Browsing by Author "Raini Hassan, Ph.D"

The substantial progress of medical imaging technology in the last decade makes it challenging for medical experts and radiologists to analyse and classify them. Medical images contain massive information that can be used for diagnosis, surgical planning, training, and research. The ability to estimate conclusions without direct human input in healthcare systems using computer algorithms is known as Artificial intelligence (AI) in healthcare. Deep Learning (DL) approaches are already being employed or exploited for healthcare purposes. There is, therefore, a need for a technique that can automatically analyze and classify the images based on their respective contents. DL algorithms open a world of opportunities, and it has been recently used for medical images analysis. Although DL techniques have demonstrated a breakthrough in medical images analysis, research still ongoing to improve the accuracy rate. This research focuses on DL in the context of analysing Magnetic Resonance Imaging (MRI) brain medical images. A comprehensive review of the state-of-the-art processing of brain medical images using DL is conducted in this research. The scope of this research is restricted to three digital databases: (1) the Science Direct database, (2) the IEEEXplore Library of Engineering and Technology Technical Literature, and (3) Scopus database. More than 400 publications were evaluated and discussed in this research. The research focus on both binary classification and multi-class classification. For binary classification, the dataset used is from the brain tumor classification project which contains tumorous and non-tumorous images, and it is available for research and development. For multi-class classification, the dataset contains T1-weighted contrast-enhanced MRI medical images from 233 patients with three types of tumours: meningioma, glioma, and pituitary which is also available for research and development. The proposed neural model is fully automatic brain tumour MRI medical images classification model that uses Convolutional Neural Network (BTMIC-CNN). The model's excellent performance was confirmed using the evaluation metrics and reported a total accuracy of 99%. It outperforms existing methods in terms of classification accuracy and is expected to help radiologists and doctors accurately classify brain tumours’ images. This study contributes to goal 3 of the Sustainable Development Goals (SDGs), which involves excellent health and well-being.

The tremendous growth of data generated daily has made the science of data compression an important and renewable field. It has become the first way to reduce the volume of data to optimize the use of storage units and accelerate the process of transferring data across various types of networks, chiefly the World Wide Web, thus reducing the cost of transport and storage. Compressed data grows with the same frequency as the data itself, which, in turn, created an urgent necessity to understand and analyze the compressed files themselves, and since efforts are focused only on inventing and developing new compression algorithms, few efforts remain trying to understand and analyze compressed files. This research invests in compressed files introducing a new way to analyze and understand compressed data from new angles. This analysis contributes to solutions to practical problems, including the problem of servers in classifying files before actually compressing them with what is known as compressibility. The issue of studying compressibility in systems servers is a sensitive and important issue, given that they provide for the optimum utilization of the physical and programmatic server resources. This research presented a new method by which server systems can distinguish between compressed files from uncompressed files on the one hand, and on the other hand, distinguish between compressed files that need more compression and those that do not need all of this in one frame. Moreover, as the randomness study programs cannot distinguish compressed data from uncompressed data in most cases, this study provided an integrated package of methods for studying the randomness of compressed files called (RTCD). This package can analyze the randomness of compressed files from new practical angles and open the way for the ability to compare compressed files with each other and distinguish between them successfully. This package includes quantitative and graphical measures all set to be standard in practice. The analysis in this study relies on the use of the Fibonacci code as a strong analytical basis capable of knowing the common characteristics of compressed files and can thus distinguish them from uncompressed files successfully. Moreover, the difference in these characteristics within the compressed files circle enables one to know the files that still need more compression. Compared to the well-known techniques that study compressibility and those that study randomness of data, this analysis shows its distinction and its ability to overcome the deficiencies of these methods.

Several studies on physiological-based human emotion have suggested that emotion causes variations in various physiological parameters. As one of the physiological parameters, heart sound signals (also referred to as phonocardiography) may infer emotions and can possibly be used for emotion recognition. For this purpose, the use of Phonocardiography (PCG) signal is substantially cheaper, and the process of acquiring the signal for heart sound analysis is comfortable as compared to other physiological measures. Capturing heart-sound signals using PCG does not require touching the surface of the human body directly. Hence it offers a convenient and practical usage in various applications of emotion recognition. Additionally, unlike the use of electrocardiography (ECG) that reflects only heartbeats through the electrically conductive system of the heart, the PCG can also reflect the muscle contraction sound of the heart. Nevertheless, the use of PCG in the emotion recognition domain is still scarce in the research literature. As such, this thesis explored usability and methods for modelling emotion recognition using PCG signals. The thesis is developed with four major phases. (i) Since PCG data for emotion recognition are not widely available, the first phase performs the creation of the corpus for both PCG and EEG, hence, the performance for both modalities can be compared. (ii) The second phase investigates the most suitable method for building a computational model for PCG-based emotion recognition. Three cepstral-based features, namely, MFCC, LFCC, and GFCC, are considered in the experiment. The DNN, XGBoost, and Decision tree are selected as the classifiers. The initial experiments of this research indicate that the best model for recognizing emotion is achieved at 87% accuracy rate by using combination of MFCC feature extraction and DNN classifier, (iii) The third phase compares PCG-based emotion recognition using heart sound signal (PCG) with EEG modality. The experimental results implied that with techniques used in phase two, the PCG signal could achieve comparatively robust performance in recognizing emotion as compared to the EEG modality. (iv) In the fourth phase, a new computational approach is proposed and implemented by incorporating signal decomposition techniques such as Empirical Mode Decomposition (EMD). As the main issue with this approach is feature dimensionality, the PCA feature reduction technique is adopted in the proposed method. The proposed method demonstrated a robust and optimal performance of a PCG-based emotion recognition model, achieving overall accuracy rate at 98%. Overall, this research has highlighted the potential use of PCG signals for emotion recognition as an alternative to other commonly discussed modalities such as EEG. Additionally, the thesis also empirically proved that with proper methods in pre-processing the signal and the right feature extraction process and the suitable classifier, the PCG signal could achieve optimal performance in recognizing emotion. As future works, the proposed approach can be used to build a wide range of practical application of emotion recognition such as Ambient Assisted Living (AAL), whereby the patient’s mental state is required to be continuously monitored.