mammograms from the Mammographic Image Analysis Society mammographic database; digital mammograms from the local database; and thermography images of the breast. Also, we present a procedure for the automatic separation of the breast region from the mammograms. Computed features based on gray-level co-occurrence matrices are used to evaluate the effectiveness of textural information possessed by mass regions. A total of 20 texture features are extracted from the region of interest. The ability of feature set in differentiating abnormal from normal tissue is investigated using a support vector machine classifier, Naive Bayes classifier and K-Nearest Neighbor classifier. To evaluate the classification performance, five-fold cross-validation method and receiver operating characteristic analysis was performed.

Introduction

Breast cancer screening with mammography has been shown to be effective for preventing breast cancer death [17]. It is very difficult to interpret mammograms. Major reasons are the very complex anatomy of the breast and the small differences in the image densities of various breast tissues, which is particularly true for dense breasts.

Many studies have focused on quantifying the textural content in the region of interest (ROI) of mammograms. Petrosian et al. [12] investigated the usefulness of texture features based on spatial gray-level dependence (SGLD) matrices for the classification of masses and normal tissues. With a dataset of 135 ROIs, the methods indicated 76% sensitivity and 64% specificity in the test step using the leave-one-out method. Kinoshita et al. [9] used a combination of shape and texture features based on gray-level co-occurrence matrices (GLCM) for the classification of benign and malignant breast lesions. They reported 81% accuracy with a dataset of 38 malignant and 54 benign lesions. Kegelmeyer [7] developed a method to detect stellate lesions in mammograms and computed Laws texture features from a map of local edge orientations. Detection results with five test images yielded a sensitivity of 83% with 0.6 false-findings per image. Mudigonda et al. [10] attempted classification of benign and malignant masses by computing gradient-based features and texture features based on GLCMs. Detection results with a total of 54 images containing 39 images from the Mammographic Image Analysis Society (MIAS) database and 15 images from a local database yielded the classification accuracy of 82.1%. Hadjiiski et al. [4] used the texture features calculated from SGLD matrices and run-length statistics matrices for classification of benign and malignant masses. With a dataset of 348 ROIs, they yielded the area under the receiver operating characteristic (ROC) curve value of 0.81. Kim and Park [8] presented a comparative study of texture-analysis methods. With a dataset of 86 ROIs, they reported the area under the ROC curve value of 0.88.

Although, mammography is considered the gold standard screening tool for the early detection of breast cancer, the performance of this procedure is less in younger women and relates to the difficulty of imaging dense breast tissue [16]. There are other breast-testing options that are more effective and safe. Thermography, also known as thermal imaging or infrared imaging is a non-invasive, non-contact system of recording body temperature by measuring infrared radiation emitted by the body surface. Infrared imaging allows detection of both breast cancer and potential breast cancer risk, which sometimes cannot be done through mammogram. The use of the thermography is based on the principle that metabolic activity and vascular circulation in both precancerous tissue and the area surrounding a developing breast cancer are almost always higher than in normal breast tissue [3].

Nowadays, thermography is becoming an increasingly popular diagnostic tool to detect various diseases. Jakubowska et al. [6] and Ng et al. [11] presented methods to segment the thermograms and to detect the malignant tumors in the breast by thermovision techniques. Tan et al. [15] used texture features to study the ocular thermograms in young and elderly subjects. Acharya et al. [1] used texture features and support vector machine (SVM) classifier to detect signs of breast cancer. With 36 images used for training and 14 thermograms used for testing, they reported the classification accuracy of 88.1%.

We first determined ROI, which involves the separation areas of the image that represents the breast from the background. The dark area near the skin line is enhanced and the pectoral muscle is filtered out, largely reducing the intensity range in the mammogram. Because of the nature of the thermographic image, the ROI is manually extracted from the original thermograms. In the present study, we focused on the texture measures based on GLCM for the classification of ROIs as normal or abnormal. To evaluate the classification efficacies of this texture-analysis method, three different classifiers were employed: an SVM classifier, Naive Bayes classifier and a K-Nearest Neighbor (k-NN) classifier. A ROC analysis was used to evaluate the classification performances of the textural features extracted by mentioned texture-analysis method.

Materials and methods

Image datasets

A comparative study of texture-analysis methods is performed for three image groups. The Mini-Mammographic Database [14] provided by the MIAS was used as the test bench. The database contains 322 different mammogram images, which were divided into seven categories according to different pathological types. We have chosen a set of 300 patterns, 187 without masses and 113 containing masses. All mammograms are digitized at 200-μm pixel edge, giving a 1024×1024 pixel resolution for each mammogram. The second data set in this study consists of automatically extracted ROIs from 300 mammograms that were randomly selected from the local database. The data set includes a mixture of abnormal (150) and normal (150) pattern. The abnormal mammograms contain at least one mass of varying size and location. The third data set was collected using non-contact thermography. Infrared thermograms were recorded using a VARIOSCAN 3021 ST sterling-cooled infrared camera from JENOPTIK (Germany) with a spectral sensitivity of 8–12 μm. The method was tested with a total of 50 images including 37 normal and 13 abnormal cases from patients who did not have or had a detected abnormality on clinical examination or breast imaging (mammography, ultrasound or magnetic resonance imaging). Abnormal thermograms mostly contain invasive ductal carcinoma in early development stage. Examination was done in a temperature-controlled room with the temperature range of 20–23°C. The patients were required to stay for a few minutes to stabilize and reduce the basal metabolic rate, which results in minimal surface temperature changes, and therefore, satisfactory thermograms.

Region of interest extraction

The inclusion of ROI extraction procedure in computer aided diagnosis systems can avoid useless processing time and data storage. The method for mammogram ROI detection is composed of several main steps, as described in the following text.

Detection of MLO view type

First, we removed the noise using a median filter. A median filter is a nonlinear digital filtering technique, widely used in digital image processing because, under certain conditions, it preserves edges while removing noise. After removing the noise, it is necessary to detect the type of MLO view: left-sided (LMLO) or right-sided (RMLO). The first step is skin line detection. The following notations are used to describe the algorithm for skin line detection:

I – the original mammographic image.

I(i,j) – pixel value in the ith row and jth column of image I.

B1 and B2 are the binary versions of original image with different threshold.

B1(i,j) and B2(i,j) – pixel value in the ith row and jth column of image B1 and B2, respectively.

Algorithm for skin line detection

If I(i,j)>5

 Then, B1(i,j)=1

 Else, B1(i,j)=0

If I(i,j)>15

 Then, B2(i,j)=1

 Else, B2(i,j)=0

Skin line = B1–B2.

Algorithm for RMLO test

Step 1: Start with first row.

Step 2: Scan from left to right side.

Step 3: If pixel is black then move to next pixel and repeat Step 3. Else, when pixel is white then go to Step 4.

Step 4: Replace current pixel with white and move to next pixel. Repeat Step 4 while current column is not the last column. If current column is the last column, go to Step 5.

Step 5: Repeat Steps 2–4 for the next row.

Step 6: R=∣I-I1∣, where I1 is image obtained using the Algorithm for RMLO test (Steps 1–5).

Algorithm for LMLO test is similar to algorithm for RMLO test. The difference is that scanning is performed from right to left side.

L=∣I-I2∣, where I2 is image obtained using the Algorithm for LMLO test (Steps 1–5).

Let mR be the mean value of R and mL the mean value of L.

If mL>mR

 View is Left

Else, View is Right.

Now, when type of MLO view is known, the system can automatically apply the background removal procedure and pectoral muscle removal procedure.

Background portion removal

The background removal procedure will be explained for the RMLO view of the mammogram. The mentioned procedure for the LMLO view is very similar to this procedure.

The first step of background removal procedure is image contrast enhancement. The image contrast is enhanced by using a simple logarithmic operation. A logarithmic operation

(1) g ( x , y ) = log [ 7 + f ( x , y ) ]  (1)

is applied to the original image f(x, y); g(x, y) is the transformed image (Figure 1A). This operation applied to the whole image significantly enhances the contrast of the regions near the breast boundary in mammograms, which are characterized by low density and poor definition of details. Then, a binarization procedure is applied to the image (Figure 1B). Every single pixel with a value greater than 128 and less than 255, is set on “1”. In other cases, pixel value is set on “0”.

Figure 1 Background removal procedure. (A) Image with enhanced contrast, (B) Binarized image, (C) Mask, (D) Extracted breast region.

Figure 1

Background removal procedure. (A) Image with enhanced contrast, (B) Binarized image, (C) Mask, (D) Extracted breast region.

To remove breast background, it is necessary to create and apply a mask. The mask is created by applying the algorithm below to the improved binary image.

Algorithm for the mask formation

Step 1: Start with first row.

Step 2: Scan from left to right side.

Step 3: If pixel is black move to next pixel and repeat Step 3. Else, when pixel is white then go to Step 4.

Step 4: Move to next pixel while pixel is white. When pixel is black then go to Step 5.

Step 5: Replace current pixel with black and move to next pixel. If current column is the last column, go to Step 6, else repeat Step 5.

Step 6: Repeat Steps 2–5 for the next row.

Masking is performed by element wise multiplication with original mammogram. Figures 1(C) and 1(D) show image mask and mammogram with removed background, respectively.

Pectoral muscle removal

Pectoral muscle tissue is usually denser than the rest of the breast. Therefore, pectoral muscle and a central part of the breast will be extracted by applying the threshold operation. Input image is mammogram with removed background.

Algorithm for the pectoral muscle removal

Step 1: Define the threshold value, based on intensity of pectoral muscle.

Step 2: If pixel value is equal or greater than the threshold, pixel retains its value. Every pixel with a value less than threshold is set on “0”.

Step 3: Start with first row and nth column, n is the first non-zero pixel in the current row.

Step 4: Scan from left to right side.

Step 5: If pixel is non-zero then replace it with zero and move to next pixel and repeat Step 5.

Else, when pixel is zero then start from the next row and nth column and repeat Step 5.

Stop the procedure when all rows are exhausted.

Step 6: Isolate the pectoral muscle by subtracting the image with the central part of the breast, obtained in Steps 3–5, from the image with extracted pectoral muscle and a central part of the breast.

Figure 2(A) shows pectoral muscle and a central part of the breast extracted by applying the threshold operation. The mammogram without background and pectoral muscle (Figure 2B), is obtained by subtracting the isolated pectoral muscle from the mammogram with removed background.

Figure 2 Pectoral muscle removal procedure. (A) Breast region with removed background after local thresholding, (B) ROI.

Figure 2

Pectoral muscle removal procedure. (A) Breast region with removed background after local thresholding, (B) ROI.

The pectoral muscle removal procedure for the LMLO type of mammogram is analogous to the here explained procedure.

Thermograms

In the case of thermographic images, the ROI is manually extracted from the original thermograms. As shown in the Figure 3(A), every part of the body has a specific color and each color indicates a certain temperature. Figure 3(B) is the corresponding grayscale image. The cropped left and right breasts are shown in parts (C) and (D) of the same figure.

Figure 3 Thermograms of patient with cancer in the left breast. (A) Original, (B) Grayscale version, (C) Cropped left breast, (D) Cropped right breast.

Figure 3

Thermograms of patient with cancer in the left breast. (A) Original, (B) Grayscale version, (C) Cropped left breast, (D) Cropped right breast.

Features extraction and classification

The presence of masses causes architectural distortion in the surrounding tissues. As a result, mammographic images possess textural information that could bear discriminant features. Some of the most commonly used texture measures are derived from the GLCM. The GLCM is created by calculating how often a pixel with the intensity value i occurs in a specific spatial relationship (i.e., in a specified orientation θ and specified distance d from each other) to a pixel with the value j. Each element (i,j) in the resultant GLCM is simply the sum of the number of times that the pixel with value i occurred in the specified spatial relationship to a pixel with value j in the input image.

In the present study, four matrices corresponding to four different directions (θ =0°, 45°, 90°, and 135°) and one distance (d=1 pixel) were computed for each selected mammogram and thermogram ROI. A pixel distance of d=1 is preferred to ensure large numbers of co-occurrences derived from the ROI. Four values were obtained for each feature corresponding to the four matrices. The mean of these four values were calculated, comprising a total of 20 GLCM features. The 20 texture descriptors extracted from GLCM texture measurement are: Angular Second Moment (Energy), Contrast, Correlation, Variance, Inverse Difference Moment (Homogeneity), Sum Average, Sum Variance, Sum Entropy, Entropy, Difference Variance, Difference Entropy, Information Measure of Correlation 1 and Information Measure of Correlation 2 proposed by Haralick [5], Autocorrelation, Dissimilarity, Cluster Shade, Cluster Prominence and Maximum Probability proposed by Soh [13] and features Inverse Difference Normalized and Inverse Difference Moment Normalized proposed by Clausi [2].

The training and testing of the classifier for textural feature set was performed using the cross-validation methodology. Classification of the normal and abnormal cases was conducted by using the three different classifiers: SVM, K-Nearest Neighbor classifier and Naive Bayes classifier with diagonal covariance matrix estimate. SVM is based on the principle of structural risk minimization, which aims at minimizing the bound on the error made by the learning machine on data unseen during training, rather than minimizing the mean square error over the data set. As a result, an SVM tends to perform well when applied to data outside the training set. The K-Nearest Neighbor is a simple yet robust classifier where an object is assigned to the class to which the majority of the nearest neighbors belong. It is important to consider only those neighbors for which a correct classification is already known. Naive Bayes classifier with diagonal covariance matrix estimate is a simple probabilistic classifier based on Bayes’ theorem with strong independence assumptions. An advantage of the Naive Bayes classifier is that it requires a small amount of training data to estimate the parameters necessary for classification.

Results and discussion

A total of 300 mammograms (187 normal and 113 abnormal) from the MIAS database, 300 digital mammograms (150 normal and 150 abnormal) from a local database and 50 thermography images of the breast (37 normal and 13 abnormal patterns) were analyzed. A comparative study of the classification accuracy is performed for the all three types of images.

The classifiers are trained and tested using five-fold cross-validation to ensure exhaustive testing with all samples. Using this technique we divided the data set at random into a set of K=5 distinct sets. Training is then performed on K-1 sets and the remaining set is tested. This is then repeated for all of the possible K training and test sets. The classification results are average of all K results.

The performance in recognition can be evaluated by three factors: accuracy (AC), sensitivity (SE) and specificity (SP) of detection. They are defined as follows:

  1. Classification accuracy is dependent of the number of samples correctly classified.

    (2) A C = T P + T N T P + F P + T N + F N  (2)

  2. Sensitivity is a proportion of positive cases that are well detected by the test.

    (3) S E = T P T P + F N  (3)

  3. Specificity is a proportion of negative cases that are well detected by the test.

    (4) S P = T N T N + F P  (4)

where, TP is the number of true positives, FP the number of false positives, TN the number of true negatives, FN the number of false negatives. The performance of the classification schemes for all image groups are summarized in Tables 13.

Table 1

Classification performance with mammograms from MIAS database.

Performance measures Mammograms from MIAS database
SVM k-NN Naive Bayes
TP 23 44 53
FP 24 49 74
FN 90 69 60
TN 163 138 113
Accuracy 62% 60.7% 55.3%
Sensitivity 20.4% 38.9% 46.9%
Specificity 87.2% 73.8% 60.4%

SVM, Support vector machine; TP, true positives; FP, false positives; FN, false negatives; TN, true negatives.

Table 2

Classification performance with mammograms from a local database.

Performance measures Mammograms from a local database
SVM k-NN Naive Bayes
TP 121 84 114
FP 20 71 32
FN 29 66 36
TN 130 79 118
Accuracy 83.7% 54.3% 77.3%
Sensitivity 80.7% 56% 76%
Specificity 86.7% 52.7% 78.7%

SVM, Support vector machine; TP, true positives; FP, false positives; FN, false negatives; TN, true negatives.

Table 3

Classification performance with thermograms.

Performance measures Thermograms
SVM k-NN Naive Bayes
TP 10 10 11
FP 3 3 5
FN 3 3 2
TN 34 34 32
Accuracy 88% 88% 86%
Sensitivity 76.9% 76.9% 84.6%
Specificity 91.9% 91.9% 86.5%

SVM, Support vector machine; TP, true positives; FP, false positives; FN, false negatives; TN, true negatives.

Results in Tables 1 and 3 indicate that there is no significant difference in classification accuracy between SVM, k-NN and Naive Bayes classifier that operate on ROIs from the MIAS database and thermogram ROIs. SVM classification of mammograms from the MIAS database has a high FN rate which resulted in a very low sensitivity. In the case of mammogram ROIs from the local database, the SVM classifier is recommended, while the k-N classifier should be avoided. Classification results of mammograms from the local database are much better than the classification results of MIAS mammograms, with an accuracy ratio of 83% according to 60.7%. A very high accuracy ratio of up to 90%, is obtained by classifying thermograms using the SVM classifier.

ROC analysis was employed to evaluate the performance of the texture-analysis method in classifying the ROIs into positive and negative ROIs. ROC curve graphs that correspond to classification of both types of mammograms and therograms are shown in Figures 46.

Figure 4 ROC graph. Classification of mammograms from MIAS database using SVM, k-NN and Naive Bayes classifiers. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 4

ROC graph. Classification of mammograms from MIAS database using SVM, k-NN and Naive Bayes classifiers. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 5 ROC graph. Classification of mammograms from the local database using SVM, k-NN and Naive Bayes classifiers. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 5

ROC graph. Classification of mammograms from the local database using SVM, k-NN and Naive Bayes classifiers. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 6 ROC graph. Classification of thermograms using SVM, k-NN and Naive Bayes classifiers. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 6

ROC graph. Classification of thermograms using SVM, k-NN and Naive Bayes classifiers. ROC, Receiver operating characteristic; SVM, Support vector machine.

Altogether, the SVM classifier has the best classification accuracy. Therefore, we showed three ROC curves obtained by using the SVM classifier for each of the three image groups in Figure 7. The area under ROC curve (AUC) is a metric that can be used to compare different analysis, in accuracy aspects. The results are considered more precise, when the AUC is large. The area under the curve of thermogram classification, shown in Figure 7, shows the prediction of highly accurate classification results in medical image diagnosis.

Figure 7 ROC graph. SVM classification of mammograms from MIAS database, mammograms from local database and breast thermograms. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 7

ROC graph. SVM classification of mammograms from MIAS database, mammograms from local database and breast thermograms. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figures 810 show the confusion matrices for the classification results obtained with SVM classification of all image groups. The confusion matrix is determined between target class and output class. The diagonal cells in each table show the number of cases that were correctly classified, and the off-diagonal cells show the misclassified cases. The blue cell in the bottom right shows the total percent of correctly classified cases (green) and the total percent of misclassified cases (red). Gray cells numbered 1 and 2 show sensitivity and specificity, respectively. Gray cells in the third column of confusion matrices represent the precision or positive predictive value (PPV) and negative predictive value (NPV). The PPVs and NPVs are the proportions of positive and negative results in classification tests that are true positive and true negative results. Mathematically, PPV and NPV can be expressed as:

Figure 8 Confusion matrix. SVM classification of mammograms from MIAS database. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 8

Confusion matrix. SVM classification of mammograms from MIAS database. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 9 Confusion matrix. SVM classification of mammograms from the local database. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 9

Confusion matrix. SVM classification of mammograms from the local database. ROC, Receiver operating characteristic; SVM, Support vector machine.

Figure 10 Confusion matrix. SVM classification of breast thermograms. SVM, Support vector machine.

Figure 10

Confusion matrix. SVM classification of breast thermograms. SVM, Support vector machine.

(5) P P V = T P T P + F P , N P V = T N T N + F N  (5)

Conclusion

The goal of this work was to compare the classification results of mammograms from the MIAS database to the classification results of mammograms from the local database, using the same GLCM features extracted from each ROI and three different classifiers: SVM, k-NN and Naive Bayes classifier. The aim of this research was also to make sure that it is possible to use breast thermograms for tumor detection using the same method. Additionally, we presented a procedure for the automatic separation the breast region (ROI) from the mammograms. Using a SVM classifier and 20 GLCM features derived from each ROI, we obtained the accuracy ratio of 88%, which corresponded to true-positive fraction of 76.9% at a false positive fraction of 8.1% for breast thermograms. In case of mammograms from the local database we obtained the accuracy ratio of 83.7%, which corresponded to true-positive fraction of 80.7% at a false positive fraction of 13.3%. The classification result obtained by classifying mammograms from the MIAS database is characterized by accuracy ratio of 62%, which corresponded to a true-positive fraction of 20.4% at a false positive fraction of 12.8%. Our study indicates that a SVM can be trained to effectively classify breast thermograms and mammograms from the local database. For the limited number of sample patterns, the experimental results for the breast thermograms are promising. However, a database that covers many more positive ROIs and negative ROIs will be needed for training of classifiers in order to apply our method to clinical situations.

Thermography is risk-free, pain-free and totally non-invasive, does not involve ionizing radiation and costs less than mammograms. Thermal imaging can detect tumorous activity as it begins to develop a blood supply to sustain its growth. Yet for mammography, the tumor has to have formed sufficient physical mass and size to be detected. However, mammograms have some benefit. Mammography identifies the location and boundaries of the tumor or mass within the breast for purposes of biopsy, lumpectomy, mastectomy, or radiation therapy. Mammograms also provide crucial feedback as to cancer growth or reduction during treatment. From the above mentioned, we can conclude that the best solution for the early breast cancer detection is a system that combines mammography and thermography. In the proposed system, the first step is making the thermography images of the breast for each patient. If the thermal image of the breast contains suspicious signs that indicate tumor, it is necessary to use mammography screening.

Future work will focus on the further improvement of classification performance by reducing the number of input features.

Authors contributions: Marina Milosevic, Dragan Jankovic and Aleksandar Peulic share first authorship of this article. All authors read and approved the final manuscript.

References

[1] Acharya UR, Ng EYK, Tan JH, Sree SV. Thermography based breast cancer detection using texture features and support vector machine. J Med Syst 2012; 36: 1503–1510. Search in Google Scholar

[2] Clausi DA. An analysis of co-occurrence texture statistics as a function of grey level quantization. Can J Remote Sens 2002; 28: 45–62. Search in Google Scholar

[3] Fok SC, Ng EYK, Tai K. Early detection and visualization of breast tumor with thermogram and neural network. J Mech Med Biol 2002; 2: 185–195. Search in Google Scholar

[4] Hadjiiski L, Sahiner B, Chan H-P, Petrick N, Helvie M. Classification of malignant and benign masses based on hybrid ART2LDA approach. IEEE Trans Med Imag 1999; 18: 1178–1187. Search in Google Scholar

[5] Haralick RM, Shanmugam K, Dinstein I. Textural features for image classification. IEEE Trans Syst Man Cybern 1973; 3: 610–621. Search in Google Scholar

[6] Jakubowska T, Wiecek B, Wysocki M, Drews-Peszynski C. Thermal Signatures for Breast Cancer Screening – Comparative Study. In: Proc. IEEE EMBS Conf Cancun Mexico 2003. Search in Google Scholar

[7] Kegelmeyer WP Jr. Evaluation of stellate lesion detection in a standard mammogram data set. Int J Pattern Recogn Artif Intell 1993; 7: 1477–1493. Search in Google Scholar

[8] Kim JK, Park HW. Statistical textural features for detection of microcalcifications in digitized mammograms. IEEE Trans Med Imag 1999; 18: 231–238. Search in Google Scholar

[9] Kinoshita SK, Marques PMA, Slates AFF, Marana HRC, Ferrari RJ, Villela RL. Detection and characterization of mammographic masses by artificial neural network. In: Karssemeijer N, Thijssen M, Hendriks J, Van Erning L, editors. Proc. 4th Int. Workshop Digital Mammography. Netherlands 1998: 489–490. Search in Google Scholar

[10] Mudigonda N, Rangayyan R, Leo Desautels J. Gradient and texture analysis for the classification of mammographic masses. IEEE Trans Med Imag 2000; 19: 1032–1043. Search in Google Scholar

[11] Ng EYK. A review of thermography as promising noninvasive detection modality for breast tumour. Intl J Thermal Sci 2009; 48: 849–859. Search in Google Scholar

[12] Petrosian A, Chan HP, Helvie MA, Goodsitt MM, Adler DD. Computer-aided diagnosis in mammography: classification of mass and normal tissue by texture analysis. Phys Med Biol 1994; 39: 2273–2288. Search in Google Scholar

[13] Soh LK, Tsatsoulis C. Texture analysis of SAR sea ice imagery using gray level co-occurrence matrices. IEEE T Geosci Remote 1999; 37: 780–795. Search in Google Scholar

[14] Suckling J, Parker J, Dance D, et al. The mammographic images analysis society digital mammogram database. Exerpta Medica 1994; 1069: 375–378. Search in Google Scholar

[15] Tan JH, Ng EYK, Acharya UR. Study of normal ocular thermogram using textural parameters. Infrared Phys Techn 2010; 53: 120–126. Search in Google Scholar

[16] Wishart GC, Campisi M, Boswell M, et al. The accuracy of digital infrared imaging for breast cancer detection in women undergoing breast biopsy. Eur J Surg Oncol 2010; doi:10.1016/j.ejso.2010.04.003. Search in Google Scholar

[17] Wittenberg Th, Wagner F, Gryanik A. Towards a computer assisted diagnosis system for digital breast tomosynthesis. Biomedizinische Technik 2012; 57(Suppl.1): 223–226. Search in Google Scholar

Received: 2014-5-20
Accepted: 2014-10-6
Published Online: 2014-10-28
Published in Print: 2015-2-1

©2015 by De Gruyter