A new approach to optical coherence tomography image analysis using machine learning

This proposal describes a novel and adventurous pump-priming (pilot) project. The project fits the BBSRC Call for developing "new approaches to the analysis and interpretation of research data in bioimaging, including the development of software tools and algorithms that address challenges arising from emerging new types of data and known problems associated with data handling".

Specifically, in this pilot project we will develop a revolutionary approach to analyzing optical coherence tomography (OCT) images, that is, to automatically extract properties that may not be visible to the eye from the images using a mathematical technique called inversion, as well as machine learning techniques and knowledge about the physics of OCT imaging.

OCT is a relatively new imaging modality, which provides non-invasive imaging of living tissues based on the principle of optical interferometry. Contrast in OCT is derived from the difference in light scattering and absorption properties of tissue structures. As optical scattering is more varied across soft tissues than either acoustic scattering or x-ray absorption, thus OCT generally provides greater contrast than computed tomography (CT) and ultrasound and can even be sensitive to tissue structural properties at the nanometer length scale. Recent advances in OCT have mainly been driven by applications in biomedical applications such as ophthalmology to study optical nerves development and diagnose glaucoma. These applications rely on the fact that most physiological changes or disease processes affect the optical properties of biological tissue, and this change in tissue optical properties provide the contrast for this imaging technology. Application of OCT is however not limited to the medical domain, e.g., it is also being used in brain imaging, developmental biology, and in tissue engineering.

The motivation of the project is the limitation of current approaches to optical coherence tomography (OCT) image analysis, which are subjective and require that the image features to be detected are visible to the eye. However, most often the vital features are invisible to the eye as biological or disease processes are complex and this complexity is reflected in the images. The project will develop a radically different approach to OCT image analysis that extracts tissue optical properties that are invisible to the eye from OCT images to address life and health science research challenges, where little pilot data exists.

Tissue optical properties are important for a wide range of applications including medical diagnosis, tissue engineering, and developmental biology. As such light scattering spectroscopy has emerged as a tool for characterizing biological tissue and cancer diagnosis by determination of the tissue optical properties. However these optical properties are not visible in optical coherence tomography (OCT) images and conventional image analysis algorithms cannot detect them. There are however established mathematical equations that describe the light transport through tissue / scattering media such as the Radiative Transfer Equation and the the Maxwell's equation. Given the tissue optical properties these equations calculate the OCT signal corresponding to the optical properties. The proposed project aims to develop a new approach to OCT image analysis by estimating the tissue optical properties given the OCT data, which is a inverse problem.

The proposed project will thus consider OCT image analysis as an inverse problem and take inspiration from the advances in solutions for inverse problems using machine learning techniques to learn the mapping/relationship between tissue optical properties and OCT data. With this mapping/relationship tissue optical properties can be predicted from new images. These optical properties can be used for classification of healthy and diseased images. The new OCT image analysis method will be validated using clinical OCT data.

The proposed research is timely in the context that currently OCT image analysis is limited to detecting visible features in the image. OCT hardware technologies have developed significantly over the past decade but software development has yet to catch up. The technology developed in this project will have a wide range of applications, with immediate application in the diagnosis of paediatric glaucoma, which is associated with poor visual outcome.

Apart from the eye conditions, recent research has used OCT image analysis for detecting early Multiple Sclerosis, early Alzheimer's Disease, and Schizophrenia. Thus the technology developed in this project can improves the accuracy of early diagnosis and can potentially be used for neurological and psychiatric diseases, as well as other medical, tissue engineering, and developmental biology studies.

The project will thus have a direct impact on healthcare with the potential of diagnosing life threatening diseases and studying retina and optical nerve development, increasing patient benefit in areas with unmet clinical needs. The project has a clear pathway towards clinical adoption and commercialization due to the involvement of both clinical investigators and industrial partners. The economic and social impact of the proposed research will be significant.