Hyperspectral tissue pre-screening and segmentation for enhanced raman-based oral cancer diagnosis

Abstract

Oro-pharyngeal cancer incidences have been increasing in recent years and late detection means that the prognosis is often poor. In spite of under-availability of trained histopathologists across the United Kingdom, there has been very little clinical translation of automated, in vivo diagnostic devices, in spite of their proven sensitivity. This is potentially due to their requiring widespread change to the established, gold standard diagnostic work˛ow. A device which could be used to support histopathology in the detection of cancer from thin tissue biopsy sections may be more easily adopted. Raman spectroscopy has been identified as a highly specific diagnostic tool but it is an extremely time consuming technique which has prevented simple clinical application and translation. In order to make Raman a realistic diagnostic aid to histopathology, a rapid pre-processing technique is required to identify regions of interest to accelerate and streamline its application. The feasibility of hyperspectral imaging (HSI) for this purpose is investigated in this project. Three systems were built consecutively; a preliminary, an improved and a more economical system, respectively. The preliminary system, utilising a Halogen source, was used for proof of principle with data collected from fluorescent dyes, blood, stained oral tissue, saliva and unstained oral tissue. Using this range of absorbent and non-absorbent, biological and non-biological samples, spectral accuracy and variations with concentration were demonstrated. Additionally principal component analysis (PCA) data denoising and k-means clustering of these hypercubes was shown to be successful, to varying degrees, even in colourless tissue samples, chiefly being sensitive to fibrous tissue and changes in cell density. The improved system garnered data with an increased SNR by utilising a powerful white light laser and with the inclusion of a com-prehensive background removal protocol. Segmentation of malignant tissue with PCA denoising and k-means clustering showed similarities with histopathologist cancer selection on corresponding H&E stained tissue sections. Several alterations to the system and software were made in order to facilitate coregistration, and quantitative comparison, between the three modes; histopathology, hyperspectral imaging and Raman spectroscopy. This ˝nal setup was built to be cost-effective for clinical appeal, with a pulsed Xenon source and fibres replacing a white light laser, a beam expander and a number of lenses. This system covered a wavelength range of 521 - 899 nm (6 nm spectral resolution) and performed with an AUROC (area under receiver operator curve) score of 0.70 on the pixel-wise segmentation of tongue squamous cell carcinoma in frozen, unstained tissue sections compared to consensus histopathology-based diagnosis, with a 2 µm spatial resolution. It per-formed similarly in comparison with Raman spectroscopy, garnering an AUROC score of 0.69, suggesting that their cancer segmentation is reasonably compatible and they could be used to good effect in a combination system. The comparison and evaluation process was vulnerable to tissue sectioning artefacts, which are unavoidable, and random fluctuations in illumination intensity and suggestions to limit the impact of these weaknesses in future work are suggested. The system and quantitative evaluation process requires very little user input and could facilitate easy development of this system, and others. This automation, and the fact that it can be used on thin, unstained tissue sections means that this system has the potential to dovetail neatly with the current histopathology-based protocol. With the suggested improvements, this hybrid HSI-Raman system could provide a second opinion, and potentially evolve to share the load and expedite the oral cancer diagnostic process and improve survival rates.

Date of Award	1 Nov 2021
Original language	English
Awarding Institution	King's College London
Supervisor	Richard Cook (Supervisor) & Susan Cox (Supervisor)