An automated mammogram classification system using modified support vector machine

Introduction

Breast cancer is a deadly disease that plagued women population in developing countries.^1,2 However, early detection of the symptoms of breast cancer could facilitate its treatment and it has been established that mammography is the best technique for this task.^3,4 This technique involves the examination of X-ray images of the breast for symptoms such as architectural distortion, calcifications, masses, etc.

Although mammography has been established to be a viable technique in the diagnosis of breast cancer, its interpretation by Radiologists’ are most times in question as different Radiologists may come up with different interpretations.^5,6 Azar⁷ argued that interpretation of mammogram is a cognitive skill which Radiologists develop over time. Nevertheless, digital image processing and machine learning approaches can proffer a solution to many of the inherent problems encounter by Radiologists when interpreting mammogram manually.

Mammogram processing (enhancement and segmentation) enhances visual interpretation. Also, intelligent classifier aids expert in detection and classification of mammograms.⁸

Methodology

The detailed steps employed in the classification task are documented in this section. It involves mammogram acquisition, mammogram processing, feature extraction and the classification models, all of which were implemented in MATLAB 2015.

Mammogram acquisition

Raw mammograms were downloaded from MIAS database.⁹ Out of 322 mammograms available in the database, 126 normal, 60 benign and 48 malignant mammograms were carefully selected. Of these, 37 normal, 23 benign, 18 malignant mammograms were used as test data while the remaining 156 mammograms were used for training the system.

Mammogram processing

Mammography uses low dose x-ray for imaging which produces a low qualitymammograms.¹⁰ In addition, Kayode et al¹¹ ascertained that image acquisition comes with lots of inherent problems which include scratches, dust artifact, scanner induced artifacts and excessive background noise which can further undermine the reliability of X-ray images, therefore, it is expedient to pre-process digital images so as to improve their quality.

Furthermore, image enhancement includes removing noise and adjusting image contrast so that the identification of key features becomes easier. In this article, image enhancement was achieved using Contrast Limited Adaptive Histogram Equalization (CLAHE). Also, segmentation was done in order to separate the needed region also known as the region of interest (ROI) from the entire breast tissue. Otsu threshold algorithm was implemented in MATLAB to extract ROI from the entire mammogram image.

Extraction of features from mammogram’s ROI

Radiologists analyze ROI found on mammograms by examining morphological features, such as size, margin and shape of the ROI; this helps them to differentiate between normal and abnormal mammograms and also to differentiate between an abnormal benign mammogram from an abnormal malignant mammogram. However, it should be noted that decisions concerning these morphological features are cognitive and greatly depend on Radiologists’ experience and opinion; therefore, it is essential to use computers to facilitate the extraction of GLCM features related to image texture, coarseness and heterogeneity, which are not necessarily seen by unaided eye.

GLCM is the most common statistical method used to compute textural features from gray level images such as mammograms. GLCM features consider the spatial relationship between the pixel of interest and its neighbouring pixels thereby providing us with textural features. Each element (m, n) in the resultant GLCM is simply the sum of the number of times that the pixel with value m occurred in the specified angular direction and distance to a pixel with value n in the input image.¹²

In this article, the thirteen GLCM features, proposed and by Haralick et al¹³ with two other features called cluster prominence and cluster shade, proposed by Tsatsoulis¹⁴ were extracted from mammograms. These GLCM features have been reported to adequately convey information about the textural characteristics of an image. The features are listed in Table 1. The detailed discussion about the features is documented in^13,14 respectively.

Table 1 Textural features extracted from the dataset

Unlike in the existing works,^15–18 where a single GLCM feature was created for each image using horizontal direction, and distance, this study is of the opinion that a single GLCM is inadequate to characterize the textural features of an image, therefore, in addition to the horizontal direction and distance two diagonals and the vertical direction at distance at distance which correspond to multiple GLCMs features at four directions and two distances 1 and 2 were computed. The following pseudocode was used to extract the features:

Objective:

1. To Extract GLCM from mammogram images

Input:

2. ROI of the enhanced image I

3. The direction to be used for creating a co-occurrence matrix

Output:

4. The GLCM extracted features vector from the input image I

Process:

5. Begin:

6. for do

7. create //GLCMs at angular directions 

8. end loop

9. for distances do

10. //sum of co-occurrence matrices

11. for each angle in do

12.

13. end loop

14. end loop

15. for i=1 to 15 do

16.

17. end loop

18. for i=1 to 15

19.

20.

21. end loop features

22. Return

23. End

Classification of mammograms

SVM was used for classification purposes. In MATLAB programming tool, SVM functions support only two classes based on its fundamental principle as a binary classifier. Due to three attributes of interest: normal, benign and malignant, SVM was trained twice, first to classify mammograms to either normal or abnormal, and second, to classify abnormal mammograms to benign or malignant. By this two-stage approach the three attributes of mammograms will be accommodated as illustrated in Figure 1.

Figure 1 Two-level classification approach using support vector machine (SVM).

Experimental results

Figure 2 illustrates the Graphic User Interface (GUI) designed for the preprocessing stage. Raw mammograms were fed into the system one after the other then CLAHE was employed for the preprocessing. Afterwards, the suspicious region otherwise known as ROI was extracted from the enhanced image; this is illustrated in Figure 3.

Figure 2 Mammogram enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE).

Figure 3 Region of interest (ROI) extraction.

Fifteen GLCM features presented in Table 1 were extracted from the ROI as illustrated with Figures 4 and 5 respectively. The averages of the features at the two distances were calculated as shown in Figure 6 while the overall averages of each of the features which serve as input to SVM classifier, were also calculated as shown in Figure 7.

Figure 4 Gray level co-occurrence matrix (GLCM) features computed at D=1.

Figure 5 Gray level co-occurrence matrix (GLCM) features computed at D=2.

Figure 6 Weighted gray level co-occurrence matrix (GLCM) features computed at distances 1 and 2.

Figure 7 Overall average gray level co-occurrence matrix (GLCM) features computed.

The first-stage classification entails categorizing the mammograms as either normal or abnormal as illustrated in Figure 8; this is termed the first-stage classification. If the image is normal the algorithm stops. However, if any abnormality is detected, the algorithm further classifies the abnormality into benign or malignant (see Figure 9); this is second-stage classification.

Figure 8 First-stage of classification: (A) normal mammogram classification; (B) abnormal mammogram classification.

Figure 9 Second-stage classification: (A) benign; (B) malignant.

System testing and performance evaluation

The model was tested to determine if the learning algorithm really performed its recognition task of differentiating normal and abnormal image on one hand and malignant and benign on the other hand.

The aim of this step is to introduce the model to the 78 unseen mammograms after it has been trained with a training dataset so as to test how well it performed.

First-stage classification

Using the expert’s classification provided alongside the dataset as the actual class, in the first-stage classification, all the 37 normal images were classified as normal (no false positive) and all 41 abnormal images were correctly classified to be abnormal (no false negative) meaning that the model can accurately detect abnormality, that is, each of the sensitivity, specificity and accuracy of the model in detecting abnormality is 100%

Second-stage classification

After the detection of abnormality, the system further classified the abnormality on the mammograms to be either “Benign” or “Malignant”. This is the second-stage classification. Out of 23 benign images, 21 were classified as truly benign while 2 are misclassified to be malignant. Also, out of 18 malignant images, 17 were classified to be malignant while an image was misclassified to be benign. These values are entered into a confusion matrix as shown in Table 2.

Table 2 The system’s confusion matrix

Results from proposed technique and existing works

Table 3 presents the results of existing works who also classified MIAS database mammograms using SVM. These are compared with the results obtained from the proposed technique. Results obtained from existing works have classified mammograms in MIAS database to either normal or abnormal mammograms or benign or malignant mammograms. However, this work reported an automated MIAS database mammograms classification in which mammogram are classified first to abnormal and normal, after which the abnormal mammograms were further classified into benign or malignant mammograms.

Table 3 Results from proposed technique and existing works

Results and discussion

For the first stage classification, the sensitivity, specificity, PPV and NPV of the model are all 100%. This means that the system is 100% accurate of differentiating abnormal mammograms from normal ones.

Using the information on the confusion matrix labeled Table 2, the second level sensitivity, specificity; PPV and NPV of the model were calculated thus:

(1)

(2)

(3)

(4)

From the performance metrics, the probability that the system would detect a malignancy among patients that have cancer is 0.944, that is, sensitivity =94.4%; the probability that the system would classify a patient’s mammogram image as benign among patients that do not have cancer is 0.913 (specificity =91.3%). Also, the probability that a mammogram would be actually malignant when it is classified to be cancerous by the system is 0.895 (PPV =89.5%) and the probability that the mammogram would be benign when the system says it is not malignant is 0.955, that is, NPV =95.5%.

Conclusion

The significant variability that occurs when interpreting the same mammogram independently by different radiologists leads to diagnostic errors. It is therefore imperative to find an improved method to aid the detection and classification of ROI on mammograms. This article has discussed an automated mammogram classification system that uses a modified SVM classification technique. Performance evaluation results obtained suggested that the system could be used as radiologists’ potential tool for supporting decision making in mammogram interpretation. This could in turn help Radiologists to make an accurate and timely decision, thereby increasing the efficiency of their diagnostic skills.