Method for differentiating between histopathological causes of kidney graft dysfunction

Abstract

A method and system for assessing a histopathological cause of kidney graft dysfunction in a subject following kidney transplantation is disclosed. The method comprises the steps of: a) generating one or more images of the native fluorescence emission from one or more exfoliated kidney cells obtained from a urine sample from the subject in a plurality of distinctive spectral channels, b) calculating, for each urinary exfoliated kidney cell, quantitative features of the autofluorescence signals in said one or more images, c) applying a feature selection methodology to the quantitative features to generate a set of optimised quantitative features; and d) applying a classification method to the optimised quantitative features to produce a multispectral profile for differentiating the histopathological cause of kidney graft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.

Description

METHOD FOR DIFFERENTIATING BETWEEN HISTOPATHOLOGICAL CAUSES OF KIDNEY GRAFT DYSFUNCTIONFIELD OF THE INVENTION

[0001] The present invention relates to multispectral or hyperspectral image analysis of urinary exfoliated kidney cells for use in differentiating between histopathological causes of kidney graft dysfunction in a subject following kidney transplantation.BACKGROUND OF THE INVENTION

[0002] The following discussion of the prior art is provided to place the invention in an appropriate technical context and enable the advantages of it to be more fully understood. It should be appreciated, however, that any discussion of the prior art throughout the specification should not be considered as an express or implied admission that such prior art is widely known or forms part of the common general knowledge in the field.

[0003] Chronic Kidney Disease (CKD) is a progressive condition which may potentially result in kidney failure, where patients require kidney replacement therapy in the form of dialysis or kidney transplantation. Whilst kidney transplantation is the definitive and often preferred option of treatment for kidney failure, optimal outcomes following kidney transplantation may be challenged due to various complications that can be defined histologically. Complications such as acute tubular necrosis (ATN) and acute rejection can lead to delayed graft function in the acute phase following transplantation whilst in the longer term, the transplanted kidney may also experience progressive dysfunction due to progressive kidney interstitial fibrosis and tubular atrophy (IFTA) from a variety of etiologies not limited to chronic allograft nephropathy, calcineurin inhibitor (CNI) toxicity, chronic rejection, progressive kidney fibrosis in the absence of immunological challenge or recurrence of rejection episode(s).

[0004] Current epidemiological data suggest delayed, or deteriorating graft function occurs in between 20 and 50% of kidney transplant cases. It is particularly important to positively identify the histopathology of those who have delayed or deteriorating graft function, as the treatment strategy for the various causes of graft dysfunction would be different. There is no validated method apart from biopsy which can definitively predict the cause of delayed or deteriorating graft function at present. Transplant biopsy isinvasive, requires hospitalization, is time-consuming and is costly. Hence, there is an unmet critical need for non-invasive diagnosis to identify kidney transplant complications.

[0005] The proximal tubule cells (PTCs) make up over 50% of the kidney mass. Dysfunction of PTCs is related to the pathogenesis of kidney transplant dysfunction. This suggests that examination of PTCs can have potential diagnostic and prognostic value in patients with delayed or deteriorating graft function. Kidney tubules are continuously exposed to glomerular filtrate, and thousands of living PTCs are excreted daily in the urine together with other cells shed from different parts of the nephron, ureters, bladder and urethra. Viable exfoliated PTCs can be isolated from urinary sediment.

[0006] To date, there remains no sensitive method to determine if exfoliated PTCs can be used to diagnose and differentiate between patients with different histopathological causes of graft dysfunction following kidney transplantation.

[0007] Cell autofluorescence is derived from native fluorophores such as collagen, elastin, tryptophan, reduced nicotinamide adenine dinucleotide (phosphate) (NAD(P)H), and flavins, which play a pivotal role in cell and tissue structure, as well as cell and tissue metabolism. This native autofluorescence can be collected using multispectral fluorescence microscopy which captures images at various emission and excitation wavelength ranges (the pair of which is referred to as a “channel”,) yielding spectrally- dependent quantitative features for each individual cell, including parameters such as average channel intensities, average channel intensity ratios, pixel standard deviations and skewness. Such data provides each cell with a unique signature, which originates from its intracellular biochemistry and organization. Past studies have noted that these cellular characteristics may be highly biologically informative and can distinguish biological conditions, including cell cycle stage, cellular inflammation, level of reactive oxygen species, presence of neoplasia and cell aging through assessment of cell autofluorescence. As these are all key pathophysiological factors to consider in kidney disease, the inventors evaluated whether multispectral autofluorescence features in urinary exfoliated PTCs can be applied to differentiate between individuals with different causes of kidney transplant dysfunction.

[0008] There is a clear need for the development of simple, reliable and accurate methods for differentiating between histopathological causes of kidney graft dysfunction in a subject following kidney transplantation, in particular at an early stage.

[0009] It is an object of the present invention to overcome or ameliorate one or more the disadvantages of the prior art, or at least to provide a useful alternative.SUMMARY OF THE INVENTION

[0010] Cell autofluorescence originates from native fluorophores such as collagen, elastin, tryptophan and reduced nicotinamide adenine dinucleotide (phosphate) (NAD(P)H), and flavins, which play a pivotal role in cell and tissue metabolism. Hyperspectral or multispectral fluorescence microscopy can be used to collect native emission data across a broad range of excitation wavelength ranges. This data gives a per cell, spectral profile of autofluorescent features. This gives each cell a unique signature, arising directly from its intracellular biochemistry and organisation, and can be used for the sensitive discrimination of specific cellular characteristics including cell cycle stage, the presence of inflammatory disease, levels of reactive oxygen species, neoplasia and aging. Hyperspectral or multispectral assessment of cell autofluorescence has been specifically shown to be sensitive to metabolic changes and oxidative stress.

[0011] The inventors have surprisingly found that using autofluorescence multispectral or hyperspectral imaging they can determine the “multispectral profile” which characterizes urinary exfoliated kidney proximal tubule cells obtained from patients.

[0012] Patients experiencing complications such as acute tubular necrosis (ATN) and graft rejection (GR), or experiencing progressive dysfunction due to non-rejection associated interstitial fibrosis and tubular (NR-IFTA), can be accurately and reliably determined from this multispectral profile, and that it is possible to use this method to differentiate between histopathological causes of kidney graft dysfunction in a subject (i.e. , recipient) following kidney transplantation.

[0013] According to a first aspect, the present invention provides a method of assessing a histopathological cause of kidney graft dysfunction in a subject following kidney transplantation, the method comprising the steps of: a) generating one or more images of the native fluorescence emission from one or more exfoliated kidney cells obtained from a urine sample from the subject in a plurality of distinctive spectral channels, b) calculating, for each urinary exfoliated kidney cell, quantitative features of the autofluorescence signals in said one or more images, c) applying a feature selection methodology to the quantitative features to generate a set of optimised quantitative features; and d) applying a classification method to the optimised quantitative features to produce a multispectral profile for differentiating the histopathological cause of kidneygraft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.

[0014] In some embodiments, the classification method is a binary classification into one of two possible classes of histopathological causes of kidney graft dysfunction selected from the group consisting of: a) acute tubular necrosis (ATN) versus graft rejection (GR); b) acute tubular necrosis (ATN) versus non-graft rejection interstitial fibrosis and tubular atrophy (NR-IFTA); and c) graft rejection (GR) versus non-graft rejection interstitial fibrosis and tubular atrophy (NR-IFTA).

[0015] In some embodiments, the the distinctive spectral channels are defined by a central excitation wavelength of between 340 nm-510 nm and an emission wavelength of between 370 nm-900 nm, with a spectral width of each spectral channel in the range of several tens of nanometers.

[0016] In some embodiments, the distinctive spectral channels are selected from:

[0017] In some embodiments, the method further comprises in relation to step b) one or more of the following steps of: b1 ) performing image pre-processing; and b2) removing correlations between the calculated quantitative features of different urinary exfoliated kidney cells.

[0018] In one embodiment, when the feature selection methodology is an entropybased feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprising:

[0019] the ratio of mean value of channel 8 and mean value of top 10% of channel 2, the ratio of mean value of top 10% of channel 31 and mean value of top 10% of channel 32, the ratio of mean value of top 10% of channel 24 and mean value of top 10% of channel 28, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 23, the ratio of mean value of channel 22 to mean value of top 40% of channel 12, the ratio of mean value of channel 27 to mean value of top 40% of channel 5, the ratio of mean value of channel 21 to mean value of top 40% of channel 25, and the ratio of mean value of channel 24 to mean value of top 40% of channel 7.

[0020] In one embodiment, when the feature selection methodology is an entropybased feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprising:

[0021] the ratio of mean value of channel 8 and mean value of top 10% of channel 2, the ratio of mean value of top 10% of channel 31 and mean value of top 10% of channel 32, the ratio of mean value of top 10% of channel 24 and mean value of top 10% of channel 28, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 23, the ratio of mean value of channel 22 to mean value of top 40% of channel 12, the ratio of mean value of channel 27 to mean value of top 40% of channel 5, the ratio of mean value of channel 21 to mean value of top 40% of channel 25, and the ratio of mean value of channel 24 to mean value of top 40% of channel 7.

[0022] In one embodiment, when the feature selection methodology is an entropybased feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA, comprising:

[0023] the ratio of mean value of top 10% of channel 24 and mean value of top 10% of channel 28, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 23, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 21 , the ratio of mean value of channel 5 to mean value of top 40% of channel 4, the ratio of mean value of channel 22 to mean value of top 40% of channel 12, the ratio of mean value of channel 21 to mean value of top 40% of channel 3, the ratio of mean value of channel 22 to mean value of top 40% of channel 15, and the ratio of mean value of channel 21 to mean value of top 40% of channel 25.

[0024] In one embodiment, when the feature selection methodology is an entropybased feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA, comprising:

[0025] the ratio of mean value of top 10% of channel 2 and mean value of top 10% of channel 5, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 27, the ratio of mean value of channel 18 and mean value of top 10% of channel 3, the ratio of mean value of top 10% of channel 5 and mean value of top 10% of channel 23, the ratio of mean value of top 10% of channel 31 and mean value of top 10% of channel 18, the ratio of mean value of channel 8 to mean value of top 40% of channel 21 , the ratio of mean value of channel 5 to mean value of top 40% of channel 4, and the ratio of mean value of channel 29 to mean value of top 40% of channel 2.

[0026] In some embodiments, the multispectral profile comprises no more than 10 optimised quantitative features when selected using a feature based selection methodology selected from the group consisting of i) a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology; ii) a particle swarm optimization feature selection methodology, iii) a Gini-index-based feature selection methodology, iv) an entropy-based feature selection methodology, v) a chi-square-based feature selection methodology, and vi) a forward feature selection methodology.

[0027] In one of these embodiments, when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learningclassifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise:

[0028] the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 16 to mean value of channel 18, the ratio of mean value of top 10% of channel 22 to mean value of channel 18, the ratio of mean value of channel 15 to mean value of top 40% of channel 6, the ratio of mean value of top 10% of channel 20 to mean value of channel 19, the ratio of mean value of channel 14 to mean value of channel 13, the ratio of mean value of top 10% of channel 34 to mean value of top 10% of channel 31 , the ratio of mean value of top 10% of channel 12 to mean value of channel 16, the ratio of mean value of channel 8 to mean value of top 40% of channel 24, and the ratio of mean value of channel 10 to mean value of channel 9.

[0029] In another of these embodiments, when the feature selection methodology is a forward feature based selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise:

[0030] the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 22 to mean value of top 10% of channel 28, the ratio of mean value of channel 6 to mean value of top 10% of channel 8, the ratio of mean value of channel 2 to mean value of top 10% of channel 34, the ratio of mean value of top 10% of channel 10 to mean value of top 10% of channel 17, the ratio of mean value of channel 5 to mean value of channel 29, the ratio of mean value of channel 16 to mean value of top 10% of channel 12, the ratio of mean value of top 10% of channel 16 to mean value of channel 17, the ratio of mean value of channel 12 to mean value of channel 9, and the ratio of mean value of top 10% of channel 15 to mean value of channel 27.

[0031] In another of these embodiments, when the feature selection methodology is a forward feature based selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA comprise:

[0032] the minimum intensity value of channel 2, the ratio of mean value of channel 16 to mean value of top 40% of channel 24, the ratio of mean value of top 10% of channel 16 to mean value of top channel 18, the ratio of mean value of channel 9 to mean valueof channel 10, the ratio of mean value of channel 14 to mean value of channel 32, the ratio of mean value of channel 24 to mean value of channel 17, the ratio of mean value of channel 14 to mean value of top 40% of channel 20, the minimum intensity of channel 13, the ratio of mean value of top 10% of channel 27 to mean value of top 10% of channel 16, and the ratio of mean value of channel 26 to mean value of top 10% of channel 17.

[0033] In another of these embodiments, when the feature selection methodology is a particle swarm optimization based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise:

[0034] the ratio of mean value of top 10% of channel 3 to mean value of channel 4, the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of channel 17 to mean value of channel 18, the ratio of mean value of channel 17 to mean value of top 40% of channel 3, the ratio of mean value of top 10% of channel 16 to mean value of channel 15, the ratio of mean value of channel 24 to mean value of top 40% of channel 4, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel 21 , the ratio of mean value of channel 17 to mean value of channel 6, the ratio of mean value of channel 8 to mean value of top 40% of channel 24, and the ratio of mean value of channel 6 to mean value of channel 4.

[0035] In another of these embodiments, when the feature selection methodology is a Gini-index based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise:

[0036] the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 17, the ratio of mean value of channel 21 to mean value of top 10% of channel 19, the ratio of mean value of top 10% of channel 21 to mean value of channel 17, the ratio of mean value of channel 4 to mean value of channel 27, the ratio of mean value of top 10% of channel 15 to mean value of top 10% of channel 19, the ratio of mean value of channel 4 to mean value of top 40% of channel 31 , the ratio of mean value of top 10% of channel 16 to mean value of top 10% of channel 15, the ratio of mean value of channel 5 to mean value of channel 19, and the ratio of mean value of channel 5 to mean value of channel 33.

[0037] In another of these embodiments, when the feature selection methodology is a chi-square based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise:

[0038] the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of channel 25 to mean value of channel 26, the ratio of mean value of top 10% of channel 15 to mean value of channel 17, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel 21 , the ratio of mean value of top 10% of channel 18 to mean value of channel 19, the ratio of mean value of channel 24 to mean value of top 40% of channel 4, the ratio of mean value of channel 5 to mean value of channel 19, the ratio of mean value of top 10% of channel 19 to mean value of channel 26, the ratio of mean value of channel 4 to mean value of channel 27, and the ratio of mean value of channel 17 to mean value of channel 18.

[0039] In another of these embodiments, when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise:

[0040] the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of channel 25 to mean value of top 10% of channel 22, the ratio of mean value of channel 22 to mean value of top 10% of channel 17, the ratio of mean value of channel 26 to mean value of channel 12, 25thpercentile value of channel 20, the ratio of mean value of top 10% of channel 3 to mean value of top 10% of channel 15, the ratio of mean value of channel 14 to mean value of channel 13, the ratio of mean value of channel 16 to mean value of top 40% of channel 3, the ratio of mean value of top 10% of channel 6 to mean value of top 10% of channel 15, and the ratio of mean value of top 10% of channel 22 to mean value of channel 33.

[0041] In another of these embodiments, when the feature selection methodology is a particle swarm optimization feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise:

[0042] the ratio of mean value of channel 23 to mean value of top 10% of channel 15, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel 21 , the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 16 to mean value of channel 15, the ratio of mean value of top 10% of channel 28 to mean value of top 10% of channel 6, the ratio of mean value of top 10% of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 19 to mean value of top 10% of channel 21 , the ratio of mean value of top 10% of channel 19 to mean value of top 10% of channel 22, and the ratio of mean value of channel 33 to mean value of top 10% of channel 2.

[0043] In another of these embodiments, when the feature selection methodology is a Gini-index based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise:

[0044] the ratio of mean value of top 10% of channel 15 to mean value of channel 27, the ratio of mean value of top 10% of channel 12 to mean value of top 10% of channel 19, the ratio of mean value of channel 2 to mean value of channel 24, the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 31 , the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of channel 12 to mean value of channel 19, the ratio of mean value of top 10% of channel 16 to mean value of channel 12, the ratio of mean value of channel 7 to mean value of channel 17, and the ratio of mean value of channel 8 to mean value of top 10% of channel 22.

[0045] In another of these embodiments, when the feature selection methodology is an entropy-based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise:

[0046] the ratio of mean value of channel 8 to mean value of top 10% of channel 33, the ratio of mean value of top 10% of channel 9 to mean value of channel 33, the ratio of mean value of channel 32 to mean value of top 40% of channel 17, the ratio of mean value of channel 7 to mean value of top 10% of channel 33, the ratio of mean value of top 10% of channel 16 to mean value of channel 19, the ratio of mean value of channel33 to mean value of top 40% of channel 17, the ratio of top 10% of mean value of channel 8 to mean value of channel 33, the ratio of mean value of top 10% of channel 12 to mean value of top 10% of channel 30, and the ratio of mean value of top 10% of channel 22 to mean value of channel 33.

[0047] In another of these embodiments, when the feature selection methodology is a chi-square based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise:

[0048] the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 34, the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of top 10% of channel 18 to mean value of top 10% of channel 17, the ratio of mean value of top 10% of channel 27 to mean value of channel 7, the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of channel 2 to mean value of channel 24, the ratio of mean value of channel 33 to mean value of channel 15, the ratio of mean value of top 10% of channel 21 to mean value of channel 32, and the ratio of mean value of top 10% of channel 4 to mean value of top 10% of channel 32.

[0049] In another of these embodiments, when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA comprise:

[0050] the ratio of mean value of channel 26 to mean value of top 10% of channel 22, the ratio of mean value of channel 28 to mean value of channel 13, the ratio of mean value of channel 28 to mean value of channel 18, the ratio of mean value of channel 28 to mean value of top 40% of channel 29, the ratio of mean value of top 10% of channel 20 to mean value of channel 12, the ratio of mean value of channel 27 to mean value of channel 17, the ratio of mean value of channel 22 to mean value of top 40% of channel 26, the ratio of mean value of top 10% of channel 3 to mean value of channel 27, the ratio of mean value of channel 24 to mean value of channel 17, and the ratio of mean value of channel 9 to mean value of channel 10.

[0051] According to another preferred embodiment, the method comprises the step of further processing one or more of the multispectral profiles obtained above by:

[0052] e) applying a second classification method to the optimised quantitative features.

[0053] In one embodiment, the second classification method is an auto machinelearning (AutoML) classifier.

[0054] In a preferred embodiment, the auto machine-learning (AutoML) classifier employed in the present invention is the open-source AutoML package, AutoGluon. Tabular (version: 0.8.3b20231023) from Amazon Web Services.

[0055] In a preferred embodiment, said urinary exfoliated kidney cells are proximal tubule cells (PTCs).According to a second aspect, the present invention provides a system for assessing a histopathological cause of kidney graft dysfunction in a subject following kidney transplantation, the system comprising: a multispectral excitation lamp configured to excite the single photon-excited autofluorescence signal of one or more exfoliated kidney cells obtained from a urine sample from the subject in a plurality of defined narrowband excitation wavelength ranges, one or more epifluorescence filters, a detector configured to detect autofluorescence of said one or more urinary exfoliated kidney cells at multiple specified wavelengths and a processing system configured to a) generate one or more images of the native fluorescence emission from the one or more urinary exfoliated kidney cells, b) calculate, for each urinary exfoliated kidney cell, quantitative features of the autofluorescence signals in said one or more images, c) apply a feature selection methodology to the quantitative features to generate a set of optimised quantitative features; and d) apply a classification method to the optimised quantitative features to produce a multispectral profile for differentiating the histopathological cause of kidney graft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.

[0056] In one embodiment, the processing system is further configured to: perform image pre-processing; calculate, for each kidney cell, quantitative features of the measured autofluorescence; remove correlations between the calculated quantitative features of different kidney cells.

[0057] Other aspects of the invention are also disclosed in the following.DEFINITIONS

[0058] In describing and claiming the present invention, the following terminology will be used in accordance with the definitions set out below. It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments of the invention only and is not intended to be limiting.

[0059] A multispectral image is an image that captures image data within specific wavelength ranges in the electromagnetic spectrum. Typically, though not necessarily, the one or more multispectral images are obtained by multispectral microscopy. In other examples, the one or more multispectral images are obtained by hyperspectral microscopy. In yet other examples, the one or more multispectral images are obtained by any other multispectral or hyperspectral imaging method.

[0060] Hyperspectral imaging uses wavelength and spatial image information for detection and classification. In one example, fluorescence images of live cells or tissues are obtained at a number of selected excitation wavelength ranges (referred to here as excitation channels), capturing their emission at multiple specified wavelength ranges (referred to as emission channels). By using optical microscopy techniques, the autofluorescence of endogenous cellular fluorophores can be observed at a single cell level providing insights into cell activities without altering them with exogenous labels.

[0061] The terms “multispectral” and “hyperspectral” are used interchangeably here. The term “multispectral” generally refers to cases where the number of excitation or emission channels is, for example, 10 or less. The term “hyperspectral” generally refers to cases where the number of excitation or emission channels is, for example, in the order of 100 or more. Typically, mathematical analysis of data acquired through multispectral or hyperspectral microscopy is identical. In this document these two terms are used interchangeably without losing generality. The presented results are for tens of channels which can be equally well described as “hyperspectral” or “multispectral”.

[0062] We additionally clarify that a “spectral channel” in fluorescence imaging is given by defining two spectral ranges, an excitation wavelength range [ex min, AeX max] and an emission wavelength range [^em,min> ^em.max] , while in reflectance imaging used in remote sensing a spectral channel is defined by a single (reflectance) wavelength range. This document is exclusively concerned with fluorescence imaging. Each spectral channel discussed here is characterised by a central (emission or excitation) wavelength. Centralexcitation wavelength is the average of Aexminandexmax. Central emission wavelength is an averagearid

[0063] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one having ordinary skill in the art to which the invention pertains.

[0064] Unless the context clearly requires otherwise, throughout the description and the claims, the terms “comprise”, “'comprising”, and the like are to be construed in an inclusive sense as opposed to an exclusive or exhaustive sense; that is to say, in the sense of “including, but not limited to”. For example, a composition, mixture, process or method that comprises a list of elements is not necessarily limited to only those elements but may include other elements not expressly listed or inherent to such composition, mixture, process or method.

[0065] The transitional phrase "consisting of” excludes any element, step, or ingredient not specified. If in the claim, such would close the claim to the inclusion of materials other than those recited except for impurities ordinarily associated therewith. When the phrase "consisting of' appears in a clause of the body of a claim, rather than immediately following the preamble, it limits only the element set forth in that clause; other elements are not excluded from the claim as a whole.

[0066] The transitional phrase "consisting essentially of" is used to define a composition, process or method that includes materials, steps, features, components, or elements, in addition to those literally disclosed, provided that these additional materials, steps, features, components, or elements do not materially affect the basic and novel characteristic(s) of the claimed invention. The term "consisting essentially of' occupies a middle ground between "comprising" and "consisting of".

[0067] Where the applicant has defined an invention or a portion thereof with an open- ended term such as "comprising", it should be readily understood that (unless otherwise stated) the description should be interpreted to also describe such an invention using the terms "consisting essentially of' or "consisting of." In other words, with respect to the terms “comprising”, “consisting of”, and “consisting essentially of”, where one of these three terms is used herein, the presently disclosed and claimed subject matter may include the use of either of the other two terms. Thus, in some embodiments not otherwise explicitly recited, any instance of “comprising” may be replaced by “consisting of” or, alternatively, by “consisting essentially of”.

[0068] While reference may be made in this disclosure to the invention comprising a combination of a plurality of elements, it is also understood that this invention is regarded to comprise combinations which omit or exclude one or more of such elements, even if this omission or exclusion of an element or elements is not expressly stated herein, unless it is expressly stated herein that an element is essential to the applicant' s combination and cannot be omitted. It is further understood that the related prior art may include elements from which this invention may be distinguished by negative claim limitations, even without any express statement of such negative limitations herein. It is to be understood, between the positive statements of applicant's invention expressly stated herein, and the prior art and knowledge of the prior art by those of ordinary skill which is incorporated herein even if not expressly reproduced here for reasons of economy, that any and all such negative claim limitations supported by the prior art are also considered to be within the scope of this disclosure and its associated claims, even absent any express statement herein about any particular negative claim limitations.

[0069] As used herein, with reference to numbers in a range of numerals, the terms "about," "approximately" and "substantially" are understood to refer to the range of -10% to +10% of the referenced number, preferably -5% to +5% of the referenced number, more preferably -1 % to + 1 % of the referenced number, most preferably -0 .1 % to +0 .1 % of the referenced number. Moreover, with reference to numerical ranges, these terms should be construed as providing support for a claim directed to any number or subset of numbers in that range. For example, a disclosure of from 1 to 10 should be construed as supporting a range of from 1 to 8, from 3 to 7, from 1 to 9, from 3.6 to 4.6, from 3.5 to 9.9, from 8 to 10, and so forth.

[0070] The terms “preferred” and “preferably” refer to embodiments of the invention that may afford certain benefits, under certain circumstances. However, other embodiments may also be preferred, under the same or other circumstances. Furthermore, the recitation of one or more preferred embodiments does not imply that other embodiments are not useful and is not intended to exclude other embodiments from the scope of the invention.

[0071] The complete disclosures of the patents, patent documents and publications cited herein are incorporated by reference in their entirety as if each were individually incorporated.

[0072] Unless expressly stated to the contrary, "or" refers to an inclusive or and not to an exclusive or. For example, a condition A or B is satisfied by any one of the following: A is true (or present), and B is false (or not present), A is false (or not present), and B is true (or present), and both A and B are true (or present).

[0073] The term "and / or" used in the context of "X and / or Y" should be interpreted as "X," or "Y," or "X and Y." Similarly, "at least one of X or Y" should be interpreted as "X," or 'Y," or "both X and Y."

[0074] The indefinite articles "a" and "an" preceding an element or component of the invention are intended to be non-restrictive regarding the number of instances (i.e., occurrences) of the element or component. Therefore "a" or "an" should be read to include one or at least one, and the singular word form of the element or component also includes the plural unless the number is obviously meant to be singular.

[0075] It will be understood that use of the term “between” herein when referring to a range of numerical values encompasses the numerical values at each endpoint of the range. For example, a temperature of between 80 °C and 150 °C is inclusive of a temperature of 80 °C and a temperature of 150 °C.

[0076] Various features of the embodiments of the invention disclosed herein are, for brevity, described in the context of a single embodiment, but may also be provided separately or in any suitable sub-combination. All combinations of the embodiments are specifically embraced by the illustrative embodiments disclosed herein just as if each and every combination was individually and explicitly disclosed. In addition, all subcombinations listed in the embodiments describing such variables are also specifically embraced by the present compositions and are disclosed herein just as if each and every such sub-combination was individually and explicitly disclosed herein.

[0077] In the foregoing paragraphs, various ratios of components have been disclosed. It will be appreciated that these ratios of components can be combined in any disclosed combination. For example, the ratio of A:B (which may be between about 100:1 and 1 :100 or any range therein), may be combined with the ratio of C:D (which may be between about 50:1 and 1 :50 or any range therein), and may be combined with the ratio of E:F (which may be between about 10:1 and about 1 :10 or any range therein).

[0078] As used herein the term “plurality” means more than one. In certain specific aspects or embodiments, a plurality may mean 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 , 12, 13, 14, 15, 16, 17, 18, 19, 20, 21 , 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 , 32, 33, 34, 35, 36, 37,38, 39, 40, 41 , 42, 43, 44, 45, 46, 47, 48, 49, 50, 51 , or more, and any integer derivable therein, and any range derivable therein.BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The aspects described above, as well as other apparent aspects, advantages, and objectives of the present invention are apparent from the detailed description below in combination with the drawing, in which:

[0080] Figure 1 shows a workflow flowchart summarizing key aspects of the study methodology. Patients are grouped into those with Acute tubular necrosis (ATN), graft rejection, or non-rejection Interstitial fibrosis and tubular atrophy (NR-IFTA) transplant complications following histopathological evaluation of the processed transplant biopsy sample by accreditation-qualified pathologists at NSW Health Pathology Laboratory;

[0081] Figure 2 shows brightfield and hyperspectral (HS) channel images from representative urinary exfoliated PTCs across the defined study groups;

[0082] Figure 3 shows multispectral differentiation of exfoliated PTCs between study participants with ATN or graft rejection (GR) in the form of: receiver operating characteristic (ROC) curves for obtained the cell classifier using the optimal spectral feature combination selected by using: (a) a random forest classifier according to method(i), and (b) a linear support vector machine-learning (SVM) classifier according to method(ii); and receiver operating characteristic (ROC) curves obtained using the AutoGluon classifier framework for comparative analysis according to method (iii), with the same feature combination which was selected by: (c) method (i), and (d) method (ii). The dashed diagonal line in the panels in Figures 3(a), (c) and (d) is the performance of a noninformative classifier area (AUC = 0.5). area under the (ROC) curve (AUC);

[0083] Figure 4 shows multispectral differentiation of exfoliated PTCs between study participants with ATN or NR-IFTA in the form of: receiver operating characteristic (ROC) curves for the obtained cell classifier using the optimal spectral feature combination selected by using: (a) a random forest classifier according to method (i), and (b) a linear support vector machine-learning (SVM) classifier according to method (ii); and receiver operating characteristic (ROC) curves obtained using the AutoGluon classifier framework for comparative analysis according to method (iii), with the same feature combination which was selected by: (c) method (i), and (d) method (ii). The dashed diagonal line in the panels in Figures 4(a), (c) and (d) is the performance of a noninformative classifier area (AUC = 0.5). area under the (ROC) curve (AUC);

[0084] Figure 5 shows multispectral differentiation of exfoliated PTCs between study participants with graft rejection or NR-IFTA in the form of receiver operating characteristic (ROC) curves for the obtained cell classifier using the optimal spectral feature combination selected by using: (a) a random forest classifier according to method (i), and (b) a linear support vector machine-learning (SVM) classifier according to method (ii); and receiver operating characteristic (ROC) curves obtained using the AutoGluon classifier framework for comparative analysis according to method (iii), with the same feature combination which was selected by: (c) method (i), and (d) method (ii). The dashed diagonal line in the panels in Figures 5(a), (c) and (d) is the performance of a noninformative classifier area (AUC = 0.5). area under the (ROC) curve (AUC);

[0085] Figure 6 shows (a) the receiver operating characteristic (ROC) curve and (b) confusion matrix obtained from a patient-to-patient level classification performance assessment, which reveals significant differences in exfoliated proximal tubule cells (PTCs) multispectral cellular features between study participants in the ATN vs Graft Rejection group, when classified using the AutoGluon classifier framework (with 5-fold cross-validation), trained using the top 3 coefficients of each principal component determined from the feature data obtained from Table 5. The dashed diagonal line in the panel in Figure 6(a) is the performance of a noninformative classifier area (AUC = 0.5). area under the (ROC) curve (AUC);

[0086] Figure 7 shows (a) the receiver operating characteristic (ROC) curve and (b) confusion matrix obtained from a patient-to-patient level classification performance assessment, which reveals significant differences in exfoliated proximal tubule cells (PTCs) multispectral cellular features between study participants in the ATN vs NR-IFTA group, when classified using the AutoGluon classifier framework (with 5-fold cross- validation), trained using the top 3 coefficients of each principal component determined from the feature data obtained from Table 6. The dashed diagonal line in the panel in Figure 7(a) is the performance of a noninformative classifier area (AUC = 0.5). area under the (ROC) curve (AUC);

[0087] Figure 8 shows (a) the receiver operating characteristic (ROC) curve and (b) confusion matrix obtained from a patient-to-patient level classification performance assessment, which reveals significant differences in exfoliated proximal tubule cells (PTCs) multispectral cellular features between study participants in the Graft Rejection (GR) vs NR-IFTA group, when classified using the AutoGluon classifier framework (with 5-fold cross-validation), trained using the top 3 coefficients of each principal componentdetermined from the feature data obtained from Table 6. The dashed diagonal line in the panel in Figure 8(a) is the performance of a noninformative classifier area (AUC = 0.5). area under the (ROC) curve (AUC); and

[0088] Figure 9 shows the receiver operating characteristic (ROC) curves obtained from a patient-to-patient level classification performance assessment, which reveals significant differences in exfoliated proximal tubule cells (PTCs) multispectral cellular features between study participants in a N = 19 cohort study in the (a) ATN vs GR group, (b) the ATN vs NR-IFTA group, and (c) the GR vs NR-IFTA group; all when classified using the AutoGluon classifier framework (with 5-fold cross-validation) and the top 8 ranked multispectral autofluorescence features determined from the feature data obtained from Table 11. The dashed diagonal line in the panels in Figures 9(a) to 9(c) is the performance of a noninformative classifier area (AUC = 0.5). area under the (ROC) curve (AUC).DETAILED DESCRIPTION

[0089] The skilled addressee will understand that the invention comprises the embodiments and features disclosed herein as well as all combinations and / or permutations of the disclosed embodiments and features.

[0090] Complications relating to delayed or deteriorating graft function following kidney transplantation are common. There is no validated method apart from transplant kidney biopsy which can accurately identify between the histopathological causes of graft dysfunction. Considering an unmet critical need for non-invasive diagnosis to reliably diagnose kidney transplant complications, the inventors have identified a novel methodology based on assessment of multispectral autofluorescence features of extracted urinary exfoliated proximal tubule cells (PTCs) from kidney transplant recipients.

[0091] The present invention is predicated in part on the inventors’ surprising discovery that multispectral / hyperspectral autofluorescence imaging of kidney cells exfoliated into urine can be used to detect, diagnose and classify stages of kidney graft dysfunction in a subject following kidney transplantation.

[0092] Accordingly, the present invention provides, for the first time, a reliable, accurate urine-based diagnostic test for detecting kidney graft dysfunction following kidney transplantation. Thus, methods described herein enable the possibility of detecting and diagnosing kidney graft dysfunction prior to the onset or manifestation of clinicalsymptoms. Accordingly, therapeutic intervention can be commenced in a pre-clinical setting which can lead to improved patient prognosis, maintaining or extending quality of life, or in delaying or preventing the onset of clinical symptoms associated with delayed or deteriorating kidney graft function following transplantation.

[0093] The skilled person will appreciate that the present invention may provide one or more significant advantages and improvements in the field and / or in view of the prior art. For example, these advantages include:• The ability to diagnose and monitor kidney transplant complications non- invasively.

[0094] What follows is a description of preferred embodiments of the present invention.

[0095] Differentiating between histopathological causes of kidney graft dysfunction in a subject following kidney transplantation

[0096] The present invention provides a method of assessing a histopathological cause of kidney graft dysfunction in a subject following kidney transplantation according to a preferred embodiment of the present invention.

[0097] Method

[0098] In its simplest form, the method comprises the steps of: a) generating one or more images of the native fluorescence emission from one or more exfoliated kidney cells obtained from a urine sample from the subject in a plurality of distinctive spectral channels, b) calculating, for each urinary exfoliated kidney cell, quantitative features of the autofluorescence signals in said one or more images, c) applying a feature selection methodology to the quantitative features to generate a set of optimised quantitative features; and d) applying a classification method to the optimised quantitative features to produce a multispectral profile for differentiating the histopathological cause of kidney graft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.

[0099] In a preferred embodiment, the histopathological causes of kidney graft dysfunction relate to kidney graft dysfunction as determined by pathologicalmeasurement of one or more of: acute tubular necrosis (ATN), graft rejection (GR), and non-rejection associated interstitial fibrosis and tubular atrophy (NR-IFTA).[000100] As is outlined in the Materials and Methods section below, live, viable kidney cells, more particularly, exfoliated proximal tubule cells (PTCs), can be extracted from a urine sample provided by the subject by immuno-magnetic separation. Once extracted, the exfoliated PTCs are resuspended in PBS buffer and then analyzed by multispectral autofluorescence imaging using the apparatus described in detail below.[000101] A number of techniques well known to those skilled in the art may be used for isolation of live cells. As exemplified herein, the urine cells may be isolated from urine by antibody selection. In a preferred embodiment, renal proximal tubular cells are extracted from the thawed urine cells (which were stored frozen at -80°C) using immune- magnetic separation, anti-CD13 and anti-Sodium-glucose linked transporter-2 (SGLT2) antibodies, and then thawed prior to use.[000102] Preferably, urine cells are collected using centrifugation and then washed twice with phosphate buffered-saline (PBS). Cells may then be stored at -80°C in a suitable medium with a cryopreservative such as either 5% or 10% Dimethyl Sulfoxide (DMSO). Preferably, the cryopreservation medium is PBS and 5% DMSO. Renal proximal tubule cells can be extracted following incubation with anti-CD13 antibody and anti-SGLT2 antibody followed by magnetic bead-based separation then assessed using multispectral and brightfield microscopy.[000103] System[000104] Measurement of the autofluorescence signals is carried out using a system comprising: a multispectral excitation lamp configured to excite the single photon -excited autofluorescence signal of one or more kidney cells from the subject in a number of defined narrowband excitation wavelength ranges, one or more epifluorescence filters, a detector configured to detect autofluorescence of said one or more urinary exfoliated kidney cells at multiple specified wavelengths.[000105] The system further comprises a processing system that is configured to a) generate one or more images of the native fluorescence emission from the one or more urinary exfoliated kidney cells, b) calculate, for each urinary exfoliated kidney cell, quantitative features of the autofluorescence signals in said one or more images, c) apply a feature selection methodology to the quantitative features to generate a set of optimised quantitative features; and d) apply a classification method to the optimised quantitativefeatures to produce a multispectral profile for differentiating the histopathological cause of kidney graft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.[000106] Autofluorescence Detection[000107] According to step a) of the method described herein, images of the native fluorescence emission produced by one or more urinary exfoliated PTCs obtained from the subject being assessed are generated in a plurality of distinctive spectral channels.[000108] These distinctive spectral channels are selected from those listed in Table 3.[000109] The cellular fluorescence emitted from the urinary exfoliated PTCs is generated by one or more endogenous cellular fluorophores, which may include, for example, nicotinamide dinucleotides such as nicotinamide adenine dinucleotide (NADH) and nicotinamide adenine dinucleotide phosphate (NADPH), flavins such as flavin adenine dinucleotide (FAD) and flavin mononucleotide (FMN), porphyrins, elastin, collagen, tryptophan and pyridoxine.[000110] The multispectral or hyperspectral imaging analysis may be sensitive to or may detect or measure the content of some of these fluorophores in the at least one cell or in one or more subcellular compartments of the cell.[000111] In one embodiment, the step of generating one or more multispectral images includes the steps of stimulating the kidney cell by irradiation with electromagnetic radiation having one or more excitation wavelengths and detecting autofluorescence of the at least one cell in an emission wavelength. In other words, there is generated one or more images of the native fluorescence emission of one or more of said kidney cells in a plurality of distinctive spectral channels defined by a central excitation wavelength of between 340 nm-510 nm and a central emission wavelength of between 390 nm-900 nm and spectral with in the order of tens of nanometres. The step of generating one or more multispectral images is typically repeated for each spectral channel.[000112] Once generated, the autofluorescence signals in these multispectral images are then measured and analyzed according to step d) of the method described herein to obtain a multispectral profile for the kidney cells of the subject being assessed.[000113] Multispectral Profiles[000114] As used herein, the term “multispectral profile" is intended to refer to the outcome of multispectral imaging followed by the further steps of image pre-processingand then calculation, for each cell, quantitative features of the measured autofluorescence image of these cells - steps performed by the processing system, as described in more detail below.[000115] The method of obtaining the multispectral profile may also involve additional steps to prepare the spectral images for quantitative analysis. For instance, the method may comprise image preprocessing steps including image smoothing to reduce the impact of Poisson’s noise and dead or saturated pixels, calibration to eliminate background autofluorescence and / or adjust for uneven illumination of the field of view and removing correlations using Principal Component Analysis (PCA) between the calculated quantitative features of different urinary exfoliated kidney cells.[000116] The method may also comprise the step of projecting, for each urinary exfoliated kidney cell, the quantitative features of the measured autofluorescence onto a new vector space produced by Linear Discriminant Analysis (LDA), which generates scatter plots that can then be compared between cell groups. These additional steps are also performed by the processing system.[000117] In some embodiments of the methods described herein, the multispectral profile comprises no more than eight (8) optimised quantitative features.[000118] For instance, and with reference to the spectral channels outlined in Table 4, when the feature selection methodology in one embodiment is an entropy-based feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprising: (i) the ratio of mean value of channel 8 and mean value of top 10% of channel 2, the ratio of mean value of top 10% of channel 31 and mean value of top 10% of channel 32, the ratio of mean value of top 10% of channel 24 and mean value of top 10% of channel 28, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 23, the ratio of mean value of channel 22 to mean value of top 40% of channel 12, the ratio of mean value of channel 27 to mean value of top 40% of channel 5, the ratio of mean value of channel 21 to mean value of top 40% of channel 25, and the ratio of mean value of channel 24 to mean value of top 40% of channel 7.[000119] Again, with reference to the spectral channels outlined in Table 4, when the feature selection methodology in another embodiment is an entropy-based featureselection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA, comprising (ii) the ratio of mean value of top 10% of channel 24 and mean value of top 10% of channel 28, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 23, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 21 , the ratio of mean value of channel 5 to mean value of top 40% of channel 4, the ratio of mean value of channel 22 to mean value of top 40% of channel 12, the ratio of mean value of channel 21 to mean value of top 40% of channel 3, the ratio of mean value of channel 22 to mean value of top 40% of channel 15, and the ratio of mean value of channel 21 to mean value of top 40% of channel 25.[000120] Again, with reference to the spectral channels outlined in Table 4, when the feature selection methodology in another embodiment is an entropy-based feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA, comprising (iii) the ratio of mean value of top 10% of channel 2 and mean value of top 10% of channel 5, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 27, the ratio of mean value of channel 18 and mean value of top 10% of channel 3, the ratio of mean value of top 10% of channel 5 and mean value of top 10% of channel 23, the ratio of mean value of top 10% of channel 31 and mean value of top 10% of channel 18, the ratio of mean value of channel 8 to mean value of top 40% of channel 21 , the ratio of mean value of channel 5 to mean value of top 40% of channel 4, and the ratio of mean value of channel 29 to mean value of top 40% of channel 2.[000121] In some embodiments, the multispectral profile comprises no more than ten (10) optimised quantitative spectral features when selected using a feature based selection methodology selected from the group consisting of i) a Minimum Redundancy Maximum Relevance (mRMR) based feature selection methodology; ii) a particle swarm optimization feature selection methodology, iii) a Gini-index-based feature selection methodology, iv) an entropy-based feature selection methodology, v) a chi-square-based feature selection methodology, and vi) a forward feature selection methodology.[000122] For instance, and with reference to the spectral channels outlined in Table 5, when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise: the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 16 to mean value of channel 18, the ratio of mean value of top 10% of channel 22 to mean value of channel 18, the ratio of mean value of channel 15 to mean value of top 40% of channel 6, the ratio of mean value of top 10% of channel 20 to mean value of channel 19, the ratio of mean value of channel 14 to mean value of channel 13, the ratio of mean value of top 10% of channel 34 to mean value of top 10% of channel 31 , the ratio of mean value of top 10% of channel 12 to mean value of channel 16, the ratio of mean value of channel 8 to mean value of top 40% of channel 24, and the ratio of mean value of channel 10 to mean value of channel 9.[000123] In another embodiment, and with reference to the spectral channels outlined in Table 6, when the feature selection methodology is a forward feature based selection methodology, and the classification method is a linear support vector machine-learning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: (i) the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 22 to mean value of top 10% of channel 28, the ratio of mean value of channel 6 to mean value of top 10% of channel 8, the ratio of mean value of channel 2 to mean value of top 10% of channel 34, the ratio of mean value of top 10% of channel 10 to mean value of top 10% of channel 17, the ratio of mean value of channel 5 to mean value of channel 29, the ratio of mean value of channel 16 to mean value of top 10% of channel 12, the ratio of mean value of top 10% of channel 16 to mean value of channel 17, the ratio of mean value of channel 12 to mean value of channel 9, and the ratio of mean value of top 10% of channel 15 to mean value of channel 27.[000124] Again, with reference to the spectral channels outlined in Table 6, when the feature selection methodology is a forward feature based selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 quantitative spectral features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA comprise (ii) the minimum intensity value ofchannel 2, the ratio of mean value of channel 16 to mean value of top 40% of channel 24, the ratio of mean value of top 10% of channel 16 to mean value of top channel 18, the ratio of mean value of channel 9 to mean value of channel 10, the ratio of mean value of channel 14 to mean value of channel 32, the ratio of mean value of channel 24 to mean value of channel 17, the ratio of mean value of channel 14 to mean value of top 40% of channel 20, the minimum intensity of channel 13, the ratio of mean value of top 10% of channel 27 to mean value of top 10% of channel 16, and the ratio of mean value of channel 26 to mean value of top 10% of channel 17.[000125] In another embodiment, and with reference to the spectral channels outlined in Table 7, when the feature selection methodology is a particle swarm optimization based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise: the ratio of mean value of top 10% of channel 3 to mean value of channel 4, the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of channel 17 to mean value of channel 18, the ratio of mean value of channel 17 to mean value of top 40% of channel 3, the ratio of mean value of top 10% of channel 16 to mean value of channel 15, the ratio of mean value of channel 24 to mean value of top 40% of channel 4, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel 21 , the ratio of mean value of channel 17 to mean value of channel 6, the ratio of mean value of channel 8 to mean value of top 40% of channel 24, and the ratio of mean value of channel 6 to mean value of channel 4.[000126] In another embodiment, and again with reference to the spectral channels outlined in Table 7, when the feature selection methodology is a Gini-index based feature selection methodology, and the classification method is a linear support vector machinelearning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise: the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 17, the ratio of mean value of channel 21 to mean value of top 10% of channel 19, the ratio of mean value of top 10% of channel 21 to mean value of channel 17, the ratio of mean value of channel 4 to mean value of channel 27, the ratio of mean value of top 10% of channel 15 to mean value of top 10% of channel 19, the ratio of mean value of channel 4 to mean value of top 40% of channel 31 , the ratio of mean value of top 10% of channel 16 to mean value oftop 10% of channel 15, the ratio of mean value of channel 5 to mean value of channel 19, and the ratio of mean value of channel 5 to mean value of channel 33.[000127] In another embodiment, and again with reference to the spectral channels outlined in Table 7, when the feature selection methodology is a chi-square based feature selection methodology, and the classification method is a linear support vector machinelearning classifier, the at least 10 optimised quantitative features differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise: the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of channel 25 to mean value of channel 26, the ratio of mean value of top 10% of channel 15 to mean value of channel 17, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel 21 , the ratio of mean value of top 10% of channel 18 to mean value of channel 19, the ratio of mean value of channel 24 to mean value of top 40% of channel 4, the ratio of mean value of channel 5 to mean value of channel 19, the ratio of mean value of top 10% of channel 19 to mean value of channel 26, the ratio of mean value of channel 4 to mean value of channel 27, and the ratio of mean value of channel 17 to mean value of channel 18.[000128] In another embodiment, and again with reference to the spectral channels outlined in Table 7, when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of channel 25 to mean value of top 10% of channel 22, the ratio of mean value of channel 22 to mean value of top 10% of channel 17, the ratio of mean value of channel 26 to mean value of channel 12, 25thpercentile value of channel 20, the ratio of mean value of top 10% of channel 3 to mean value of top 10% of channel 15, the ratio of mean value of channel 14 to mean value of channel 13, the ratio of mean value of channel 16 to mean value of top 40% of channel 3, the ratio of mean value of top 10% of channel 6 to mean value of top 10% of channel 15, and the ratio of mean value of top 10% of channel 22 to mean value of channel 33.[000129] In another embodiment, and again with reference to the spectral channels outlined in Table 7, when the feature selection methodology is a particle swarm optimization feature selection methodology, and the classification method is a linearsupport vector machine-learning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of channel 23 to mean value of top 10% of channel 15, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel 21 , the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 16 to mean value of channel 15, the ratio of mean value of top 10% of channel 28 to mean value of top 10% of channel 6, the ratio of mean value of top 10% of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 19 to mean value of top 10% of channel 21 , the ratio of mean value of top 10% of channel 19 to mean value of top 10% of channel 22, and the ratio of mean value of channel 33 to mean value of top 10% of channel 2.[000130] In another embodiment, and again with reference to the spectral channels outlined in Table 7, when the feature selection methodology is a Gini-index based feature selection methodology, and the classification method is a linear support vector machinelearning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of top 10% of channel 15 to mean value of channel 27, the ratio of mean value of top 10% of channel 12 to mean value of top 10% of channel 19, the ratio of mean value of channel 2 to mean value of channel 24, the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 31 , the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of channel 12 to mean value of channel 19, the ratio of mean value of top 10% of channel 16 to mean value of channel 12, the ratio of mean value of channel 7 to mean value of channel 17, and the ratio of mean value of channel 8 to mean value of top 10% of channel 22.[000131] In another embodiment, and again with reference to the spectral channels outlined in Table 7, when the feature selection methodology is an entropy-based feature selection methodology, and the classification method is a linear support vector machinelearning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of channel 8 to mean value of top 10% of channel 33, the ratio of mean value of top 10% of channel 9 to mean value of channel 33, the ratio of mean value of channel 32 to mean value of top 40% of channel 17, the ratio of mean value of channel 7 to mean value of top 10% of channel 33, the ratio of mean value of top10% of channel 16 to mean value of channel 19, the ratio of mean value of channel 33 to mean value of top 40% of channel 17, the ratio of top 10% of mean value of channel 8 to mean value of channel 33, the ratio of mean value of top 10% of channel 12 to mean value of top 10% of channel 30, and the ratio of mean value of top 10% of channel 22 to mean value of channel 33.[000132] In another embodiment, and again with reference to the spectral channels outlined in Table 7, when the feature selection methodology is a chi-square based feature selection methodology, and the classification method is a linear support vector machinelearning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 34, the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of top 10% of channel 18 to mean value of top 10% of channel 17, the ratio of mean value of top 10% of channel 27 to mean value of channel 7, the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of channel 2 to mean value of channel 24, the ratio of mean value of channel 33 to mean value of channel 15, the ratio of mean value of top 10% of channel 21 to mean value of channel 32, and the ratio of mean value of top 10% of channel 4 to mean value of top 10% of channel 32.[000133] In another embodiment, and again with reference to the spectral channels outlined in Table 7, when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the at least 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA comprise: the ratio of mean value of channel 26 to mean value of top 10% of channel 22, the ratio of mean value of channel 28 to mean value of channel 13, the ratio of mean value of channel 28 to mean value of channel 18, the ratio of mean value of channel 28 to mean value of top 40% of channel 29, the ratio of mean value of top 10% of channel 20 to mean value of channel 12, the ratio of mean value of channel 27 to mean value of channel 17, the ratio of mean value of channel 22 to mean value of top 40% of channel 26, the ratio of mean value of top 10% of channel 3 to mean value of channel 27, the ratio of mean value of channel 24 to mean value of channel 17, and the ratio of mean value of channel 9 to mean value of channel 10.[000134] Auto-Machine Learning[000135] According to another preferred embodiment, the method further comprises the step of:[000136] e) applying a second classification method to the optimised quantitative features of the multispectral profile for further differentiating the histopathological cause of kidney graft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.[000137] In one embodiment, the second classification method is an auto machinelearning (AutoML) classifier.[000138] Examples of such classifiers include Linear Classifiers (such as Logistic Regression, Naive Bayes Classifier, Fisher's Linear Discriminant, Perceptron), Support Vector Machines, Decision Trees (including Boosted Trees and Random Forest), Neural Networks, Quadratic classifiers Kernel estimation (such as Nearest Neighbor) and many others. The classifiers may not be perfectly accurate and may misclassify some of the healthy subjects being assessed as having a degree of kidney dysfunction and some of the subjects presenting with kidney dysfunction as healthy; this is, however, typical and standard in the subject literature. The degree of this misclassification is captured in the receiver operator characteristics (ROC) curve. The overall quality of classifiers is captured in the area under the ROC curve (AUC); a perfect classifier has an AUC value =1 , a classifier which is no better than a random draw has an AUC value =0.5.[000139] In a preferred embodiment, the auto machine-learning (AutoML) classifier employed in the present invention is the open-source AutoML package, AutoGluon. Tabular (version: 0.8.3b20231023) from Amazon Web Services, which specifically employs the Tabular Dataset module for tabular data preprocessing and the Tabular Predictor module for model training and prediction.[000140] AutoGluon. Tabular integrates multiple traditional machine learning classifiers, including tree models, neural networks, k-nearest neighbors (KNN), support vector machine-learning (SVM), and linear models. AutoGluon Tabular enhances performance through weighted averaging and multi-layer stacking, automatically identifying the optimal combination of models and their corresponding weights.[000141] Specifically, AutoGluon. Tabular is utilized to apply the alternative classifiers available in this open-sourced software to one or more of the optimised quantitativefeature combinations that have been determined in some of the embodiments described above (see Table 4, Table 5 and Table 6).MATERIALS and METHODS[000142] A summary workflow flowchart describing the key aspects of the present study methodology is illustrated in Figure 1.[000143] Study participant recruitment and ethical considerations[000144] The study included adult individuals of either sex aged between 18 and 75 years under the care of the Royal North Shore Hospital, New South Wales, Australia, with previous kidney transplantation(s) and with a clinically indicated transplant kidney biopsy (either as a protocol or as-needed biopsy) performed on the day of urine sample collection. Informed consent was obtained from all study participants. Data collection in this study was carried out in accordance with relevant local guidelines and regulations, and collection of human data was approved by the human ethics committee at Royal North Shore Hospital (Ref: HREC / 17 / HAWKE / 471 ) and University of New South Wales, Sydney, Australia (Ref: HC180710).[000145] Procurement of allograft kidney biopsy tissue[000146] The procurement of allograft kidney biopsy tissue was performed at the Royal North Shore Hospital. Prior to commencing the procedure, written consent was obtained from study participants to collect the pre-biopsy urine sample for the purposes of the study, and obtain the allograft kidney biopsy tissue. Allograft kidney biopsy was then performed in an ultrasound-guided manner using a sterile technique. This was repeated 2 or 3 times depending on whether an adequate sample, subject to the operator’s evaluation, was obtained. A dressing was then applied to the site of needle insertion, and a further ultrasound scan was performed to check for any immediately occurring complications. Patients were required to lie flat for 4 hours with a further 1 -2 hours sitting on bed. They were monitored by nursing staff during this time period for hemodynamic stability prior to discharge.[000147] Evaluation of allograft kidney biopsy tissue to determine study participant groups[000148] Tissue obtained from allograft biopsies was subsequently transferred to the histopathology department at Royal North Shore Hospital and urgently assessed as per standard protocol - with majority of the sample processed for light microscopic evaluationvia paraffin-embedded sections, supplemented by special and immune histochemical (IHC) stains and some reserved for immunofluorescence and electron microscopic studies if indicated. Light microscopy assessment included a minimum of two hematoxylin and eosin (H&E), two periodic acid-Schiff (PAS), two Masson’s trichrome (trichrome), and two Jones methenamine silver (silver) stains in complementary fashion. H&E stains provided a general overview of all structures, cytoplasmic and nuclear features, PAS stains serving to highlighted tubular and glomerular basement membranes, trichrome stains accentuated fibrous tissue and fibrin, if present, and silver stains highlighted the glomerular and tubular basement membranes, and also sclerosis. An IHC or immunofluorescence stain for C4d was also routinely employed to evaluate for antibody- mediated rejection. The biopsy assessment was conducted and reported by accredited- qualified pathologist assessors in NSW Health Pathology Laboratory, Department of Anatomical Pathology, Northern Sydney Local Health District, Sydney, Australia.[000149] Study participants were divided into three groups, based on findings from their allograft biopsy detailing the likely cause of delayed or deteriorating graft function. This study aimed to differentiate between patients reported with acute tubular necrosis (the ATN group), graft rejection or NR-IFTA in transplant biopsy (two groups) using multispectral assessment of cell autofluorescence.[000150] Demographic and clinical characteristics of study participant groups[000151] Study participants’ demographic alongside clinical and biochemical data were acquired from the Royal North Shore Hospital electronic medical record, summarized using appropriate descriptive statistics and compared between the three groups. For variables with symmetric normal distributions, the mean and standard deviation were reported. For variables that were skewed or ordinal, the median and interquartile range were used for summarization. Proportions were also presented for categorical variables. Continuous variables between the groups were compared using the Analysis of Variance (ANOVA) test (if normally distributed) or the Kruskal-Wallis test (if the distribution was not normal). Categorical variables were compared using the Chi- square test or Freeman-Halton extension of the Fisher’s exact test accounting for sparsely distributed data.[000152] A total of 30 study participants were included, including 10 individuals with acute tubular necrosis (ATN), 10 individuals with graft rejection, and 10 individuals with non-rejection associated interstitial fibrosis and tubular atrophy (NR-IFTA). Table 1outlines the demographic and clinical characteristics for individual study participants between the three groups. Table 2 presents the overall cohort profile and compares demographic and clinical characteristics between the three groups. The mean age in the ATN and NR-IFTA groups appear to be significantly older compared to the graft rejection group [55.0 (ATN) vs 39.2 (rejection) vs 54.1 years (NR-IFTA), p=0.028], whilst the median number of days between kidney transplantation and allograft biopsy was significantly much higher in the NR-IFTA group compared to the other two groups [5 (ATN) vs 96 (rejection) vs 1939 days (NR-IFTA), p<0.001 ], as expected. Otherwise, there were no statistically significant differences in the demographic and clinical characteristics between the groups.[000153] Collection of urine samples and extraction of urinary exfoliated proximal tubule cells[000154] Using urine bottles with capacity of up to 200ml, spot urine samples were collected from study participants. Each collected urine sample was placed on ice immediately after collection and were centrifuged for 20 minutes at 4°C to collect exfoliated cells then washed twice with phosphate buffered-saline (PBS). Prior to extraction of PTCs, the initially processed sample was stored at -80°C in Phosphate buffer saline (PBS) with 10% Dimethyl Sulfoxide (10% DMSO). Specific PTCs expressing CD13, SGLT-2 and angiotensinogen were extracted via a validated specific immuno-magnetic separation method. Extracted cells were subsequently resuspended in 200pl of PBS on a 13mm glass bottom petri dish (Cell E&G, USA, and #GDB0004-200), and viewed in brightfield microscopy. Quantification of the number of exfoliated PTCs in brightfield microscopy was also performed.[000155] Multispectral autofluorescence imaging of extracted urinary exfoliated proximal tubule cells from allograft kidneys[000156] Multispectral autofluorescence imaging was conducted using a custom- made autofluorescence microscopy system, built by adapting a standard fluorescence microscope (Olympus iX83TM). The light source (from Quantitative Pty Ltd, Australia) in this single photon autofluorescence microscope provides a number of defined narrowband (± 5 nm) excitation wavelength ranges through low power LEDs and several epifluorescence filter cubes producing defined spectral channels (34 channels). The channels span wide excitation (340 nm - 510 nm) and emission (390 nm-900 nm) wavelength ranges, (Table 3) provides the details of the spectral channels. A 40* oil (NA1.15) objective, with transmission in the UV range was used for imaging. All images were captured by a Nuvu™ EMCCD camera HNu 1024 cooled to -65°C to reduce sensor- induced noise in the images (1024 x 1024 pixels).[000157] Image acquisition times of up to 5 seconds per channel, with three times averaging was applied to optimize image quality (i.e. signal-to-noise ratio). A sequence of fluorescence images for each spectral channel was collected for each sample area resulting in a “data block”). Each data block is supplemented by a brightfield image showing cell morphology. Representative brightfield and multispectral channel cell images from each group are presented (in Figure 2). In addition to cell imaging, reference images were captured of water as background fluorescence, and calibration fluid (30 pM NADH mixed with 18 pM FAD) displaying fluorescence across all channels. The calibration fluid was additionally measured on a calibrated spectrofluorometer (Flouromax 4) providing reference values. Manual cell segmentation was performed on brightfield images, generating masks outlining the PTCs.able 1. Demographic and clinical characteristics for individual study participants with post -transplant ATN, graft rejection nd NR-IFTABMR: Antibody-mediated rejection; ACR: Albumin creatinine ratio; ATN: Acute tubular necrosis; BCR: B-cell rejection; CNI: Calcineuri nhibitor; DM: Diabetes mellitus; eGFR: Estimated glomerular filtration rate; F: Female; FSGS: Focal segmental glomerular sclerosis; ID dentification; IFTA: Interstitial fibrosis and tubular atrophy; M: Male; Macro: Macroalbuminuria; Micro: Microalbuminuria; TCR: T-ce ejectionable 2. Baseline demographic and clinical characteristics of the study groups (n = 30)BMR: Antibody-mediated rejection; ACR: Albumin creatinine ratio; ATN: Acute Tubular Necrosis; BCR: B-cell rejection; CNI: Calcineurinnhibitor; eGFR: Estimated glomerular filtration rate; IFTA: Interstitial fibrosis and tubular atrophy; IQR: Interquartile range; SD: Standar eviation; TCR: T-cell rejection1p values are calculated using the ANOVA test (if normally distributed) or Kruskal-Wallis test (if the distribution is non-parametric) fo ontinuous variables, and the Chi-square test or Freeman-Halton extension of the Fisher’s exact test accounting for sparsely distribute ata for categorical variablesable 3. Details of the spectral channels used in this study for spectral imagingM: Electron-multiplying gain[000158] Cellular image preparation and feature extraction for analysis[000159] Following image acquisition, the spectral images were prepared for quantitative analysis by applying image smoothing to reduce the impact of noise, subtracting background autofluorescence, and implementing a calibration method. Image smoothing minimized the impact of Poisson’s noise and dead or saturated pixels. The calibration mechanism involved eliminating background fluorescence, adjusting for uneven illumination of the field of view, and aligning the multispectral images with reference values from the Fluoromax 4.[000160] As indicated above, the step of measuring autofluorescence from the cell includes the steps of performing image pre-processing; calculating, for each cell, quantitative features of the measured autofluorescence; removing correlations between the calculated quantitative features of different cells; and projecting, for each cell, the quantitative features of the measured autofluorescence onto a new vector space. The step of removing correlations may use Principal Component Analysis (PCA). The new vector space may be produced by Linear Discriminant Analysis (LDA). The use of a preprocessing PCA step is one way to avoid numerical problems in the later LDA stage when calculating a within-group sum of squares and cross products matrix which turns out to be singular if the input variables are linearly correlated. This PCA stage is unsupervised, and it uses the covariance matrix derived from all of the calculated feature vectors.[000161] PCA is necessary to produce a decorrelated version of the feature data, and this prevents numerical problems in later stages of the analysis. This decorrelated version of data leaves the important information about cell differences whilst removing data correlations. This procedure transforms the original basis vectors in the feature space into specific new basis vectors that are rotated with respect to the original basis vectors. The original dataset is transformed in such a way that the features become maximally decorrelated. This dataset still retains the meaning and the information content of the original features. Further data analysis proceeds in this new PCA-decorrelated space.[000162] The LDA stage uses prior knowledge of class assignment through data labelling and attempts to find a projection that optimally separates the data based on second order statistics through the use of Fisher’s statistical distance criterion.[000163] Image Pre-processing[000164] In one embodiment, the pre-processing steps include taking first set of reference images (Step A), primary denoising with removal outliers (spikes or dips) (StepB), image smoothing (Step C), removing background fluorescence (Step D) measurement of calibration fluid (Step E), background illumination flattening (Step F), measurement of second set of reference images and spectra (Step G), and cell segmentation (Step H). All this was carried out without changing the mathematical structure of the dataset. The pixel identification (image number, pixel coordinates, spectral channel etc.) are separately retained for the reconstruction of two-dimensional fluorophore abundance maps.[000165] Step A - Taking a reference image: At the beginning of each experiment, a reference image of water is taken using the multispectral microscope system. This reference image is then used to pre-process the sample images.[000166] Step B - Removing outliers (spikes): Abnormal behavior of sensor pixels in combination with high EM sensor gain may cause random sharp spikes or sharp dips in the image. To remove these outliers, a ‘threshold limiting window’ was scanned over all the images to locate these spikes or dips. Then these specific data points are replaced with the values interpolated from immediately adjacent nine pixels.[000167] Step C - Image smoothing: The main sources of noise from a camera include illumination independent and illumination-dependent noise. The illumination independent noises (e.g. dark-current shot noise, readout noise etc.) are minimized by using low sensor temperature (below -65°C). Illumination-dependent noise (e.g. photon shot noise, clock induced charge noise, EM gain register noise - if applicable etc.) is considered as multiplicative temporal and spatial noise. The overall noise in autofluorescence images was a combination of illumination dependent noise which was approximately Poissonian, while the noise from the illumination independent sources could be modelled as a Gaussian noise. Gaussian noise is used in simulation as a proxy for the overall noise, because the Poisson’s noise amplitude cannot be modified independently from the signal. A customized wavelet filter is used to remove the image noise for smoothing, which facilitated improved capture of spectral information from a signal compared with standard frequency spectra produced by Fourier analysis. The wavelets help to divide the signal into different scale components and thus these customized wavelet filters prove to be a computationally efficient method of capturing textural information from filters or banks of filters with attractive attributes with potentially lossless coverage of the frequency spectrum.[000168] Step D - Removing background autofluorescence: The images are also affected by the unavoidable autofluorescence signals from the microscope slide, petri dishes, dirt on sensors etc. These signals make additive contributions to all images. To remove these contributions, two hyperspectral images are taken of water in the petri dish used for imaging. The smoothed average of these two images is denoted by B(k, i). This smoothed average image, different for each channel, was subtracted from each sample image in this specific channel.[000169] Step E - Measurements of calibration fluid images: The microscope system was calibrated by taking hyperspectral images of a “calibration fluid”. This can include, for example, and calibration fluid (e.g. 30 pM NADH mixed with 18 pM FAD) which in these experiments. Its composition can be adjusted so that the spectrum of the calibration fluid has non-zero response across all the spectral channels. The smoothed image of the calibration fluid is denoted byCraw(fc, i).[000170] Step F - Image flattening: Finally, the raw sample image, yraw(fc, i), is corrected by using the averaged and smoothed background image B(k, i). Furthermore, the smoothed image of the calibration fluid was used to correct for the somewhat uneven (approximately Gaussian) illumination of the field of view. This was done by dividing the sample image in each channel (after subtracting the smoothed water image) by the relevant smoothed image of the calibration fluid. These corrections are specified in the equation below:[000171] Step G - Measurement of a second set of reference images and spectra: In the case when fluorophore unmixing is required, the relationship between standard fluorimetry and hyperspectral / multispectral microscopy of a set of reference pure fluorophore compounds must be obtained. The pure fluorophores are diluted to approximately physiological concentrations in the micromolar range. Their fluorescence spectra in the wavelength ranges corresponding to each pair of the excitation / emission channels are measured using standard fluorimetry and the same samples are then imaged using a multispectral / hyperspectral microscope, in all pairs of excitation / emission channels. The reference images and spectra are then utilized for fluorophore unmixing which may be carried out in the current context as per the publication “Statistically strong label-free quantitative identification of native fluorophores in a biological sample" bySaabah B. Mahbub, et al., Scientific Reports, volume 7, article number: 15792(2017). Fluorophore unmixing may or may not be required to identify quantitative features of relevance.[000172] Step H - Cell segmentation: In obtaining the multispectral profile, a quantitative feature may be a spatial average of channel intensity for each cell divided by the cell area for each cell. To calculate this example quantitative feature, the pixels corresponding to a specific cell may be identified by a procedure known as “cell segmentation”, and the corresponding autofluorescence signal measured within those specific pixels added across each pixel, separately for each spectral channel used. The area of the cell is then calculated by counting the pixels belonging to that cell. The two values thus obtained are then divided producing average autofluorescence intensity in that cell in this spectral channel. In other examples, the quantitative feature may be a mean, median, variance, kurtosis, or other Haralick feature calculated for each channel separately, or various derivative features combining individual channel features, such as channel ratios (the ratio of average cell intensities in two different channels) derived from the measured multispectral autofluorescence images for that specific cell. These features are separately calculated in each spectral channel. The features may be defined in a way that reflects specific biology of the cell under investigation, for example average content of specific fluorophores of relevance to a particular cellular pathway (such as bound NADH, free NADH, FAD, flavins, cytochrome C and many others), or for example characteristics of the mitochondria such as their perinuclear location or specific shape, or shape distribution. Then the average channel intensity for each cell, or any other quantitative channel feature or the set of features for each cell, may be used to assign a type or class label such as “acute tubular necrosis (ATN)", “graft rejection (GR)’” or “nongraft rejection interstitial fibrosis and tubular atrophy (NR-IFTA)".[000173] It may also be possible to use non-cellular features for classification, in particular pixel features, where features or average features for each cell are not calculated on a per cell basis, but the pixel data for the classes or groups are used. For example, one could use raw pixel autofluorescence signals or secondary features such as width of pixel autofluorescence signal distributions in each of the channels as quantitative features. Alternatively, one could use quantitative features which incorporate some cellular identifications but do not imply taking cellular averages such as, for example, raw pixel autofluorescence signals, only from the largest 10% of cells.[000174] The feature space may then be transformed to a “new vector space” as per paragraph below to optimally present cell group separation.[000175] In some examples, the new vector space is produced by Linear Discriminant Analysis (LDA) (see, for example, J. Ye, “Characterisation of a family of algorithms for generalized discriminant analysis on under sampled problems", J. Mach. Learn. Res. 6 (2005) 483-502). The “new” vector space means a vector space whose set of basis vectors differs from the set of basis vectors of the original feature space. In other examples, the new vector space may be produced by alternative methods, including rotation under subjective manual control.[000176] This means that the set of quantitative features thus obtained for each cell can then be projected, onto a vector space that optimally discriminates or separates the data based on this class assignment. This projection may be done by using LDA. The dimensionality of this new space and hence the number of new canonical variables is P - 1 , where P is the number of unique classes assigned to the data (for example, if trying to separate two classes “healthy” and “sick”, then P=2 and the projection is onto a 1 - dimensional space). The data are projected onto specific directions determined by LDA, these directions based on the actual cell data. The coordinates of each cell are now expressed in terms of these canonical variables, sometimes called “spectral variables”, and reflect the distance measured along these specific directions. Two out of P - 1 directions may be selected to generate scatterplots to aid in visualizing the data.[000177] Therefore, in the methods of the invention, P cell classes are initially chosen and, by using LDA, the original feature space and the vectors representing the cell features are projected onto a new, lower dimensional space. Its dimension is given by the number of groups of cell classes to be distinguished, less 1 .[000178] In some examples, three classes of cells may be used, so that after LDA the spectra of these cells can be depicted as points on two-dimensional plots. This two- dimensional spectral space produced by LDA is one of the examples of “canonical spectral spaces” that are convenient for visualization. Its basis vectors are orthogonal and may be aligned with the axes in two-dimensional plots for visual representation.[000179] The LDA method ensures that the new space is optimized to provide the best degree of separation between selected cell classes (such as, for example, cells from different patients). In some examples, in order to quantify the distinctiveness between selected pairs of cell clusters, the LDA analysis may be performed again on each pair ofcell cluster data projecting them onto a one-dimensional line. The Kolmogorov-Smirnov or alternative statistical tests such as t-test may then be applied to gauge and compare the similarity of the pair of clusters. In some examples, the maximum Fisher statistical distance may also be calculated. This is a measure of cluster closeness which is sensitive to cluster means and takes account of the data dispersion.[000180] In some examples, data for additional cells, cell groups, and / or patients may be plotted together with the previously obtained cell autofluorescence data as transformed by methods described herein. In this approach the new data are projected on the vector space optimized to provide best separation of the original groups, but not necessarily the new groups formed by integrating the previous groups with new data. Although there is no mathematical certainty that optimum separation will be achieved for such blended datasets, a clear separation may often be achieved in the case when class distinction results are statistically strong with small p-values.[000181] Bioinformatic analysis may identify one or more quantitative features of the cells which, in combination, enable distinguishing cell ensembles of healthy patients from those of patients presenting with kidney dysfunction.[000182] Pixel intensity values are defined for a given excitation spectral channel by the measured autofluorescence intensity at each pixel in the multispectral image. Using these values, and having segmented the cells, it is possible to calculate quantitative cellular features such as mean, median, variance, kurtosis, or other features known to those skilled in the art. It is important for some of these features to be divided by the cell area calculated, for example as the number of pixels belonging to that cell. Therefore, a set of quantitative features may be calculated on the basis of pixel intensities for each cell captured by the multispectral image.[000183] Pixel intensity ratios are defined for a given pair of excitation spectral channels by the ratio of the measured autofluorescence intensity at each pixel of the first pair of excitation / emission channels with respect to the measured autofluorescence intensity at each pixel of the second pair of excitation / emission channels. Using these vectors for each specific cell, it is possible to calculate quantitative features such as mean, median, variance, kurtosis, or other quantitative features. Therefore, a multitude of quantitative features derived from pixel intensity ratios may be calculated for each cell captured in the multispectral image. In analogy to this example, more involved pixelfunctions of alternative types may also be calculated, and related cellular features produced.[000184] The calculated quantitative feature vectors for each cell may be arranged in a P by Q matrix, where P is the number of cells and Q is the number of quantitative features for each cell. The data may then undergo further processing, by PCA or LDA or alternatives as described above, causing the new variables to satisfy group variance maximization criteria.[000185] The uncorrelated variables (post-PCA) can further be used for discriminatory analysis. The discriminatory analysis provides another set of variables maximizing the separation between pre-specified groups. The number of variables returned by PCA, and discriminatory analysis is equal to the number of statistical features, however, in some examples, only some of the variables may be plotted for data visualization.[000186] In this example, kidney cells are selected manually by using a superimposed brightfield image. The normalized autofluorescence intensity in each cell is documented in the yki matrix.[000187] The PTCs in the images were manually segmented.[000188] Following image preparation and segmentation, image features across all spectral channels were computed for each segmented cell. A total of 7703 hand-crafted, quantitative features per cell were generated in MATLAB 2023b using an in-house generated code. The features included cell-average intensities in each channel and channel intensity ratios.[000189] Machine learning analysis to differentiate between multispectral autofluorescence signals in exfoliated proximal tubule cells between individuals across the study groups[000190] Binary classification of cells from different groups of patients was carried out using three different methods: (i) random forest classifier, (ii) linear support vector machine-learning (SVM) classifier and (iii) by applying the AutoGluon machine learning software from Amazon Web Services to explore and rank alternative classifiers.[000191] In each of these methods, multispectral autofluorescence images of PTCs were quantitatively analyzed by using optimized small feature sets to avoid overfitting.[000192] For method (i), the best 8 features (Table 4) were selected by using an entropy-based approach.able 4. Optimal spectral feature combination selected to differentiate between urinary exfoliated proximal tubule cells in the three groups using method (i)[000193] In the linear SVM classification of method (ii), features were selected using forward feature selection or Minimum Redundancy Maximum Relevance (MRMR), resulting in the AUC value of 0.87 (for ATN versus graft rejection), AUC=0.87 (for ATN versus NR-IFTA) and AUC=0.77 (for graft rejection versus NR-IFTA) respectively. The AutoGluon classifier optimization for the same features resulted in AUC values of 0.88 (for ATN versus graft rejection), AUC=0.86 (for ATN versus NR-IFTA) and AUC=0.72 (for graft rejection vs NR-IFTA).[000194] The inventors explored 6 different feature selection methods: (1 ) MRMR feature selection; (2) particle swarm optimization to select the largest Fisher distance; (3) Gini-index based feature selection; (4) entropy-based feature selection; (5) chi-square- based feature selection and (6) forward feature selection.[000195] For method (ii), the inventors found that for the comparison between the ATN versus graft rejection, MRMR was found to be the optimal feature selection strategy yielding 10 specific features, as listed in Table 5.able 5. Optimal spectral feature combination selected to differentiate between urinary exfoliated proximal tubule cells in the ATN v raft rejection groups by MRMR feature selection and using a linear SVM classifier in method (ii)[000196] On the other hand, when comparing between the ATN versus NR-IFTA, as well as between the graft rejection versus NR-IFTA groups, the forward feature selection was the optimal strategy in method (ii) for yielding a set of different 10 features (Table 6).able 6. Optimal spectral feature combination selected to differentiate between urinary exfoliated proximal tubule cells in the ATN vs raft NR-IFTA and graft rejection vs NR-IFTA groups by forward feature selection and using a linear SVM classifier in method (ii)TN: Acute tubular necrosis; MRMR: Minimum Redundancy Maximum Relevance; NR-IFTA: Non-rejection associated IFTA - Interstitial ibrosis and tubular atrophy; PTC: Proximal tubule cells; SVM: linear Support Vector Machine-Learning[000197] For method (iii), automated machine learning (autoML) was used to train all classifiers. For instance, suitable autoML classifiers for training purposes include multiple linear, tree-based algorithms, neural networks, and k-nearest neighbors (KNN), including Naive Bayes, KNeighbors, Logistic Regression, Decision Tree, Random Forest, Extra Tree, Support Vector Machines (SVMs), AdaBoost (Adaptive Boosting), XGBoost (extreme Gradient Boosting) and LightGBM (Light Gradient Boosting Machine), CatBoost, MXNet Neural Network, and FastAI Neural Network.[000198] According to a preferred embodiment, the classifiers were selected as initial models, based on screening results and practical consideration. These models were combined or trained on multiple stratified folds using bagging and stack assembling techniques with an auto-ML library.[000199] As indicated above, the inventors utilized AutoGluon. Tabular (version: 0.8.3b20231023), specifically employing the TabularDataset module for tabular data preprocessing and the TabularPredictor module for model training and prediction.[000200] In each of methods (i), (ii) and (iii), 10-fold cross-validation was conducted to validate the results, with performance metrics measured using the Area Under Curve (AUC) value in the Receiver Operating Characteristic (ROC) curves.[000201] Thus, for each of the comparative analyses, feature combination subsets generated from the alternative feature selection methods that were not the most optimal in terms of AUC performance but generated classification performance results of AUC > 0.70, are detailed in Table 7.able 7. Alternative spectral feature combinations selected to differentiate between urinary exfoliated proximal tubule cells in the ATN raft rejection and NR-IFTA groups by method (ii)TN: Acute tubular necrosis; GR: Graft rejection; NR-IFTA: Non-rejection associated Interstitial fibrosis and tubular atrophy; MRMR: Minimum Redundancy Maximum Relevance; TC: Proximal tubule cells[000202] In method (iii), the AutoGluon. TabularDataset and TabularPredictor modules were utilized to explore alternative classifiers available in this software on each of the feature combinations from Table 4, Table 5 and Table 6.[000203] Patient-level classification to differentiate between individuals across the study groups by multispectral autofluorescence features[000204] A patient-to-patient level classification model was developed using mean values calculated from the cell features for each study participant. In developing patientlevel classification models, the feature data from Table 5 and Table 6 was used. Average cellular feature values were calculated for each study participant. This resulted in a data set consisting of a total of ten (10) patient features for each of the ten study participants per group.[000205] Each feature combination was fed into the AutoGluon. Tabular model for training. The performance was comprehensively assessed using receiver operating characteristic (ROC) curves, area under the curve (AUC) scores, and confusion matrices.[000206] To prevent overfitting, Principal Component Analysis (PCA) was performed to reduce the number of patient features to the top three (3) most important principal components per patient. The resulting feature table was prepared by specifying which columns represented the features and which represented the label. This table was then fed into the AutoGluon Tabular model for training.[000207] Classification was conducted using 5-fold cross-validation with AutoGluon. Tabular The classification performance was comprehensively assessed using the mean AUC scores obtained across all folds, represented as ROC curves and confusion matrices.[000208] To clarify the derivation of the principal components (PCs) according to this process, the main steps of PCA are outlined below together with an explanation of how each step contributes to generating these components. PCA is a we 11 -recognized, widely- used data analysis technique that helps capture the major variance in a dataset with reduced dimensions.[000209] The steps are as follows:1. Standardize the Data (from either Table 5 or 6): First, each feature was centred by subtracting the mean and, if necessary, scale to unit variance. Thisensures that features with larger scales do not dominate the principal components.2. Compute the Covariance Matrix: The covariance matrix captures relationships between features, quantifying how each pair of features varies together. This step is crucial, as it lays the foundation for understanding which feature combinations hold the most variance.3. Calculate Eigenvalues and Eigenvectors: The eigenvalues show the magnitude of variance in the direction of each eigenvector. Eigenvectors, on the other hand, point to the directions of maximum variance, forming the principal components.4. Sort Eigenvalues and Select Principal Components: The eigenvalues were arranged in descending order, then the top k (in this case, k=3 for the top three principal components) were selected to capture the most variance in the data. The corresponding eigenvectors form the new feature space.5. Transform the Data: Finally, the original data was projected onto the selected eigenvectors (principal components), creating a transformed dataset with reduced dimensions that retains maximum variance.[000210] The coefficients of each principal component (PC1, PC2, and PC3) for each of the group-to-group comparisons (i.e. ATN vs Graft Rejection, ATN vs NR-IFTA, Graft Rejection vs NR-IFTA) at a patient-to-patient classification level are shown in Tables 8a, 8b and 8c, using the top 3 PCA features calculated from Table 5, Table 6 and Table 6, respectively.[000211] Information about the corresponding PC1, PC2 and PC3 coefficients for these same group-to-group comparisons is shown in Tables 9a, 9b and 9c, respectively.able 8a Top 3 Coefficients of each principal component (PC1, PC2, and PC3) for ATN vs Graft rejection (GR)able 8b Top 3 Coefficients of each principal component (PC1, PC2, and PC3) for ATN vs NR-IFTAable 8c Top 3 Coefficients of each principal component (PC1, PC2, and PC3) for Graft Rejection (GR) vs NR-IFTAable 9a Information Associated with the top 3 Coefficients of each principal component (PC1, PC2, and PC3) forATN vs Graftejection (GR)able 9b Information Associated with the top 3 Coefficients of each principal component (PC1, PC2, and PC3) forATN vs NR-IFTAable 9c Information Associated with the top 3 Coefficients of each principal component (PC1, PC2, and PC3) for Graft Rejection (GR) vs NR-IFTA[000212] RESULTS[000213] Multispectral autofluorescence imaging was completed for 161 cells in total (51 cells in the ATN group, i.e. , on average 5.1 cells / patient; 60 cells in the graft rejection group (GR), i.e., on average 6.0 cells / patient; 50 cells in the NR-IFTA group, i.e., on average 5.0 cells / patient).[000214] Differentiation of multispectral autofluorescence signals in exfoliated proximal tubule cells between study participants presenting with ATN and graft rejection (GR)[000215] Multispectral imaging and subsequent analysis using method (i) for binary classification of cells, revealed significant differences in multispectral cellular feature patterns between exfoliated PTCs from study participants with ATN versus graft rejection (GR) with an AUC value of 0.97±0.03 (Figure 3a).[000216] Upon selecting the optimal feature combination (Table 5) using method (ii), the inventors were able to categorize PTCs between these two groups with an AUC value of 0.87±0.03 (Figure 3b).[000217] The analysis using AutoGluon. Tabular in method (iii) revealed significant differences in multispectral cellular features between exfoliated PTCs from study participants in these groups, with an AUC value of 0.95±0.05 (Figure 3c), when using features from Table 4 and AUC value of 0.88±0.14 (Figure 3d) when using the features from Table 5.[000218] The patient-level classification model, when using the top 3 PCA features calculated from Table 5, revealed significant differences in exfoliated PTC multispectral cellular feature patterns between study participants with ATN versus graft rejection (GR) with an AUC value of 0.85±0.20 (Table 10). The corresponding ROC and confusion matrix for the ATN versus graft rejection (GR) group are shown in Figure 6a and Figure 6b, respectively.able 10. Multispectral differentiation of exfoliated PTCs between study participants with ATN vs graft rejection (GR), ATN vs NR-IFTA and graft rejection (GR) vs NR-IFTA - mean AUC values for obtained patient classifier based on performing principa component analysis (PC A) on the feature data initially derived from method (ii)UC Area under the curve; ATN: Acute tubular necrosis; GR: Graft rejection; NR-IFTA: Non-rejection associated Interstitial fibrosis anubular atrophy; PTC: Proximal tubule cells[000219] Differentiation of multispectral autofluorescence signals in exfoliated proximal tubule cells between study participants presenting with ATN and NR-IFTA[000220] Multispectral imaging and subsequent analysis using method (i) for binary classification of cells revealed significant differences in multispectral cellular feature patterns between exfoliated PTCs from study participants with ATN versus NR-IFTA, with an AUC value of 0.84±0.11 (Figure 4a).[000221] Upon selecting the optimal feature set (Table 6) using method (ii), the inventors were able to categorize PTCs between the groups with an AUC value of 0.87±0.06 (Figure 4b).[000222] The analysis using AutoGluon. Tabular in method (iii) revealed significant differences in multispectral cellular feature patterns between exfoliated PTCs from study participants in the two investigated groups, with an AUC value of 0.92±0.08 (Figure 4c) when using features from Table 4 and AUC value of 0.86±0.11 (Figure 4d) when using the features from Table 6.[000223] The patient-level classification model revealed significant differences in exfoliated PTC between study participants with ATN versus NR-IFTA groups. Using the top 3 PCA features calculated from Table 6, an AUC value of 0.82±0.29 was obtained (Table 10). The corresponding ROC and confusion matrix for the ATN versus NR-IFTA group are shown in Figure 7a and Figure 7b, respectively.[000224] Differentiation of multispectral autofluorescence signals in exfoliated proximal tubule cells between study participants presenting with graft rejection (GR) and NR-IFTA[000225] Multispectral imaging and subsequent analysis using method (i) for binary classification of cells revealed significant differences in multispectral cellular feature patterns between exfoliated PTCs from graft rejection versus NR-IFTA, with an AUC value of 0.70±0.15 (Figure 5a).[000226] Upon selecting the optimal feature combination (Table 6) using method (ii), the inventors were able to categorize PTCs between the groups with an AUC value of 0.77±0.07 (Figure 5b).[000227] The analysis using AutoGluon. Tabular in method (iii) revealed significant differences in multispectral cellular features between exfoliated PTCs from study participants classified in the two groups investigated here, with an AUC value of0.91 ±0.10 (Figure 5c) when using features from Table 4 and an AUC value of 0.72±0.20 (Figure 5d) when using the features from Table 6.[000228] The patient-level classification model revealed differences in exfoliated PTC multispectral cellular feature patterns between study participants from graft rejection versus NR-IFTA groups. The top 3 PCA features calculated from Table 6 produced an AUC value of 0.55±0.19 (Table 10). The corresponding ROC and confusion matrix for the graft rejection (GR) versus NR-IFTA group are shown in Figure 8a and Figure 8b, respectively.[000229] DISCUSSION[000230] This study was able to demonstrate an excellent degree of discrimination at a cell-to-cell classification level between urinary exfoliated PTCs extracted from individuals with various histopathological complications following kidney transplantation - namely between ATN versus graft rejection and ATN versus NR-IFTA.[000231] Group differences were very distinctive and several independent analysis methodologies yielded comparable results clearly confirming these differences. When comparing between the ATN versus graft rejection and ATN versus NR-IFTA groups, clear discrimination has also been demonstrated at a patient-to-patient classification level, indicating the translational potential of our methodology for clinical application within this context.[000232] Cell exfoliation is an active biochemical process that has been linked to the homeostasis of mammalian organs. Exfoliation occurs where extracellular matrix components that usually tightly connect between cells within a structure break off, with live and dead external cells removed from the epithelial luminal surface. Cell exfoliation is thought to be under the control of cell metabolism, and the properties of kidney cells exfoliated into urine are now deemed to be a potentially useful indicator of kidney pathology. Cell exfoliation in the context of kidney transplantation is less well studied.[000233] The urine-derived renal epithelial cells (URECs) that are frequently obtained in the early stages following kidney transplantation have been shown to conventionally be derived from proximal tubules. Voided URECs are known to have high proliferating and inflammatory properties, suggesting their potential role in prognosticating posttransplant ischemia-reperfusion and acute kidney injury (AKI) states as well as their immunomodulatory potential.[000234] Innovation of novel urinary biomarkers and subsequent development of precision assessment methods to differentiate between the histopathological causes of graft dysfunction such as ATN versus graft rejection (GR) over the past decade have otherwise been associated with the utilization of various molecular biomarkers such as urinary multi-omics markers (for example, urinary transcriptom ics and proteomics) and urinary exfoliated extracellular vesicles, in which their application in post-transplant monitoring are increasingly promising.[000235] To the inventors’ knowledge, this report is the first to evaluate multispectral autofluorescence characteristics in exfoliated human transplanted cells from the kidney.[000236] In addition to the novelty of pursuing multispectral autofluorescence analysis in urinary exfoliated PTCs, this study also involves the application of artificial intelligence (Al) machine-learning based analysis to predict causes of graft dysfunction post kidney transplantation.[000237] Comparative Artificial Intelligence Analysis of Urinary Proximal Tubule Cell Autofluorescence in Kidney Transplant Dysfunction[000238] The following study provides a comparison of Al analysis of urinary proximal tubule cell (PTC) autofluorescence in kidney transplant dysfunction.[000239] The study applies the previous multispectral autofluorescence imaging and machine-learning framework for the cohort of N = 10 to a larger sample size (N = 19) of kidney transplant recipients to evaluate if similar classification performance outcomes could be achieved.[000240] Methodology[000241] Image-based features (histology masks and their corresponding MATLAB output data) were merged into a feature matrix, invalid values (Inf / NaN) were removed, and statistical outliers were excluded (e.g. z-score > 3).[000242] The feature vectors were normalized using MinMax scaling, and data were split 80 / 20 for training and test data. The training data were then used to train a Random Forest classifier, using 5-fold cross-validation with area under the curve (AUC) used as the scoring metric.[000243] After initial training, the eight (8) features with the highest importance ranking were selected and the model was re-trained on these features (repeating ROC-AUC analysis). In the Results, the term ‘CV AUC’ refers to the mean area under the curve (AUC) across the 5 cross-validation folds on the 80% training data.[000244] AutoGluon was then applied as a second independent automated machinelearning library.[000245] For each binary comparison, an abbreviated data set was built including only the eight (8) Random Forest-selected features and the class label and the same 80 / 20 train-test split was applied. AutoGluon was trained only on the training partition (with internal validation) and the held out 20% partition was reserved for independent evaluation of model performance (AUC, accuracy, precision and recall).[000246] This ensured that no information from the held-out samples was used during model fitting or hyperparameter optimization, and that all reported metrics are based on models using the selected eight-feature panels.[000247] Feature Summary[000248] The top features identified for the N=19 cohort study are indices identified as follows.[000249] For the ATN vs GR task, the top features identified for the N=19 cohort study are indices [140, 138, 76, 34, 65, 77, 74, 58],[000250] For the ATN vs NR-IFTA task, the top features identified for the N=19 cohort study are indices [138, 140, 11 , 137, 143, 77, 105, 91 ].[000251] While, for GR vs NR-IFTA, the top features identified for the N=19 cohort study are indices [34, 90, 33, 92, 67, 78, 89, 46],[000252] The top eight (8) features (by feature index) for each of these classification tasks in the N=19 cohort study are shown in Table 11.Table 11AUC: Area under the curve; ATN: Acute tubular necrosis; GR: Graft rejection; NR-IFTA: Non-rejection associated Interstitial fibrosis and tubular atrophy.[000253] In essence, the integers in Table 11 (e.g. 140, 138, 76) are index numbers into the pre-defined 145-dimensional feature matrix (F0-F144) which is produced by the urinary proximal tubule cell (PTC) multispectral autofluorescence pipeline.[000254] Each feature index refers to a pre-defined quantitative descriptor computed from a single time-point image, including intensity and texture features.[000255] The top 8 features in the N = 19 cohort study for each comparison therefore refer to the eight (8) most informative members of this fixed F0-F144 feature library, not arbitrary labels.[000256] The formulas and spectral channel definitions for these features are as described above, with the feature indices most often selected being those determined using the Random Forest classifier as most predictive for each classification task.[000257] Classification Tasks and Results[000258] All three binary comparisons (ATN vs GR, ATN vs NR-IFTA, GR vs NR- IFTA) have been tested.[000259] Model performance for each task is presented in Table 12.[000260] For each task, we report the 5-fold AUC (mean over folds) for the Random Forest model in the text and key held-out test metrics (AUC, accuracy, precision, recall) from the final AutoGluon model in Table 12.[000261] The AUC values in Table 12 correspond to performance of the AutoGluon models on the independent 20% held-out test set using the eight selected features for each task.[000262] Figure 9 shows the (ROC) curves obtained from a patient-to-patient level classification performance assessment, which reveals significant differences in the exfoliated proximal tubule cells (PTCs) multispectral cellular features between study participants in the N = 19 cohort study in the (a) ATN vs Graft Rejection (GR) group, (b) the ATN vs NR-IFTA group, and (c) the GR vs NR-IFTA group, when classified using the AutoGluon classifier framework (with 5-fold cross-validation) and the top 8 ranked multispectral autofluorescence features determined from the feature data obtained from Table 11[000263] As shown in Figure 9 (a), the Random Forest model for the ATN vs GR group of the N=19 cohort analysis had a mean CV AUC of 0.782 (based on the top 8features, see Table 11), and achieved a test AUC of 0.9186 (accuracy 0.8333, precision 0.8333, recall 0.8824).[000264] As shown in Figure 9 (b), the Random Forest model for the ATN vs NR- IFTA group of the N=19 cohort analysis had a mean CV AUC of 0.791 , and achieved a final test AUC of 0.8867 (accuracy 0.7188, precision 0.6667, recall 0.8750).[000265] While in Figure 9 (c), the Random Forest model for the GR vs NR-IFTA group of the N=19 cohort analysis had a mean CV AUC of 0.7611 , achieving a final test AUC of 0.9378 (accuracy 0.7667, precision 0.7000, recall 0.9333).[000266] All the results shown below in Table 12 are reported using AutoGluon models trained using the eight (8) highest-ranked features for each task. The AUC values in Table 12 show the area under the receiver operating characteristic (ROC) curve on the independent 20% held-out test set; accuracy, precision and recall are calculated from these held-out test predictions.Table 12ATN: Acute tubular necrosis; GR: Graft rejection; NR-IFTA: Non-rejection associated Interstitial fibrosis and tubular atrophyComparative Summary[000267] In summary, the inventors observed that the N = 19 (large) cohort study replicated the high AUC values observed for the smaller N = 10 cohort study across most tasks, but was not identical.[000268] Notably, the AUC for the N = 10 cohort study was slightly higher for ATN vs GR (0.9545 vs 0.9186), but N = 19 had a higher recall at a loss of precision (0.882 vs 0.818).[000269] For ATN vs NR-IFTA, both the N=10 and N=19 cohorts generated high AUCs (~0.92 vs 0.89) and very high accuracy, but the precision for the N=19 (large) cohort was much lower (0.667 vs 0.889).[000270] In comparing GR vs NR-IFTA tasks, the N=10 cohort study delivered modest results with an AUC of 0.894 and low accuracy and recall, while the N=19 cohort study demonstrated an improvement with an AUC of 0.9378 with substantially higher accuracy and recall.[000271] This is consistent with previous work showing that, in small-sample, highdimensional biomedical datasets, adding more samples and slightly changing class balance can alter the apparent performance and calibration of otherwise-identical models.[000272] CONCLUSION[000273] In summary, the inventors employed an entropy-based feature selection methodology combined with either (i) a Random Forest classifier, (ii) a linear support vector machine-learning (SVM) classifier, or (iii) an automated-machine-learning (AutoML) classifier, for use in processing images of the native fluorescence emission produced by exfoliated proximal tubule cells (PTCs) obtained from the urine sample of a recipient of a kidney transplant for the purpose of differentiating between histopathological causes of kidney graft dysfunction following kidney transplantation.[000274] The Random Forest classification method (i) resulted in cell classification with an area under the curve (AUC) value of 0.97±0.03 (for ATN versus graft rejection), AUC=0.84±0.11 (for ATN versus NR-IFTA) and AUC=0.70±0.15 (graft rejection versus NR-IFTA).[000275] The linear support vector machine-learning (SVM) classification method (ii) resulted in cell classification with an area under the curve (AUC) value of 0.87±0.03 (for ATN versus graft rejection), AUC= 0.87±0.06 (for ATN versus NR-IFTA) and AUC= 0.77±0.07 (for graft rejection versus NR-IFTA).[000276] While the AutoML classification method (iii) using the AutoGluon. Tabular software from Amazon Web Services, resulted in cell classification with AUC values of 0.95±0.05 (for ATN versus graft rejection), AUC=0.92±0.08 (for ATN versus NR-IFTA) and AUC=0.91 ±0.10 (for graft rejection versus NR-IFTA), respectively.[000277] Applying a patient-level classification model using the optimal sets of multispectral autofluorescence features derived as above, the inventors were also able to discriminate between study participants with an AUC value of 0.85±0.20 (for ATN versus graft rejection), an AUC value of 0.82±0.29 (for ATN versus NR-IFTA) and an AUC value of 0.55±0.19 (for graft rejection vs NR-IFTA).[000278] The findings provided herein demonstrate that measurement of the autofluorescent features from urinary exfoliated PTCs could be used to differentiate between patient groups with acute tubular necrosis (ATN), graft rejection and NR-IFTA in kidney transplant recipients. The accuracy of classification was excellent at a cellular level, and high at a patient level.[000279] Given the retrieval of human urine is non-invasive, convenient to obtain and given our methodology in extracting and assessing urinary exfoliated PTCs via multispectral autofluorescence imaging is robust, it is foreseeable that this technique could be implemented in post-transplant settings as a clinical decision-making tool to guide diagnosis and patient management in the future.[000280] Although the invention has been described with reference to specific examples, it will be appreciated by those skilled in the art that the invention may be embodied in many other forms, and in particular features of any one of the various described examples may be provided in any combination in any of the other described examples. Various modifications and alterations to this invention will become apparent to those skilled in the art without departing from the scope and spirit of this invention. It should be understood that this invention is not intended to be unduly limited by the illustrative embodiments and examples set forth herein and that such examples and embodiments are presented by way of example only with the scope of the invention intended to be limited only by the claims set forth herein as follows.

Claims

CLAIMS1. A method of assessing a histopathological cause of kidney graft dysfunction in a subject following kidney transplantation, the method comprising the steps of: a) generating one or more images of the native fluorescence emission from one or more exfoliated kidney cells obtained from a urine sample from the subject in a plurality of distinctive spectral channels, b) calculating, for each urinary exfoliated kidney cell, quantitative features of the autofluorescence signals in said one or more images, c) applying a feature selection methodology to the quantitative features to generate a set of optimised quantitative features; and d) applying a classification method to the optimised quantitative features to produce a multispectral profile for differentiating the histopathological cause of kidney graft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.

2. The method according to claim 1 , wherein the classification method is a binary classification into one of two possible classes of histopathological causes of kidney graft dysfunction selected from the group consisting of: a) acute tubular necrosis (ATN) versus graft rejection (GR); b) acute tubular necrosis (ATN) versus non-graft rejection interstitial fibrosis and tubular atrophy (NR-IFTA); and c) graft rejection (GR) versus non-graft rejection interstitial fibrosis and tubular atrophy (NR-IFTA).

3. The method of claim 1 , wherein the distinctive spectral channels are defined by a central excitation wavelength of between 340 nm-510 nm and an emission wavelength of between 370 nm-900 nm, with a spectral width of each spectral channel in the range of several tens of nanometers.

4. The method of claim 1 or 2, wherein the distinctive spectral channels are selected from:

5. The method according to any one of claims 1 to 3, wherein in step b), the method further comprises one or more of the following steps of: b1 ) performing image pre-processing; and b2) removing correlations between the calculated quantitative features of different urinary exfoliated kidney cells.

6. The method according to any one of claims 1 to 5, wherein when the feature selection methodology is an entropy-based feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprising: the ratio of mean value of channel 8 and mean value of top 10% of channel 2, the ratio of mean value of top 10% of channel 31 and mean value of top 10% of channel 32, the ratio of mean value of top 10% of channel 24 and mean value of top 10% of channel 28, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 23, the ratio of mean value of channel 22 to mean value of top 40% of channel 12, the ratio of mean value of channel 27 to meanvalue of top 40% of channel 5, the ratio of mean value of channel 21 to mean value of top 40% of channel 25, and the ratio of mean value of channel 24 to mean value of top 40% of channel 7.

7. The method according to any one of claims 1 to 5, wherein when the feature selection methodology is an entropy-based feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA, comprising: the ratio of mean value of top 10% of channel 24 and mean value of top 10% of channel 28, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 23, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 21 , the ratio of mean value of channel 5 to mean value of top 40% of channel 4, the ratio of mean value of channel 22 to mean value of top 40% of channel 12, the ratio of mean value of channel 21 to mean value of top 40% of channel 3, the ratio of mean value of channel 22 to mean value of top 40% of channel 15, and the ratio of mean value of channel 21 to mean value of top 40% of channel 25.

8. The method according to any one of claims 1 to 5, wherein when the feature selection methodology is an entropy-based feature selection methodology, and the classification method is a random forest classifier, the multispectral profile comprises a combination of no more than 8 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA, comprising: the ratio of mean value of top 10% of channel 2 and mean value of top 10% of channel 5, the ratio of mean value of top 10% of channel 26 and mean value of top 10% of channel 27, the ratio of mean value of channel 18 and mean value of top 10% of channel 3, the ratio of mean value of top 10% of channel 5 and mean value of top 10% of channel 23, the ratio of mean value of top 10% of channel 31 and mean value of top 10% of channel 18, the ratio of mean value of channel 8 to mean value of top 40% of channel 21 , the ratio of mean value of channel 5 to mean value of top 40% of channel 4, and the ratio of mean value of channel 29 to mean value of top 40% of channel 2.

9. The method according to any one of claims 1 to 5, wherein the multispectral profile comprises no more than 10 optimised quantitative features when selected using a feature based selection methodology selected from the group consisting of i) a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology; ii) a particle swarm optimization feature selection methodology, iii) a Gini-index-based feature selection methodology, iv) an entropy-based feature selection methodology, v) a chi- square-based feature selection methodology, and vi) a forward feature selection methodology.

10. The method according to claim 9, wherein when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise: the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 16 to mean value of channel 18, the ratio of mean value of top 10% of channel 22 to mean value of channel 18, the ratio of mean value of channel 15 to mean value of top 40% of channel 6, the ratio of mean value of top 10% of channel 20 to mean value of channel 19, the ratio of mean value of channel 14 to mean value of channel 13, the ratio of mean value of top 10% of channel 34 to mean value of top 10% of channel 31 , the ratio of mean value of top 10% of channel 12 to mean value of channel 16, the ratio of mean value of channel 8 to mean value of top 40% of channel 24, and the ratio of mean value of channel 10 to mean value of channel 9.11 . The method according to claim 9, wherein when the feature selection methodology is a forward feature based selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 22 to mean value of top 10% of channel 28, the ratio of mean value of channel 6 to mean value of top 10% of channel 8, the ratio of mean value of channel 2 to mean value of top 10% of channel 34, the ratio of mean value of top 10% of channel 10 to mean value of top 10% of channel 17, the ratio of mean value of channel 5 to mean value of channel 29, the ratio of mean value of channel 16 to meanvalue of top 10% of channel 12, the ratio of mean value of top 10% of channel 16 to mean value of channel 17, the ratio of mean value of channel 12 to mean value of channel 9, and the ratio of mean value of top 10% of channel 15 to mean value of channel 27.

12. The method according to claim 9, wherein when the feature selection methodology is a forward feature based selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA comprise: the minimum intensity value of channel 2, the ratio of mean value of channel 16 to mean value of top 40% of channel 24, the ratio of mean value of top 10% of channel16 to mean value of top channel 18, the ratio of mean value of channel 9 to mean value of channel 10, the ratio of mean value of channel 14 to mean value of channel 32, the ratio of mean value of channel 24 to mean value of channel 17, the ratio of mean value of channel 14 to mean value of top 40% of channel 20, the minimum intensity of channel13, the ratio of mean value of top 10% of channel 27 to mean value of top 10% of channel 16, and the ratio of mean value of channel 26 to mean value of top 10% of channel 17.

13. The method according to claim 9, wherein when the feature selection methodology is a particle swarm optimization based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise: the ratio of mean value of top 10% of channel 3 to mean value of channel 4, the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of channel 17 to mean value of channel 18, the ratio of mean value of channel17 to mean value of top 40% of channel 3, the ratio of mean value of top 10% of channel 16 to mean value of channel 15, the ratio of mean value of channel 24 to mean value of top 40% of channel 4, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel 21 , the ratio of mean value of channel 17 to mean value of channel 6, the ratio of mean value of channel 8 to mean value of top 40% of channel 24, and the ratio of mean value of channel 6 to mean value of channel 4.

14. The method according to claim 9, wherein when the feature selection methodology is a Gini-index based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative featuresfor differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise: the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 17, the ratio of mean value of channel 21 to mean value of top 10% of channel 19, the ratio of mean value of top 10% of channel 21 to mean value of channel 17, the ratio of mean value of channel 4 to mean value of channel 27, the ratio of mean value of top 10% of channel 15 to mean value of top 10% of channel 19, the ratio of mean value of channel 4 to mean value of top 40% of channel 31 , the ratio of mean value of top 10% of channel 16 to mean value of top 10% of channel 15, the ratio of mean value of channel 5 to mean value of channel 19, and the ratio of mean value of channel 5 to mean value of channel 33.

15. The method according to claim 9, wherein when the feature selection methodology is a chi-square based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus GR comprise: the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of channel 25 to mean value of channel 26, the ratio of mean value of top 10% of channel 15 to mean value of channel 17, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel 21 , the ratio of mean value of top 10% of channel 18 to mean value of channel 19, the ratio of mean value of channel 24 to mean value of top 40% of channel 4, the ratio of mean value of channel 5 to mean value of channel 19, the ratio of mean value of top 10% of channel 19 to mean value of channel 26, the ratio of mean value of channel 4 to mean value of channel 27, and the ratio of mean value of channel 17 to mean value of channel 18.

16. The method according to claim 9, wherein when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of channel 25 to mean value of top 10% of channel 22, the ratio of mean value of channel 22 to mean value of top 10% of channel 17, the ratio ofmean value of channel 26 to mean value of channel 12, 25thpercentile value of channel20, the ratio of mean value of top 10% of channel 3 to mean value of top 10% of channel 15, the ratio of mean value of channel 14 to mean value of channel 13, the ratio of mean value of channel 16 to mean value of top 40% of channel 3, the ratio of mean value of top 10% of channel 6 to mean value of top 10% of channel 15, and the ratio of mean value of top 10% of channel 22 to mean value of channel 33.

17. The method according to claim 9, wherein when the feature selection methodology is a particle swarm optimization feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of channel 23 to mean value of top 10% of channel 15, the ratio of mean value of top 10% of channel 17 to mean value of top 10% of channel21 , the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 16 to mean value of channel 15, the ratio of mean value of top 10% of channel 28 to mean value of top 10% of channel 6, the ratio of mean value of top 10% of channel 24 to mean value of channel 15, the ratio of mean value of top 10% of channel 19 to mean value of top 10% of channel 21 , the ratio of mean value of top 10% of channel 19 to mean value of top 10% of channel 22, and the ratio of mean value of channel 33 to mean value of top 10% of channel 2.

18. The method according to claim 9, wherein when the feature selection methodology is a Gini-index based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of top 10% of channel 15 to mean value of channel 27, the ratio of mean value of top 10% of channel 12 to mean value of top 10% of channel 19, the ratio of mean value of channel 2 to mean value of channel 24, the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 31 , the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of channel 12 to mean value of channel 19, the ratio of mean value of top 10% of channel 16 to mean value of channel12, the ratio of mean value of channel 7 to mean value of channel 17, and the ratio of mean value of channel 8 to mean value of top 10% of channel 22.

19. The method according to claim 9, wherein when the feature selection methodology is an entropy-based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of channel 8 to mean value of top 10% of channel 33, the ratio of mean value of top 10% of channel 9 to mean value of channel 33, the ratio of mean value of channel 32 to mean value of top 40% of channel 17, the ratio of mean value of channel 7 to mean value of top 10% of channel 33, the ratio of mean value of top 10% of channel 16 to mean value of channel 19, the ratio of mean value of channel 33 to mean value of top 40% of channel 17, the ratio of top 10% of mean value of channel 8 to mean value of channel 33, the ratio of mean value of top 10% of channel 12 to mean value of top 10% of channel 30, and the ratio of mean value of top 10% of channel 22 to mean value of channel 33.

20. The method according to claim 9, wherein when the feature selection methodology is a chi-square based feature selection methodology, and the classification method is a linear support vector machine-learning classifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between ATN versus NR-IFTA comprise: the ratio of mean value of top 10% of channel 4 to mean value of channel 24, the ratio of mean value of top 10% of channel 15 to mean value of channel 34, the ratio of mean value of channel 15 to mean value of top 40% of channel 28, the ratio of mean value of top 10% of channel 18 to mean value of top 10% of channel 17, the ratio of mean value of top 10% of channel 27 to mean value of channel 7, the ratio of mean value of channel 24 to mean value of channel 15, the ratio of mean value of channel 2 to mean value of channel 24, the ratio of mean value of channel 33 to mean value of channel 15, the ratio of mean value of top 10% of channel 21 to mean value of channel 32, and the ratio of mean value of top 10% of channel 4 to mean value of top 10% of channel 32.21 . The method according to claim 9, wherein when the feature selection methodology is a Minimum Redundancy Maximum Relevance (MRMR) based feature selection methodology, and the classification method is a linear support vector machine-learningclassifier, the 10 optimised quantitative features for differentiating the histopathological cause of kidney graft dysfunction between GR versus NR-IFTA comprise: the ratio of mean value of channel 26 to mean value of top 10% of channel 22, the ratio of mean value of channel 28 to mean value of channel 13, the ratio of mean value of channel 28 to mean value of channel 18, the ratio of mean value of channel 28 to mean value of top 40% of channel 29, the ratio of mean value of top 10% of channel 20 to mean value of channel 12, the ratio of mean value of channel 27 to mean value of channel 17, the ratio of mean value of channel 22 to mean value of top 40% of channel 26, the ratio of mean value of top 10% of channel 3 to mean value of channel 27, the ratio of mean value of channel 24 to mean value of channel 17, and the ratio of mean value of channel 9 to mean value of channel 10.

22. The method according to any one of claims 1 to 21 , further comprising the step of: e) applying a second classification method to the optimised quantitative features of the multispectral profile for further differentiating the histopathological cause of kidney graft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.

23. The method according to claim 22, wherein the second classification method is an auto machine-learning (AutoML) classifier.

24. The method according to any one of claims 1 to 23, wherein said urinary exfoliated kidney cells are proximal tubule cells (PTCs).

25. A system for assessing a histopathological cause of kidney graft dysfunction in a subject following kidney transplantation, the system comprising: a multispectral excitation lamp configured to excite the single photon-excited autofluorescence signal of one or more kidney cells obtained from a urine sample from a subject in a number of defined narrowband excitation wavelength ranges, one or more epifluorescence filters, a detector configured to detect the native fluorescence emission from said one or more urinary exfoliated kidney cells at multiple specified wavelengths, and a processing system configured to: a) generate one or more images of the native fluorescence emission from the one or more urinary exfoliated kidney cells, b) calculate, for each urinary exfoliated kidney cell, quantitative features of the autofluorescence signals in said one or more images,c) apply a feature selection methodology to the quantitative features to generate a set of optimised quantitative features; and d) apply a classification method to the optimised quantitative features to produce a multispectral profile for differentiating the histopathological cause of kidney graft dysfunction in the subject from a plurality of histopathological causes of kidney graft dysfunction.

26. The system according to claim 25, wherein the processing system is further configured to: perform image pre-processing; calculate, for each kidney cell, quantitative features of the measured autofluorescence; remove correlations between the calculated quantitative features of different kidney cells; and project, for each kidney cell, the quantitative features of the measured autofluorescence onto a new vector space.

Images