Magnetic Resonance Imaging (MRI) is one of the important resources for identifying abnormalities in the human brain. This work proposes an effective Multi-Class Classification (MCC) system using Binary Robust Invariant Scalable Keypoints (BRISK) as texture descriptors for effective classification. At first, the potential Region Of Interests (ROIs) are detected using features from the accelerated segment test algorithm. Then, non-maxima suppression is employed in scale space based on the information in the ROIs. The discriminating power of BRISK is examined using three machine learning classifiers such as
The growth of abnormal cells in the brain is referred as brain tumour or brain cancer. Some tumours are cancerous and others are non-cancerous. Magnetic Resonance Imaging (MRI) images are often used in medical domain for the diagnosis of various diseases. Several machine learning algorithms such as Naive Bayes (NB), Decision Tree (DT), Gaussian, and Support Vector Machine (SVM) along with radial basis function kernel are reviewed in [
A brain tumor investigative system is built in [
Automatic brain tumour detection is analyzed in [
The whole image around the skull is used for the classification in [ Developing an efficient Binary Robust Invariant Scalable Keypoints (BRISK) descriptor tool that describes the key points in the MRI brain image to point out the affected region. Additional specific image features are extracted and utilized to further improve the system’s performances. Applying multi-class classifier to categorize the images into normal or one of the abnormalities such as glioma, meningioma, and pituitary.
The main objective is to develop machine learning algorithms such as
In this section, the proposed MCC system for classifying the MRI images into four classes such as normal, glioma, meningioma, pituitary using machine learning classifiers is discussed. The workflow of the proposed MCC system is depicted in
Feature extraction is a method utilized to recognize or extract key parameters from input images as initial information to obtain the novel data. In this work, key pointer detection method is applied to detect the features and also locate the extracted features via key points. Especially, BRISK algorithm is utilized which is an attribute point recognition as well as depiction approach with scalable invariance along with turning round. Moreover, this detector contest attributes among two images via tuning the parameters up to 200 values. The group of key points comprises of points cultures image positions linked with floating point scaled principles. The BRISK descriptor is collected as two-fold string through adding the outcomes of effortless intensity of image comparison trials. This intensity comparison helps to enhance the image descriptiveness.
For every undersized couple, it obtains smoothened intensity of sampling points and then finds whether the first pixel of smoothened intensity is greater than second pixel. If the first pixel is greater than second pixel then the descriptor is considered as one or otherwise 0. Hence this descriptor is suitable for feature extraction on images even in very short pairs.
The image features extracted from MRI brain images are image area, equivalent diameter, orientation, minor & minor axis, perimeter, minimum intensity, maximum intensity level and mean intensity level. The features which are extracted from the MRI brain images are described in
Features | Description |
---|---|
Area | The image area estimated using formula |
Equivalent diameter | Defines the image diameter with equivalent sectional area Estimated using formula |
Orientation | Orientation defines the image rotation in both clockwise and counter clockwise position. |
Minor axis | Represents vertical axis |
Major axis | Represents horizontal axis |
Perimeter | Sum up all the side length of tumour image |
Minimum intensity | Signifies darker intensity values at every position |
Mean intensity | Calculates the overall image intensities by adding intensities of every pixel in image. |
Maximum intensity | Signifies lighter intensity values at every position |
In this sub-section, three different machine learning algorithms which are used to classify the input MRI brain images into different classes of tumours are discussed.
The other distance metrics used in
All learning algorithms incorporate with both training and testing stages. Hence we have to follow these steps to perform KNN algorithm Calculate the appropriate distance metric. During training stage, in accordance with extracted features, accumulate the training images as a duo ( Estimate the distance among novel feature vector of an image as well as all training image too during testing stage. Also, this algorithm breeds the nearest data pixels to the unlabeled portion on the images. Finally based on voting the classification of images has done.
SVM is one of the supervised machine learning approaches to perform classification task on images. This method has frequently established to improve the classification results when compared to other pattern identification approaches. The SVM based systems are extremely smart for distinguishing different patterns in the data’s or images. Moreover, SVM method try to establish finest isolating hyper plane among several classes by finding support vectors which are located at the boundary of every hyper plane [ Class 1: Class 2:
Here the two classes such as class 1 and class 2 are separated using hyper plane parameters defined by a vector
To identify the hyper plane, v and v0 will be evaluated in such a manner that
The objective of hyper plane is to put down the highest boundary among classes. The vectors related with SVM should mention to detect the optimal hype plane which categorizes the classes. The support vectors reclined on two hyper planes that are matching to best possible one [
After rescaling of hyper plane parameters
Subject to
Here S refers to the split in training images which points non-zero Lagrangian multipliers. Such training images are known as support vector. The cost function which helps to amalgamate the highest boundary as well as lowest error by means of special variables named as
SVM plots the input image represented in vectors (x) into highly dimensional image attributes and in this case, finest isolation of hyper plane is created in that space. Such plotting generates additional complications to the dilemma. The representation of inner product function is defined in
But K (x, y) represents kernel function. The binary optimization issues can be created as in
And finally the consequential classifier becomes
For object based image investigation, multi label classification is used in [
To work with non-linear data, kernel trick is employed. Different kernels such as quadratic, Radial Basis Function (RBF) and polynomial are used in this work with the standard linear kernel. Their definitions are as follows:
where
The automated segmentation as well as classification of brain stroke images is discovered using RF classifier that attains highest accuracy in [
This algorithm is working better while the input dataset is bulky. The forest is highly vigorous while larger quantity of decision trees utilized during decision production procedure. Initially the image dataset is split into two halves with analogous formation. The edge function using RF algorithm is defined as,
Here I refer to Indication function h1 (X), h2(X), …, hn(X) are classifiers with ensembling method where X and Y are random vectors. The function that denotes error which is expressed in
Now the edge function equation for RF is rewritten in
Here
The RF method is a time-saving approach that combines numerous decision trees into a single tree for the best prediction accuracy. The step by step procedures for RF algorithm are as follows: A novel image is constructed from original given brain MRI image by means of sampling as well as reducing the 1/3 rd portion of row images. Now the algorithm is trained to produce novel images from sample reduction and also evaluates balanced error. At every pixel of the image first column is chosen from total number of columns available. Many decision trees develop concurrently and then ultimate output is predicted through gathering of every decision to attain better accuracy classification.
The performances of the proposed MCC system for classifying MRI brain images are assessed using a public database. It is freely downloadable from [
Tumour types | Training samples | Testing samples | Total samples |
---|---|---|---|
Glioma tumour | 826 | 100 | 926 |
Meningioma tumour | 822 | 115 | 937 |
Pituitary tumour | 827 | 74 | 901 |
Normal | 395 | 105 | 500 |
Total | 2870 | 394 | 3264 |
In this study, the performance of the MCC system is estimated by means of performance metrics such as accuracy, sensitivity and specificity. These measures are obtained by forming confusion matrix from the outcomes of the classifiers such as True Positive (TP), False Positive (FP), True Negative (TN) and False Negative (FN).
Predicted value (PV) | |||||
---|---|---|---|---|---|
Classes | Normal (N) | Glioma (G) | Meningioma (M) | Pituitary | |
Actual value (AV) | Normal | ||||
Glioma | |||||
Meningioma | |||||
Pituitary | |||||
where |
Parameters | Description | For normal class |
---|---|---|
TP | The AV and PV should be same | |
FP | sum of values in the corresponding column except the TP. | |
FN | sum of values in the corresponding rows except the TP | |
TN | Sum of all values except TP, FN and TN |
After computing the parameters such as TP, FP, TN and FN, the following
Three machine learning algorithms such as
PV | Parameters | ||||||||
---|---|---|---|---|---|---|---|---|---|
Classes | N | G | M | P | TP | FP | FN | TN | |
AV | N | 92 | 6 | 3 | 4 | 92 | 15 | 13 | 274 |
G | 4 | 91 | 2 | 3 | 91 | 16 | 9 | 278 | |
M | 6 | 6 | 100 | 3 | 100 | 9 | 15 | 270 | |
P | 5 | 4 | 4 | 61 | 61 | 10 | 13 | 310 |
It can be seen from
PV | Parameters | ||||||||
---|---|---|---|---|---|---|---|---|---|
Classes | N | G | M | P | TP | FP | FN | TN | |
AV | N | 100 | 2 | 1 | 2 | 100 | 5 | 5 | 284 |
G | 1 | 97 | 0 | 2 | 97 | 6 | 3 | 288 | |
M | 2 | 3 | 109 | 1 | 109 | 3 | 6 | 276 | |
P | 2 | 1 | 2 | 69 | 69 | 5 | 5 | 315 |
It can be seen from
Predicted value (PV) | Parameters | ||||||||
---|---|---|---|---|---|---|---|---|---|
Classes | N | G | M | P | TP | FP | FN | TN | |
Actual value (AV) | N | 105 | 0 | 0 | 0 | 105 | 1 | 0 | 288 |
G | 1 | 98 | 0 | 1 | 98 | 1 | 2 | 293 | |
M | 0 | 0 | 115 | 0 | 115 | 0 | 0 | 279 | |
P | 0 | 1 | 0 | 73 | 73 | 1 | 1 | 319 |
It can be seen from
This work proposes an efficient BRISK descriptor based MCC system for the diagnosis of brain cancer. It uses MRI images for the classification of three different brain cancers such as glioma, meningioma and pituitary along with normal class. The image based features such as image area, orientation, equivalent, perimeter, diameter, minor and major axis, min, max mean intensity are also extracted and added to BRISK descriptor to increase the MCC system performances. Three machine learning classifiers;