Non-invasive cancer characterisation using autofluorescence imaging
thesisposted on 28.03.2022, 17:38 by Abbas Habibalahi
Auto fluorescence imaging plays a special role in cancer detection, as it is capable of recognising aspects of the chemical composition of tissue based on spectral signatures of naturally fluorescent compounds. Various components naturally present in cells and tissue have auto-fluorescence properties, including Porphyrins (PPIX), Nicotinamide adenine dinucleotide (NADH), and Flavins, whose contents are modified in cancer due to a transformation in cell metabolism. In particular, a lack of iron or ferrochelatase in tumours results in a change of PPIX concentration relative to the normal host tissue. Thus, quantification of these components and studying their variation provides valuable insights into the diagnosis and characterisation of cancer cells and tissue. Conventionally, auto-fluorescence imaging technology has been limited to a few costly channels (n<4) employed in some microscope methodologies such as fluorescence lifetime imaging (FLIM). Consequently, these technologies can monitor only a limited number of fluorophores. However, in this study, a newly designed spectral imaging microscope that employs tens of different channels (n=38) was used. This system uses light excitation with a number of narrow band ranges of wavelength and collects the native fluorescence emission of the sample at specific wavelengths. A combination of excitation /emission wavelength bands forms a spectral channel, and a number of such channels (n=38) were used in this work. The sample is imaged in each of these channels, to acquire separate spectral images. This represents an advance over traditional auto-fluorescence imaging systems. The availability of multiple channels makes it possible to survey the overall biochemical composition of the tissue, in addition to detecting specific markers to identify the tissue state. First, the newly - designed non - invasive auto - fluorescence multispectral imaging methodology was applied to detect Ocular surface squamous neoplasia (OSSN), with a view to a future clinical ophthalmological application. The aim was to distinguish between normal and neoplastic tissue in fixed human samples and sophisticated data analysis was applied to meticulously extract the spectral signature. Two different classification frameworks were deployed, namely intra - and inter - patient classification, to consider aspects of patients' variability and quantify the spectral signature of OSSN. Using machine learning methods, an approach was also introduced for objective assessment of boundary detection. This technique creates a false colour map which can be rapidly generated in quasi - real time and used for intraoperative assessment. The neoplastic boundaries predicted by employing machine learning methods were validated and assessed by an anatomical pathologist. The approach introduced in this study has the potential to reduce the incidence of eye biopsies, prevent therapy delays and make treatment more effective. Using such cutting-edge technology in auto-fluorescence imaging led us to employ a number of channels simultaneously to diagnose or monitor diseases with high accuracy. However, generating a large data set based on tens of spectral images may also increase the possibility of having irrelevant channels which carry very little discriminatory information for a specific diagnostic application. Consequently, such multi spectral imaging needs to be optimised in terms of the number of channels.Different known and unknown factors may influence the usefulness of the channels for a specific application, which cannot be determined by prejudgment. Hence, the best way for channel selection is to employ all of the channels for the detection and then determine which ones are the most relevant channels. In this study, an advanced methodology using a combination of swarm intelligence and cluster analysis was developed to discover rich and informative spectral channels for differentiating normal and diseased (OSSN) tissue.First, discrimination analysis was applied to find normal and diseased clusters and then a criterion function was defined to minimise the within-cluster variance while maximising the between-cluster variance. Such a criterion function was optimised using three different swarm intelligence methodologies including particle swarm intelligence (PSO), differential evolution (DE) and ant colony optimisation (ACO). Finally, depending on the required accuracy and criticality of the application, the richest subsets with a few channels were proposed(5 channels). Optimising the number of channels resulted in more efficient instrumentation in terms of equipment (5 out of 38 channels), acquisition time(80% less acquisition time)andcomputation complexity. Moreover, the metabolic heterogeneity of melanoma cancer cells and fibroblast were considered in this stud y. Multi - spectral auto - fluorescence imaging was employed for evaluation of melanoma cells and fibroblast in order to produce discriminative information. Unlike typical auto - fluorescence imaging techniques that consider only a few features, a variety of bio logically relevant quantitative information was extracted from spectral images. Such a powerful analysis helped capture different aspects of the spectrum in a single cell resolution. Different features including intensity, first order and second order feat ures, textural features, and different statistical measures of pixel value s were quantitatively analysed . After selection of the most indicative features, a discriminative analysis was undertaken to distinguish melanoma cells from fibroblast efficiently. This was then followed by an examination of melanoma cells derived from a patient under treatment. Using unsupervised data processing, spectral features from the channels were de - correlated based on the principal component analysis (PCA). Then the data were quantitatively assessed using hierarchical clustering. Consequently, this study also successfully demonstrates (AUC>0.9) the possibility of obtaining information about melanoma cells and their environment to monitor their behaviour and discriminate them from normal skin cell types. Such a methodology may open a new way for cell therapy, regenerative medicine, personalized immunotherapy and cancer treatment.