Data-driven microwave imaging and processing techniques for monitoring neurodegenerative diseases

Abstract

Neurodegenerative diseases, such as Alzheimer’s disease (AD), affect cognitive functions and behaviour due to the progressive loss of neurons in the human brain. Unfortunately, although medicines can help relieve some of the symptoms, there is currently no cure for AD. Therefore, it is crucial to detect AD at earlier stages, to stop or slow its progression. Currently, computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) are being used clinically to diagnose AD. These imaging modalities can provide information about malignant brain tissue’s shape, location, and size. However, they have many limitations. For instance, CT scans involve a higher dosage of ionising radiation and are slower and more complex.Conversely, MRI generates high-quality images with a good spatial resolution. The downside of MRI includes high equipment cost, system complexity, and longer time required for data collection. Therefore, there is a need for wearable, non-invasive and portable imaging devices that can monitor the progression of these diseases. It will allow patients to be treated more conveniently at their homes.

In recent years, microwave imaging systems (MIS) have been widely investigated for medical imaging such as breast cancer and brain stroke due to their potential to provide low-cost, portable and non-ionising imaging devices. The detection and monitoring of neurodegenerative diseases is a recent application of microwave imaging. A complete MIS consists of hardware and software components of equal importance for accurate image reconstructions. The hardware components include the wearable device, Vector Network Analyzer (VNA), and a personal computer (PC). The software components, which are the topic of this thesis, include signal processing and imaging techniques. The collected microwave signals are processed to generate images showing the object(s) under test. Numerous processing and imaging algorithms have been proposed for breast tumour and brain stroke detection in recent years. These processing and imaging techniques give promising results; however, they have several shortcomings that needs to be addressed before they can be used to detect or monitor small changes in the brain caused by neurodegenerative diseases. This thesis investigates the potential of data-driven signal preprocessing and microwave imaging techniques to detect and monitor neurodegenerative diseases, especially AD.

The first stage of the study investigate and evaluate the existing confocal microwave imaging technique. Various algorithms originally used for brain stroke and breast tumour detection are implemented and considered to detect the AD changes in the brain. An integrated microwave image reconstruction (MIR) algorithm is proposed based on a modified form of Microwave Imaging via Space-time (MIST). Novel preprocessing techniques for clutter removal are integrated with the proposed imaging algorithm. The algorithm is validated on real lamb brain phantoms that imitate whole-brain atrophy. The results show that the proposed algorithm can detect and locate whole-brain atrophy with reasonable accuracy. The second stage of the study is the development of a frequency-based imaging algorithm that solves multiple issues with time-domain radar-based algorithms such as signal transformation, multi-speed and multipath. A frequency-based multistatic imaging algorithm is proposed that integrates each component’s frequency response to form an image. The algorithm compensates for both the amplitude and phase of the scattered signals. It resolves dispersion and attenuation issues and therefore outperforms the timeshift based algorithms. The algorithm is evaluated on both simulation and experimental data obtained from brain phantoms that mimic the atrophy and CSF changes. Quantitative analysis and comparison with conventional radar-based algorithms show that the proposed imaging algorithm can detect both brain atrophy and changes in the thickness of CSF layers more accurately.

The third stage of the study explores how to optimise the image reconstruction algorithms. Multistatic imaging algorithms such as multistatic MIST and delay multiply and sum (DMAS) cannot produce images in an acceptable time due to intensive computation and sequential execution. A novel distributed approach based on Apache Spark, which adopts both data and algorithm in parallel, is proposed for efficient image reconstruction. The proposed parallel approach could be used to accelerate any computationally intensive algorithm on a cluster having master-slave architecture.The last stage of the study investigates ML-based methods to aid in classifying and monitoring different stages of neurodegenerative disease based on physiological and pathological changes in the brain. A novel data augmentation method is proposed to generate synthetic data for ML algorithms. Two feature extractions strategies are stipulated and compared. Multiple ML techniques are trained using raw data, manually and automatically extracted features to validate how well each algorithm could classify the AD stages. The results indicated that the proposed ML-based method could be used to monitor AD at its early stages.