Detection Of Abnormal Tissues In Magnetic Resonance Images

INTRODUCTION

The fundamental element in the human body is the cell which are continuously created and destroyed to maintain human health. Any unbalance in the process of new cell formation and old cell destruction leads to accumulation of excessive cells which lead to formation of masses generally called tumours. A tumour is technically a malignant intracranial solid neoplasm which is generally because of atypical, unrestrained cell division. [1] Tumours can form in any part of the body and can be benign which are harmless or malignant which can be cancerous.

The malignant tumours can prove to be fatal if not identified in the early stages of development. Early discernment of Brain Tumour is a difficult task as symptoms appear in the advanced stages of tumour. To provide a better result in brain tumour detection, Magnetic Resonance Image is most widely preferred. [2] To give better discovery of tumour without influencing an ordinary tissue is also an exceptionally difficult process. The purpose of medical imaging techniques is to ease the process of medical diagnosis by imaging the internal human body.

[3] The central nervous system of humans contains brain in the forefront, that is, in the anterior part. When patients describe symptoms likely to be caused by tumour, they are made to undergo various diagnostic methods to determine the causal factor. [4] Diagnostic methods involve biopsy or imaging (MRI or CT scan).

Pathologists check for the occurrence of tumour in a specimen of brain tissue when performing a biopsy to test for presence of abnormality. Biopsies are indicative of presence and pathology of tumour, but when surgery is to be performed, the extent of the tumour as well as its exact location in the brain should also be known, which becomes more apparent from an MRI scan of the patient.

MRI is a safer method in general as compared to CT scan because it eliminates use of harmful degenerative radiations. [5] Traditionally, diagnosis of medical images is performed manually which is highly dependent on the perception and judgement of the physician while analysing the required region. [6]

One of the most common difficulties faced by such technicians is high extent of similarity and very minute differences when the real and malignant part is compared in the gray scale MRI and also the labour intensive readings owing to the immense number of images to be studied for correct diagnosis. [7]

Hence, there arises a need for a method which allows for detection of abnormality without involving the factor of human error as well as involving machine learning. For transmitting, storing, retrieving, printing, processing, and displaying medical imaging information, the international standard DICOM, which stands for Digital Imaging and Communications in Medicine, is used. DICOM allows interoperability of medical imaging information and integration of image-acquisition devices from different manufacturers. [8] The standard specifies a set of services to store the media, format of the file, and a directory like structure to access medical. Files complying with DICOM standard have a bit depth of 16 bit with a ‘dcm’ extension and can be read in MATLAB thereby allowing for further processing. The DICOM images of a normal brain and that with tumour are depicted in Fig.1 and Fig.2 respectively.

The DICOM images obtained need to be pre-processed to obtain a similar size of images for comparative analysis. Pre processing also involves converting the .dcm format DICOM images to gray scale for the creation of their Gray Level Co-occurrence Matrix. The statistical tool to examine texture is known as GLCM. GLCM accounts for the spatial relationship among pixels. Occurrences of pixel pairs with similar values and spatial relationships are evaluated and calculated to create the GLCM which then characterizes the image’s texture features. [9] The texture features extracted from the GLCM are Contrast, Correlation, Energy, Homogeneity and Entropy. Feature extraction helps in reducing the dimensionality of the image so it can be compared or classified. The texture feature of contrast measures the variations in the gray-level co-occurrence matrix at the local or individual pixel level calculated as per equation (1).

The texture feature of correlation gives the combined probability of occurrence of the predefined pixel pairs and is found using equation (2). [10]

METHODOLOGY

The basic difference in the images of normal and abnormal brain arises from the resonance provided by the cells that form the tumour and hence can be depicted better by non-invasive MRI images which are pre-processed to a similar size and are used for extraction of texture features. The steps needed to be followed in order to detect abnormality present in the MRI images are given in the block diagram shown in Fig.3 below.

131445-10795

Fig.3: Algorithm for detection of abnormality

The texture feature of energy results in the sum of all elements squared in the GLCM and is an indicator of uniformity as the angular second moment. It can be obtained by the formula given in equation (3). [10]

= ? , (P ) (3)

The texture feature of homogeneity measures the proximity of the distribution of GLCM elements to the diagonal of the GLCM and can be calculated as per the following formula given in equation (4). [10]

= ? ,()(4)

1482090-24574500

The texture feature of entropy can be computed by the formula indicated in equation (5).

= ? , ?ln(P )P (5)

Here, x and y are the spatial coordinates of the pixel in the

GLCM of size K and is the probability of occurrence of the pixel denoted by x and y.

The extracted features are then used for training the SVM classifier. The supervised learning algorithm that is SVM analyses and classifies the data as a non-probabilistic binary classifier and projects the classification of a new sample in a particular category. [11]

Machine learning allows for division of data set into training and testing and once trained as per the decision theory can classify test data in the category that it should belong to. SVM creates a decision boundary called the hyper plane which distinguishes one class from another, the Normal from Abnormal in this case. The segregation of test data into normal and abnormal tests the accuracy of SVM model constructed on the basis of the training sample. The misclassification error given after testing is indicative of SVM accuracy and can be determined for all the features in the dataset.[12]

RESULTS

The texture features extracted from the GLCM of each image in the sample of images have been differentiated into five cases based on the abnormality type in the sample of images and for each case one normal image is compared with five abnormal images for each feature and tabulated in Table.1.

Table.1 Texture Features extracted from the data set

Case Type Contrast Correla Energy Homogene Entropytion ity I Normal 2.2723 0.052 1.559 0.0249 6.6597

Abnormal 2.1638 0.0034 1.5784 0.0277 6.958

Abnormal 2.0134 0.0035 1.6239 0.0268 6.8663

Abnormal 2.084 0.0029 1.6241 0.0271 6.7876

Abnormal 1.8478 0.0489 1.6949 0.0284 6.7175

Abnormal 1.8928 0.047 1.5644 0.0287 7.0244

II Normal 2.5099 0.0071 1.4734 0.0249 6.5689

Abnormal 2.4141 0.0025 1.736 0.0259 6.9954

Abnormal 2.4732 0.0036 1.6021 0.0252 6.701

Abnormal 2.3772 0.0063 1.5917 0.0261 6.7137

Abnormal 2.4002 0.0018 1.694 0.0256 6.9779

Abnormal 2.271 0.0042 1.5752 0.0261 6.9708

III Normal 2.9791 0.052 1.299 0.0245 6.8942

Abnormal 2.0204 0.0085 1.3585 0.028 7.1366

Abnormal 1.9172 0.0048 1.3793 0.0283 7.1363

Abnormal 1.9542 0.0019 1.3716 0.0281 7.0017

Abnormal 2.0447 0.0038 1.3464 0.0273 7.1204

Abnormal 2.0927 0.0251 1.3542 0.0254 6.9111

IV Normal 2.6184 0.0196 7.2459 0.0243 7.0265

Abnormal 2.7552 0.0141 9.2043 0.0252 7.172

Abnormal 2.7759 0.0027 7.8468 0.0249 7.5112

Abnormal 2.7493 0.0153 9.6926 0.0251 7.5064

Abnormal 2.7174 0.0062 9.8728 0.0249 7.3821

Abnormal 2.6700 0.0064 7.9929 0.0245 7.601

V Normal 2.6536 0.025 1.127 0.0233 7.4908

Abnormal 2.5361 0.003 1.231 0.0246 7.3134

Abnormal 2.2945 0.0029 1.2398 0.0261 7.3848

Abnormal 2.0316 0.0029 1.2257 0.0265 7.2861

Abnormal 2.4515 0.0058 1.2667 0.0254 7.0256

Abnormal 2.6852 0.0069 1.255 0.0246 6.6039

The extracted feature data is depicted in Fig.4 in the form of a stem diagram showing the feature values of abnormal cases with a transparent block indicating the normal levels of each feature in each of the five cases. Each transparent block represents the range of ‘Normal’ values while the relevant feature value for ‘Abnormal’ case can be comparatively analysed from the individual coloured stem plots for the different cases.

-4445-8890

Fig.4:Comparative representation of texture feature values

Fig.4 show that the abnormal values obtained are in a specific range only which is either all greater or all smaller than the normal values associated with those particular features. The cases show adequate discrimination in values for these features to be supportive of a clear classification.

Two-third of the samples of the data set is used for training while one -third of the samples have been used for testing the classifier. Since SVM classification depicts the features in a two-dimensional space with each feature being one of the dimensions, two of the features taken at a time are used to train the SVM and further test remaining samples against the trained SVM model. The misclassification error is determined for each feature taken against each of the other features and has been tabulated in Table.2.

3791585200660

Table.2 Inter-comparison of misclassification error of all features with each other

FEATURE Contrast Correlation Energy Entropy Contrast 0.3000 0.0667 0.2667 0.2667 Correlation 0.0667 0.5333 0.2667 0.3000 Energy 0.2333 0.3000 0.2667 0.1667 Entropy 0.2667 0.2667 0.1333 0.4000 Fig.7: SVM plot of Contrast vs. Homogeneity

The comparison Table.2 shows that when each feature is classified with each other feature, the lowest misclassification error is observed between Contrast and Correlation. Hence, Contrast and Correlation can be considered the most discriminatory features which shall be most prominent in distinguishing the sample of images into the Normal or Abnormal category for classification of tumour tissue. Their SVM plot is given in Fig.5.

347980-1090295

Fig.8: SVM plot of Contrast vs. Entropy The above graphs indicate the closeness of data to the hyper plane determined from the support vectors and thus indicate efficiency and accuracy of classification. Also, Energy and Entropy form the secondary set of features which do not play a deciding role in this type of classification adding to the element of generalisation in pattern recognition. Fig.5: SVM plot of Contrast vs. Correlation The extracted texture feature data is further analysed by A few SVM plots of the training and testing data for the other features plotted against Contrast are shown in Fig.6, 7 and 8.

38925593980

changing the linear decision boundary to a non-linear boundary which is more flexible as per the data and can provide better classification. The texture feature data of Contrast and Correlation is used to test the efficiency of SVM when using two different kernel functions – Linear and Gaussian. Their respective accuracies are given in Table.3 and the Gaussian SVM plot is shown in Fig.9. The box plot for the entire dataset is given in Fig.10 with its one way variance table in Table.4 which is used for data interpretation.

Table.3 Accuracy of SVM with different Kernel Functions

Kernel Function Misclassification Error Accuracy

Linear 0.1667 83.33%

Gaussian 0.0667 93.33%

Fig.6: SVM plot of Contrast vs. Energy

Fig.9: Contrast vs. Correlation for Gaussian Kernel

219710-2297430

Fig.10: Box plot of all texture features

Table.4 ANOVA Table for the dataset

Source SS dF MS F Prob>F

Columns 994.72 4 248.68 138.47 1.76516e-48

Error 260.41 145 1.796 Total 1255.13 149 The box plot in Fib.10 shows discrimination of data from the median line and the first feature of contrast appears to have the most discrimination as is also determined from the SVM classification. The value of probability of error given in column 6 of Table.4 shows that the features all combined together also have very low possibility of error and high discrimination.

range. The calculation and comparison of accuracies with Linear and Gaussian kernel functions are indicative of a 93% efficiency of SVM classification with a non-linear hyper plane. Hence the features extracted are accurate and the classification is efficient. The low p-value of 1.76516e-48 from the box plot is indicative of the discriminatory nature of data and thus being capable of accurate prediction of the test data thereby classifying a tumour type of abnormality in tissues. Thus using this method for tumour classification shall reduce possibility of human error and improve machine testing for detection of abnormal tissue.

REFERENCES

1. Suraj S. Gawande, Vrushali Mendre, Brain tumour diagnosis using image processing: A survey, IEEE International Conference On Recent Trends in Electronics Information & Communication Technology, 2017
2. Ghanavati, Sahar, Junning Li, Ting Liu, Paul S. Babyn, Wendy Doda, and George Lampropoulos, Automatic brain tumour detection in Magnetic Resonance Images, IEEE International Symposium on Biomedical Imaging, 2012.
3. WR Hendee, ER Ritenour, Medical Imaging Physics, Fourth Edition, pg 2, 356, 468, 2003
4. El-Sayed A, El-Dahshan, Heba M. Mohsen, Kenneth Revett, Abdel-Badeeh M., Computer-aided diagnosis of human brain tumour through MRI: A survey and a new algorithm, Elsevier Expert Systems with Application, Volume 41, Issue 11, 2014
5. Kidwell C.S., Chalela J.A., Saver J.L., et al., Comparison of MRI and CT for Detection of Acute Intracerebral Haemorrhage, Journal of American Medical Association, 2004
6. Anam Mustaqeem, Ali Javed, Tehseen Fatima, An Efficient Brain Tumour Detection Algorithm Using Watershed & Thresholding Based Segmentation, International Journal of Image, Graphics and Signal Processing, 2012
7. Rohini Paul Joseph, C. Senthil Singh, M.Manikandan , Brain Tumour MRI Image Segmentation and Detection in Image Processing, International Journal of Research in Engineering and Technology, 2014
8. S. Shankar Bharti, Proceedings of the Fourth International Conference on Signal and Image Processing, Springer Nature, Volume 2, pg 212, 2013
9. Abdalla, Alla Sh., Islam A. Yusuf, Sahar Haj Ali A. Mohammed, Meinas A. Mahmoud, and Zeinab A. Mustafa, A Computer-Aided Diagnosis System for Classification of Lung Tumours , Journal of Clinical Engineering, 2015
10. Rao, M. Akshatha, Dharmanna Lamani, Rekha Bhandarkar, and T. C. Manjunath, Automated detection of diabetic retinopathy through image feature extraction, International Conference on Advances in Electronics Computers and Communications, 2014
11. J. Zhou, K. L. Chan, V. F. H. Chong, S. M. Krishnan, Extraction of Brain Tumour from MR Images Using One-Class Support Vector Machine, Engineering in Medicine and Biology 27th Annual Conference, 2013
12. Suthaharan S., Support Vector Machine, Machine Learning Models and Algorithms for Big Data Classification. Integrated Series in Information Systems, vol 36, Springer, Boston, 2016