Abstract

Endoscopic imaging, an essential subfield of biomedical imaging, is the focus of the present study. While X-Ray, Ultrasound, and MRI have been used widely, Optical coherence tomography (OCT) has recently received much interest owing to its potential use in non-destructive tissue imaging and industrial component testing. OCT is a tomographic technique that produces cross-sectional images of objects with a resolution of 2 to 10 µm.
Recently, Liverpool university developed a Line Field OCT system to improve scanning speed while reducing scanning distortion errors and motion abnormalities. However, contemporary OCT employ refractive optics for scanning and traditional spectrometers for data analysis. The fundamental shortcoming of refractive optics is chromatic aberration, particularly in OCT, where a broadband light source is utilized. Also, traditional spectrometers have nonlinearity in k-space, which reduces the signal sensitivity.
This thesis aims to design a reflective optics-based line scan (LS-OCT) with cylindrical optics where 2D cross-sectional imaging data can be obtained without requiring a mechanical scanner. Chromatic aberration is eliminated with the use of reflective optics. Further, a novel linear k-space spectrometer has been developed as a part of this thesis to reduce the signal sensitivity drop-off. The scanner and spectrometers design include an analytical study with MATLAB and optical modeling with ZEMAX.
The design is optimized for wavelength range of 830 ± 100 nm. The scanning system is designed to provide a scan range of 2  2  2 mm, and the designed scanner is 30% smaller than a similar design in literature, while providing higher image quality within the scan range. Multiple linear k-space spectrometers are designed and analyzed as part of this work. The optimization is performed to maximize linearity and image quality while keeping the size of the spectrometer minimum.
Finally, on the data analysis aspect of the thesis, texture identification approach based on the Deep Recurrent Neural Network (DRNN) model is presented. For this purpose, different geometrical defects are 3D printed and imaged with an OCT. From the images, training is performed for defect identification. The performance of the various training approaches with different datasets for texture recognition is assessed, a two-layer LSTM network obtained the best outcome (97 % accuracy).