Pattern recognition and tomographic reconstruction with Terahertz Signals for applications in biomedical engineering.

Abstract

Over the last ten years, terahertz (THz or T-ray) biomedical imaging has become a modality of interest due to its ability to simultaneously acquire both image and spectral information. Terahertz imaging systems are being commercialized, with increasing trials performed in a biomedical setting. Advanced digital image processing algorithms are greatly need to assist screening, diagnosis, and treatment. Pattern recognition algorithms play a critical role in the accurate and automatic process of detecting abnormalities when applied to biomedical imaging. This goal requires classification of meaningful physical contrast and identification of information in images, for example, distinguishing between different biological tissues or materials. T-ray tomographic imaging and detection technology contributes especially to our ability to discriminate opaque objects with clear boundaries and makes possible significant potential applications in both in vivo and ex vivo environments. The Thesis consists of a number of Chapters, which can be grouped in to three parts. The first part provides a review of the state-of-the-art regarding THz sources and detectors, THz imaging modes, and THz imaging analysis. Pattern recognition forms the second part of this Thesis, which is represented via combining several basic operations: wavelet transforms and wavelet based signal filtering, feature extraction and selection, along with classification schemes for THz applications. Signal filtering in this Thesis is achieved via wavelet based de-noising. The ultrafast pulses generated terahertz time-domain spectroscopy (THz-TDS), which is demonstrated to justify their decomposition in the wavelet domain as it can provide better de-noising performance. Feature extraction and selection of the terahertz measurements rely on observed changes in pulse amplitude and phase, as well as scattering characteristics of several different types of powder samples under study. Additionally, three signal processing algorithms are adopted for the evaluation of the complex insertion loss function of such samples as lactose, mandelic acid, and dl-mandelic acid: (i) standard evaluation by ratioing the sample with the background spectra, (ii) a subspace identification algorithm, and (iii) a novel wavelet packet identification procedure. These system identification algorithms enable THz measurements to be transformed to features for THz pattern recognition. Meanwhile, a novel feature extraction method involving the use of Auto Regressive (AR) and Auto Regressive Moving Average (ARMA)models on the wavelet transforms of measured T-ray pulse responses of ex vivo osteosarcoma cells as well as other biomedical materials is presented. Classification schemes are carried out via simple and robust schemes, such as the linear Mahalanobis distance classifier, and the non-linear Support Vector Machine (SVM) classifier. In particular, SVMs are used as a learning scheme to achieve the identification of two classes of RNA samples and multiple classes of powered materials. Coherent terahertz detection hardware—THz time-domain spectroscopy (THz-TDS)—is used to obtain all the data for validation of these classification schemes. The past decade has witnessed the tremendous development of terahertz instruments for detecting, storing, analysing, and displaying images. Terahertz time-domain spectroscopy (THz-TDS) is a broadband technique that generates and detects THz radiation in a synchronous and coherent manner. By contrast, the newly developed THz quantum cascade laser is a narrow-band radiation source that provides potential for realising compact systems; they produce image data with higher average power levels. The third part of this Thesis discusses methods to improve the capability of both broad and narrow-band terahertz imaging, driven by computer-aided analytical techniques. A wavelet based reconstruction algorithm for terahertz computed tomography is represented to show how this algorithm can be used to rapidly reconstruct the region of interest (ROI) with a reduction in the measurements of terahertz responses, compared with a standard filtered back-projection technique. These reconstruction algorithms are applied to the analysis of acquired experimental data and to locally recover the two dimensional (2D) and three-dimensional (3D) structures of several optically opaque objects. Moreover, a segmentation technique based on two dimensional wavelet transforms is investigated for the identification of different materials from the reconstructed CT image.