THz TECHNOLOGY AND MACHINE LEARNING FOR ENHANCEMENT OF OPTICAL COHERENCE TOMOGRAPHY IMAGERY

Optical Coherence Tomography (OCT) was discovered in 1991 utilizing low-coherence light to capture micrometre resolution which typically operates in the near-infrared light region. Ophthalmology has heavily relied on OCT for retinal scans in the medical field due to its high resolution. Despite great resolution, non-invasive and lower costs, OCT’s main disadvantage is the introduction of noise artefacts produced in images due to the low coherence light used. Noise artefacts are the main reason for the minimal application of research in dentistry. Therefore, there is a large knowledge gap regarding denoising, and tooth decay detection in OCT images using Machine Learning (ML) effectively. Thus, this project aims to provide a software and hardware solution to improve OCT images to aid in the early detection of tooth decay and carries. Starting with the denoising, detection and deconvolution of OCT images. For hardware solutions different ways are explored for utilising the inherent long-coherence nature of THz source to build a cleaner image. Regarding denoising, the proposed framework revolves around transformers integrated into an autoencoder (TEAR). This is to remove different types of noise artefacts in the regions of interest (ROI) provided by the attention gates (AG). With a new loss function along with TEAR, which combines a sliding box, contrast-to-noise ratio (CNR), peak signal-to-noise ratio (PSNR) and mean squared error (MSE). It compares the CNR and MSE between predicted and denoised paired images, to evaluate the focused regions the AGs choose. The proposed method improved the PSNR by 27.9 dB, CNR by 6.3 dB, and ENL by 120.8 dB for a dental dataset. Focusing on speckle reduction, gated-filtering super-resolution enhancer (GFSRE) with a backbone of denoising CNN (DnCNN). DnCNN was critically evaluated and modified to produce high-quality resolution enhancement by promoting an adaptive region- based enhancer and hard thresholding through attention gating. MKSF block was integrated into the DnCNN to ensure the removal of the different types of noise artefacts in the ROIs provided by the AGs. The proposed method improved lateral and axial PSF by 17%. Using denoised images lesion detection is overcome by the framework Transformer Enhanced Detection (TED). TED is designed to detect the different types of anomalies commonly present in OCT images for diagnostic purposes, aiding clinical interpretation. The proposed method improved the accuracy of detection by 16-22% and Intersection over Union (IOU) by 10%, in two dentistry datasets. Moving on to hardware solutions, THz imaging via a Vector Network Analyzer (VNA) was compared to OCT imaging for dentistry applications. Using a VNA, the tooth phantom is placed in an imaging setup and S21, S11 analysis was performed to simulate a B- scan to mimic an OCT image. A dispersive complex signal amplitude was measured at successive A-scan points (as for the OCT). In the configuration used, VNA resolved to 0.05 mm lateral resolution and could locate the region, width, and depth of decay