Towards Designing Improved Pathological Brain Detection System
using Machine Learning Approaches

Abstract

Brain diseases are recognized to be the most predominant cause of death among individuals with different age groups across the globe. Early detection and treatment are hence of utmost importance for precluding the severity of the brain diseases and improving the patient’s quality life. Magnetic resonance imaging (MRI), a noninvasive 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. The manual diagnosis based on the visual inspection of MRI scans is cumbersome, time-consuming and requires skilled supervision. Thus, it has become imperative to automate the detection process using advanced image processing and machine learning techniques to provide fast, reliable and accurate decisions. The development of automated computer-aided detection/diagnosis (CADe/CAD) systems has made remarkable progress in the past decade. However, there exists a sufficient room for improved automation, applicability and improved accuracy. The existing CADe systems are directed towards solving binary as well as multiclass pathological brain (PB) detection problem and hence, the CADe system is termed, by the research community, as pathological brain detection (PBD) system.
This dissertation focuses on the design of automated PBD systems to assist radiologists
to validate their clinical diagnosis. Our research primarily involves the proposition of different feature extraction methodologies and classifiers to detect pathological brains from MR images.
Two hybrid schemes are presented to classify the brain MR images as normal or pathological. In the first scheme, discrete wavelet transform combined with probabilistic
principal component analysis have been exploited for feature vector computation, whereas, in the second scheme, a stationary wavelet transform based feature descriptor is proposed to represent each MR image. A symmetric uncertainty ranking filter is further applied in the later scheme to derive an optimal set of features. Finally, two ensemble classifiers, namely, Ada boost with random forest and support vector machine are employed separately to obtain the accuracy of the schemes.
To alleviate the bottleneck caused by wavelet features, two feature extraction methodologies have been proposed to efficiently represent the brain MR images. The first methodology considers fast curvelet transform (FCT) to capture multi-directional features from MR images, while in the second methodology, a discrete ripplet-II transform based feature set is developed to capture features along arbitrary directions. A set of new hybrid classifiers incorporating extreme learning machine and contemporary meta-heuristic techniques, are designed in isolation for classifying pathological brains from normal brains.
Multiclass classification of pathological brain MR images is comparatively more
challenging and thus, has become an active area of research in the domain of biomedical image processing. Two multiclass PB detection frameworks are developed to classify the MR images into five categories. A texture based feature descriptor is proposed using curvelet transform to extract salient features from MR images. A kernel extension of random vector functional link network (KRVFL) is suggested to perform multiclass classification which on the other hand eliminates the need for selection of common hyperparameters and improves the generalization performance at faster learning speed. The second framework, along with feature extraction and classification, instigates an angle modulated sine cosine algorithm for optimal feature selection.


The problem of manual choice of suitable features and classifiers for the task of multiclass PB detection is addressed in the subsequent chapter with the application of deep learning algorithms. A fast and efficient autoencoder based deep neural network model is designed to learn high-level features directly from the input images. This research also focuses on the development of a deep convolutional neural network model to classify multiclass pathological brains. The proposed deep architectures facilitate end-to-end learning and help in achieving a promising performance.
Extensive set of experiments have been carried out separately using binary as well as
multiclass brain MR datasets to evaluate the effectiveness of each proposed framework.
Different performance measures have been utilized to compare the proposed schemes with the standard existing methods. Experimental results confirm the superiority of proposed approaches over existing schemes for both binary and multiclass PB detection task. The deep learning based schemes are found to be the most accurate and efficient models for the detection of pathological brain and therefore, these models can be utilized as the supportive tools by physicians to verify their screening.