Abstract
Automated detection of brain abnormalities through magnetic resonance imaging (MRI) has made a significant stride in the past decade. The feature extractors exploited in the literature suffer from issues like limited directional selectivity and high dimensionality, and the classifiers used have critical drawbacks like slow learning speed, poor computational scalability, and trivial human intervention. The fast curvelet transform (FCT) and ripplet-II transform (R2T) provides improved discriminant ability and high directional selectivity. Extreme learning machine (ELM), a randomized learning algorithm for single layer feed-forward neural network, has received significant attention as it provides good generalization performance at much faster speed. In this chapter, the authors compare the effectiveness of two feature extractors based on FCT and R2T along with different ELM algorithms. These schemes have been evaluated on three brain MR datasets and comparative analyses have been made on several combinations of methods. Finally, the potential of the best scheme is compared to the state of the art.
TopIntroduction
Brain disorders are one of the principal sources of death and are very often seen in people of different age groups across all corners of the earth. As estimated by the National Brain Tumor Foundation (NBTF) in the United States, brain tumors are the cause of one-fourth of all cancer deaths in children (El-Dahshan et al., 2014). The World Health Organization (WHO) in 2014 estimated that approximately 2, 50,000 individuals were diagnosed with primary brain tumors per year across the world. Brain diseases are broadly categorized into four classes: cerebrovascular diseases (stroke), neoplastic diseases (brain tumor), infectious diseases, and degenerative diseases (Gudigar et al., 2019a). These diseases cause serious issues and may sometimes prompt to death. Moreover, the prevalence of some of the brain diseases is increased with age. Therefore, early treatment is indispensable to not only prevent the severity of the diseases but also improve the patient's quality life.
Magnetic resonance imaging (MRI), a non-invasive neuroimaging technique, has been profoundly used in diagnosing brain and nervous system abnormalities because of its ability to provide high resolution of soft brain tissues (Shi et al., 2018). In practice, the brain abnormalities are diagnosed manually by the radiologists and the manual diagnosis at a larger scale based on the visual inspection of MRI is tedious, time-consuming, and very expensive. It needs the expertise from neurologists, physicians, and radiologists. Besides, manual interpretation by clinicians suffers from inter- and intra-reader variability and may be non-repeatable. Thus, there is a strong demand to design an automated diagnosis system with the help of image processing and machine learning techniques to accurately detect the brain abnormalities.
The development of automated computer-aided detection/diagnosis (CAD) systems for screening brain abnormalities has made remarkable progress in the past decade with the exploitation and design of various feature descriptors and statistical classifiers (Wang et al., 2016; El-Dahshan et al., 2014). A summary of the existing related works is presented in Table 1.
Table 1. A summary of the related works on brain abnormalities detection
| Author | Feature | Classifier | Number of features | Images |
|
Chaplot et al., 2006
| Discrete wavelet transform (DWT) | SVM | 4761 | 52 |
|
Zhang et al., 2010
| DWT and principal component analysis (PCA) | FNN + ACPSO | 19 | 160 |
|
El-Dahshan et al., 2010
| DWT + PCA | KNN, FNN | 7 | 70 |
|
Zhang et al., 2011
| DWT + PCA | FNN + SCG,
FNN + SCABC | 19 | 66 |
| Zhang et al. 2015b | Wavelet packet Shannon
entropy (WPT-SE) | Generalized eigenvalue proximal SVM (GEPSVM) | 16 | 66, 160 and 255 |
| Zhang et al., 2015a | Wavelet entropy (WE) and Hu moment invariants (HMI) | Generalized eigenvalue proximal SVM (GEPSVM) | 14 | 66, 160 and 255 |
|
Wang et al., 2015
| Fractional Fourier entropy
(FRFE) and Welch’s t-test | Twin SVM | 12 | 66, 160 and 255 |
| Zhang et al., 2015b | Stationary wavelet transform
(SWT) and PCA | GEPSVM | 7 | 66, 160 and 255 |
|
Zhou et al., 2015
| Wavelet entropy (WE) | Naive Bayes | 7 | 64 |
|
Zhang et al., 2016
| Minkowski-Bouligand
dimension features | FNN + improved PSO | 5 | 255 |
|
Nayak et al., 2016
| DWT and probabilistic PCA | AdaBoost with random forest | 13 | 66, 160 and 255 |
|
Zhang et al., 2018
| Pseudo Zernike moment | Kernel SVM | 200 | 66, 160 and 255 |
|
Gudigar et al., 2019a
| Shearlet transform + Texture | SVM + PSO | 15 | 612 |
ACPSO: adaptive chaotic PSO, FNN: feed-forward neural network, SCG: scaled conjugate gradient
Key Terms in this Chapter
Magnetic Resonance Imaging (MRI): It is a noninvasive medical imaging technique which has been extensively used to detect brain related disorders.
Discrete Ripplet-ii Transform (DR2T): It has the ability to extract features along multiple scales and arbitrary directions.
Extreme Learning Machine (ELM): It is a simple but effective learning technique for single-hidden layer feed-forward neural networks that offers better generalization performance at faster training speed.
Fast Curvelet Transform (FCT): It has the ability to derive features along multiple scales and directions.