Optical coherence tomography and machine learning applied to biological samples

Abstract

Optical coherence tomography (OCT) is a non-invasive imaging technique which can acquire twoand three-dimensional images from samples with micrometer resolution. The developments in OCT over the past thirty years have enhanced the potential of OCT to become the diagnostic and quality assessment standard in other fields than just ophthalmology. The structural and functional parameters from OCT data have the potential to determine the properties of various biological samples. However the transition of OCT technology from research labs to industry and/or clinical environment requires innovative solutions that can analyse OCT data in real time with minimal to null error. My thesis introduces a new approach to increase the efficiency of OCT in studying biological samples using machine learning. Machine learning is an important outcome of artificial intelligence that helps to find underlying relationship among datasets and draw inferences without the need of any explicit programming. The work of the thesis is inspired by real world problems which requires modern day solutions. The first part of the thesis focuses on a meat industry problems that have remained unsolved for decades – the need for a technique which can measures the intramuscular fat content (IMF) and tenderness in meat in real time without destroying the sample. The thesis identifies optical properties– attenuation and birefringence, which can be used to study IMF content and tenderness in meat. Furthermore we demonstrate how machine learning techniques can be implemented to derive information from a relatively small OCT dataset in a completely automated way to quantify the IMF content and tenderness in meat samples in near real time. The adaptability and robustness of combining OCT with machine learning are further demonstrated by studying another sample – human skin. The research focuses on diagnosing the occurrence of a rare autoimmune disease called systemic scleroderma (SSc) in human skin. The approach used is similar to the method used in meat study, proving the adaptability of our method to other biological samples. Attenuation of light was the parameter used to identify changes in skin when infected by SSc. Our approach of using machine learning to make automated diagnosis was tested on the dataset and an effective classification model to differentiate normal and SSc infected skin was developed, proving the reliability of the method. Our study also provides an alternative to the traditional approach of segmenting skin layers to identify differences in normal and diseased skin. The thesis demonstrates how machine learning can improve the impact of OCT and make the diagnostic performance of OCT more universal with minimal human errors during diagnosis.