Towards label-free non-invasive autofluorescence multispectral imaging for melanoma diagnosis

This study focuses on the use of cellular autofluorescence which visualizes the cell metabolism by monitoring endogenous fluorophores including NAD(P)H and flavins. It explores the potential of multispectral imaging of native fluorophores in melanoma diagnostics using excitation wavelengths ranging from 340 nm to 510 nm and emission wavelengths above 391 nm. Cultured immortalized cells are utilized to compare the autofluorescent signatures of two melanoma cell lines to one fibroblast cell line. Feature analysis identifies the most significant and least correlated features for differentiating the cells. The investigation successfully applies this analysis to pre-processed, noise-removed images and original background-corrupted data. Furthermore, the applicability of distinguishing melanomas and healthy fibroblasts based on their autofluorescent characteristics is validated using patient cells with the same evaluation technique. Additionally, the study tentatively maps the detected features to underlying biological processes. This research demonstrates the potential of cellular autofluorescence as a promising tool for melanoma diagnostics.

with advanced melanoma [1,2].The risk of developing melanoma can be influenced by genetic risk factors such as fair skin, light hair color, blue eyes, and freckles coupled with prolonged exposure to sun light and UV radiation [3,4].These factors contribute to high incidence rates in Australasia, Europe, and North America which accounted for 275 000 out of the 325 000 reported melanoma cases in 2020 [5,6].However, when diagnosed early, treatment advances such as immunotherapy improved the 5-year survival rate to exceed 90% [2,5,7,8].The standard skin screening procedure for skin lesions including melanoma is visual inspection through dermoscopy, and its accuracy depends on the examining physician's knowledge and experience [9].In addition, rare melanoma types with atypical features such as hypopigmented amelanotic melanoma are particularly difficult to diagnose with standard methods [10][11][12].As a result, some skin changes may remain undetected without being referred to a follow-up biopsy for a conclusive histological assessment which are widely acknowledged as the definite benchmark for tumor diagnosis [13].Therefore, noninvasive accurate imaging technologies for rapid standardized melanoma assessment would be valuable to overcome these limitations.Optical detection methods such as reflectance confocal microscopy [14], optical coherence tomography [15], multiphoton imaging [16], and stepwise twophoton excited fluorescence [17] have been investigated for melanoma diagnostics.The listed techniques can improve melanoma diagnosis by providing additional information and increased resolution.However, for example, confocal imaging and multiphoton imaging are timeconsuming, costly, and prone to motion artifacts [18][19][20].Furthermore, a trained and experienced personal is required for proper handling and interpretation of the images, and there tend to be accessibility issues to certain body parts [19].Significant effort has also been invested in non-invasive image-based melanoma detection using conventional skin images analyzed by machine learning and artificial intelligence.However, varying accuracies for melanoma diagnostics have hindered its widespread clinical application [21,22].These studies highlight the need to develop a more reliable, cost-effective, and objective technique before replacing skin examination with the unaided eye and dermoscopy as standard methods of melanoma diagnosis [23,24].
The transformation of a normal cell into a cancerous one tends to be accompanied by changes and dysfunctions in cell metabolism [25].A common phenomenon with cancer cells is the Warburg effect, where aerobic glycolysis and glucose uptake are increased, and this has a role in cancer cell proliferation [26][27][28].The metabolic fingerprint of a cell can be reflected through native fluorophores present within the cell, making them potentially valuable biomarkers [29].Essential autofluorescent co-enzymes contributing to the cell metabolism include nicotinamide adenine dinucleotide (NAD(P)H) and flavins [26,30,31].The native fluorescence signal of NAD(P)H has excitation maxima at 290 and 351 nm and emission maxima at 440 and 460 nm [26].The autofluorescent emission of flavin adenine dinucleotide (FAD), an important flavins representative, peaks at 535 nm induced by 450 nm excitation light [26].These native fluorophores have been used to monitor metabolic processes due to their involvement in cellular respiration [32].In particular, the optical redox ratio (here defined as the ratio of the optical signals of flavins and NAD(P)H) can reflect the activity of the electron transport chain and ATP production [26,33].Changes in these fluorophores and the redox ratio may indicate a dysregulated cellular metabolism typical in cancer, and therefore, have high potential to be useful in oncology applications [26].
Essential autofluorophores potentially useful as biomarkers for skin cancer diagnostics include collagen, elastin, flavins, melanin, and NAD(P)H [34][35][36][37][38]. Published reports focusing on the assessment of skin lesions include work of Lihacova et al. [39] who observed reduced autofluorescence in melanoma compared to benign lesions and basal cell carcinoma (BCC) linked to altered collagen structures.Despite its efficacy in distinguishing melanoma, their technique proved ineffective for BCC.Shifting focus to non-melanoma skin cancers, Giovannacci et al. [40] used a fiber-based spectroscopy approach which similarly indicated reduced autofluorescence for both BCC and squamous cell carcinoma (SCC) compared to healthy skin.Romano et al. [41] employed fluorescence lifetime imaging (FLIM) to compare healthy tissue and BCC, focusing on the native fluorophores collagen, NADH, and FAD.They observed reduced fluorescence intensity and lifetimes across all channels as well as increased texture variability in malignant lesions.The tissue classification achieved an area under the receiver operating characteristic curve (AUC) of 0.82.Recent cell experiments by Garbarino et al. [42] comparing SCC cells versus normal keratinocytes supported these finding revealing lower autofluorescence intensity for SCC related to FAD, NAD(P)H, lipo-pigments, and porphyrins.
An altered ratio of free to bound NADH dependent on the size of the melanoma lesion was reported in Pastore et al. [43] who combined multiphoton microscopy and FLIM to study melanoma on a mouse model.They observed strong NADH signals with minimal detectability of FAD autofluorescence.Metabolic changes in melanoma cells which depend on culture conditions have also been reported in Ayuso et al. [44].Co-culturing melanoma cells with fibroblasts and keratinocytes resulted in a higher redox ratio compared to a single cell culture.
Whilst progress in using native fluorophores for diagnostic purposes has been made, challenges associated with utilizing autofluorescence for skin screening purposes include a corrupted field of view by blood, bacteria, inflammation, or scarring.Such hindered fields of view can therefore result in false positives [45].Nevertheless, Tamoši unas et al. [46] highlighted the use of autofluorescence in analyzing complex autofluorescence dynamics in postoperative scar tissue of removed BCC utilizing NADH levels to localize tumor reoccurrence.
These studies collectively highlight the potential of native fluorescence in effectively distinguishing various skin conditions, particularly melanomas.Additionally, the research emphasizes the significance of a multichannel analysis, recognizing the involvement of multiple native fluorophores.Furthermore, the importance of transferability into a cost-effective device with clinical applicability is emphasized.
The present study investigates the application of multispectral imaging of cellular autofluorescence for enhanced melanoma diagnostics.The technique exposes cells to various excitation wavelengths of light, where the emitted native fluorescence can be captured in distinct wavelength ranges, specific to each individual fluorophore.In this way, an expanded dataset is obtained by combining the different combinations of excitation and emission in a hyperspectral data cube, extending the information that can be inferred [47].We show that this imaging method can distinguish two melanoma cell lines and fibroblasts by detecting subtle changes in cell metabolism.Image processing has been used here to identify quantitative features, serving as a molecular fingerprint for cells.A feature selection algorithm has been designed to choose the most informative features for classifying healthy and cancerous cells, utilizing only the most relevant channels, and thereby facilitating accelerated measurements.Furthermore, the study assesses the requirement of image pre-processing on immortalized melanoma and fibroblast cells which hinder clinical translation.Clinical applicability of label-free non-invasive hyperspectral imaging of cellular autofluorescence is supported by a demonstration of successful melanoma diagnosis on patient melanoma samples and fibroblasts.
142BR cell line was subcultured and maintained in the complete culture medium Minimum Essential Medium Eagle (EMEM + non-essential amino acids, Sigma-Aldrich) combined with 2 mM glutamine (Gibco), 15% fetal bovine serum (FBS, Gibco), and 5 mL penicillin/streptomycin (P/S; 100 U/mL; Gibco).The melanoma cells COLO679 were cultured in RPMI 1640 with L-Glutamine (Gibco), 10% FBS, and 5 mL P/S.For A375, DMEM (high glucose, pyruvate, no glutamine; Gibco) was used along with 15% FBS and 5 mL P/S.Cells were incubated at 37 C 5% CO 2 incubator.Passaging of cells was performed once the confluency reached 80%.Cells were washed with phosphate buffered saline (PBS) and trypsinized with TrypLE (GIBCO, Australia, Catalog No: 12563-029).Following incubation with trypsin for 5 min at 37 C, the complete medium was added to trypsinized cells.The cell suspension was centrifuged at 500 g for 5 min.After removing the supernatant, the cell pellet was resuspended in the complete medium.Cell viability testing was performed using Trypan blue 0.4% (Sigma-Aldrich, Australia, Catalog No: T8154).

| Patient cells
Clinical samples of 10 patients were received under the permission of the Macquarie University Human Research Ethics Committee, reference number 5201400458.Upon receipt, the patient cells were cultured in Dulbeco's Modified Eagle Medium containing 10% heat inactivated FBS (Sigma-Aldrich), 11.25 mM glutamine (Gibco, Thermo Fisher), and 10 mM HEPES (Gibco) and maintained at 37 C and 5% CO 2 .

| Preparing cells for spectral imaging
To perform hyperspectral imaging, the cells were seeded into 35 mm plastic culture dishes with 18 mm well and # 1.5 cover slip bottoms (Cell E&G, USA, and # GDB0004-200).Each dish was seeded with 1 mL of cells (10 5 cells/mL) and incubated at 37 C, 5% CO 2 for 48 h to reach 70% confluence.
The cells were washed three times with PBS prior to imaging, then 4 mL of Hanks Balanced Salt Solution (HBSS) was added into each dish.A total of 6 dishes of 142BR (3xP27, 3xP28, 2xP30), 5 dishes of COLO679 (3xP11, 2xP13, 2xP21), and 5 dishes of A375 (5xP7) were analyzed resulting in a minimum of 300 imaged cells per group.
The same cell preparation protocol was used for imaging patients' cells.

| Experimental set-up
Multispectral imaging of cellular autofluorescence was carried out using fluorescence microscopes customized to measure cellular autofluorescence designed by Quantitative Pty Ltd.To this aim, regular fluorescence Olympus IX83 microscopes were retrofitted with a multi-LED light source (±5 nm) and filter cubes containing excitation, edge, and emission filters as described in [48,49] and schematically shown in Figure S1.This allows for measurements within an excitation range of 340-510 nm and emission ranging between 396 nm and 1200 nm.A cooled, highly sensitive camera reduced the noise when imaging the generated autofluorescence signals on a 40Â oil objective (NA 1.35).The experimental setup enabled the collection of brightfield images and multispectral images.
For immortalized cells, the imaging system was coupled with a Nüvü™ EMCCD camera HNü 1024.The imaging protocol defined a set of 28 spectral channels (see Table S1 for a list of channels).The channels covered excitation wavelengths ranging from 345 nm to 505 nm and captured emission wavelengths above 391 nm.To ensure optimal signal-to-noise ratios, exposure times and image averaging were adjusted individually for each channel.
For clinical samples, a multispectral microscope with an Andor IXON 885 EMCCD camera was employed.The imaging protocol for clinical samples utilized 38 channels (see Table S2).The excitation wavelengths ranged from 340 nm to 510 nm, while the emission filters covered the range of 420 nm to 650 nm.
To account for potential uneven illumination of the field of view and relate the multispectral images to reference fluorescence values measured by a FluoroMax 4 spectrofluorometer (HORIBA Scientific, Japan), a calibration fluid containing 3.75 μM NADH and 1.24 μM FAD was imaged producing calibration files.Moreover, a water image was used to eliminate background artifacts during image pre-processing, ensuring the accuracy and reliability of the experimental results.

| Image pre-processing
The multispectral images contain a range of artifacts such as cosmic ray-induced spikes, readout noise, Poisson's noise and background autofluorescence, in addition to the cellular autofluorescent signal.Therefore, these artifacts had to be removed before evaluation.High outlier values were eliminated based on the intensity histogram and replaced by the mean value of the surrounding pixels [50].Wavelet filtering with hard thresholding removed noise enhancing the image quality [29].For every channel i, Equation (1) was computed for each pixel (x, y) to receive a flattened and calibrated image Cell cal x, y, i ð Þto be smoothed in the subsequent step [29].
Water s and Cal s denote the respective smoothed water and calibration file and Cell displays the raw cell file.f i ð Þ is the calibration factor from the calibrated reference measurement with the sum of all channels normalized to 1. Shifting by the median value of manually defined background areas removed the remaining traces of autofluorescent signatures outside cells.
As this procedure would impede clinical applicability, the experiment on immortalized cells was additionally run without image pre-processing.

| Feature analysis
The feature analysis conducted in this study is directed towards the classification of data into the distinct groups of healthy and diseased cells, with a particular focus on identifying different melanoma subtypes when applicable.The analysis was performed on pre-processed images of manually segmented immortalized and patient cells, as well as the original images of immortalized cells.For each cell, the features mean intensity, channel ratios and products were generated.Furthermore, the mean value of the brightest 10% of pixels and statistical measures such as pixel variance, skewness, kurtosis, and entropy were computed, resulting in 924 features for immortalized cell lines and 1634 features for patient cells.A detailed mathematical description of the feature computation is given in the supplementary material of [48].To ensure robust data analysis, potential outlier cells are eliminated by removing the lower and upper one percentile of the overall performance values from the dataset.These cells are characterized by features that deviate from the general pattern of the dataset.
The implemented feature selection algorithm, as illustrated in Figure S2, identifies the most informative features for data grouping, leading to reduced exposure times and mitigating overfitting issues.The algorithm combines two methods: the ANOVA significance test with a designated feature significance threshold of p < 0:05 [51], and the Pearson correlation method which considers r-values between À0.3 and +0.3 as uncorrelated [51,52].In this study, the r-value is further restricted to the negligible range of À0.1 to +0.1 to ensure feature dissimilarity.The algorithm achieves an optimal sub-selection of features meeting these conditions by searching for a feature combination that minimizes the sum of their p-values, with a smaller r-value serving as a secondary constraint in the selection process.
For pre-processed immortalized cells, the two most informative features were selected, and the three most revealing features were chosen for the non-processed immortalized cells and the biologically variable patient cells.A support vector machine (SVM) classifier was trained using the selected features.This non-probabilistic binary classifier searches the maximal margin between two classes and marks the center with the separating hyperplane which optimally dividing the data [53].60% of the data was utilized to train the model while the remaining 40% of the dataset was allocated for validation, employing the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC).

| RESULTS AND DISCUSSION
For the immortalized cells 142BR (healthy fibroblasts), COLO679 (melanoma) and A375 (amelanotic melanoma), the algorithm identified two channel ratios as the most informative features (see Table 1).Their feature distribution is displayed in the box plots in Figure 1C,D.Figure 1A relates the two selected features demonstrating three distinct classes with each data point representing one observation.The ROC curve (Figure 1B) shows high values for the AUC with 142BR, COLO679, and A375 scoring, respectively 96%, 97%, and 100%.
Furthermore, the algorithm was applied to the original multispectral images without image pre-processing.The increased number of outliers in the feature distribution (shown in the boxplots of the three selected features in Figure 2C-E) indicates higher data variability than in preprocessed images.The three selected features are plotted in Figure 2A forming three separate classes.The predictive power of the classifiers is AUC = 89% for 142BR, AUC = 85% for COLO679, and AUC = 97% for A375.
Similar calculations were performed on clinical patient cells, where three distinctive features were selected as excellent representatives, as depicted in the box plots in Figure 3C-E.The 3D visualization in Figure 3A shows a clear grouping of normal healthy skin cells and melanoma cells.The AUC of 92.8% indicates strong classification accuracy (Figure 3B).
The measured cellular autofluorescence by multispectral imaging is largely determined by the fluorescence signals NAD(P)H and flavins.The redox ratio (flavins/NAD (P)H) or its inverse is represented as one selected feature in each of the three analysis: For immortalized cells T A B L E 1 Specifications of selected feature IDs (FIDs) for immortalized cells with (column 1) and without (column 2) image preprocessing and selected features for patient cells (column 3).analyzed after pre-processing (Figure 1), Feature ID 1 (FID 1) displays flavins/NAD(P)H, whereas the inverse can be found in FID 3 of immortalized cells without preprocessing (Figure 2).For patient cells (Figure 3), the redox ratio is displayed in FID 3 and its inverse in FID 2. Variation in the ratio of these autofluorophores can provide insight into the metabolism of the cells, including altered rates of glucose catabolism, and therefore, the Warburg effect observed in cancer cells [33], having the potential to be used as a marker of early cancer detection.The substantial variation in the optical redox ratio correlates with disparities observed in prior investigations examining melanomas [44].

FID
The experiments on immortalized and patient cells presented here indicate that autofluorescence may be a valuable optical biomarker for melanoma diagnosis.Based on a simple data analysis, distinct groupings of healthy and melanoma cells could be identified using a limited number of features (2 features for pre-processed immortalized cells, 3 features without pre-processing and patient cells).This reduced number of features also limits the necessary combinations of excitation-emission wavelengths within the hyperspectral data cube, referred to as spectral channels.Features distinguishing immortalized cells involve four spectral channels (ex 490 nm, em 558-604 nm/ex 371 nm, em 398-504 nm; ex 397 nm, em 665-1200 nm/ex 430 nm, em 665-1200 nm).For distinguishing healthy and cancerous cells without pre-processing, data from six spectral channels is utilized (ex 385 nm, em 558-604 nm/ex 403 nm, em 611-900 nm; ex 403 nm, em 558-604 nm/ex 476 nm, em 558-604 nm; ex 345 nm, em 396-437 nm/ex 490 nm, em 398-504 nm).Patient cells require five spectral channels (ex 391 nm, em 454-495 nm; ex 368 nm, em 573-613 nm/ex 391 nm, em 454-495 nm; ex 413 nm, em 573-613 nm/ex 373 nm, em 454-495 nm).The limited number of channels offers benefits for clinical applicability by reducing imaging time and data collection, consequently, allowing for a higher throughput and increased patient comfort.
We found that multispectral imaging of cellular autofluorescence achieves a highly accurate differentiation between fibroblasts and two melanoma types on immortalized cell lines both with and without pre-processing.In particular, amelanotic melanomas which are generally detected late due to their hypopigmentation were classified nearly perfectly in pre-processed images.This is important because image pre-processing steps, particularly background subtraction, is not feasible in in-vivo settings.Consequently, we analyzed unprocessed images without background subtraction to simulate real-world conditions which also resulted in three clearly separated groups omitting the need of the prior steps.Our evaluation showed that cancer autofluorescence is so significantly different from autofluorescence of healthy tissue that the inclusion of background autofluorescence from other sources such as the microscope-although methodologically suboptimal-does not affect diagnostic results, as apparent in strong data separation.This highlights the potential of multispectral autofluorescence as an effective melanoma diagnostic indicator with practical clinical utility.
The technology studied in this work has the potential to yield a simple handheld diagnostic tool complementing standard dermoscopy.It could easily be integrated with existing imaging methods offering advantageous additional information.Extensive research has focused on image classification for melanoma detection, often with varying accuracies [21,22].Combining AI-based melanoma detection with autofluorescent characteristics could enhance early-stage cancer diagnosis and address misclassification challenges.For transferring this technology to human skin tissue, the impact of cell immortalization [54] and potential interference of melanin absorbance across the visible wavelength range [55] when imaging hyperpigmented melanoma must be considered.

F I G U R E 1
Results of feature analysis after pre-processing for multispectral imaging of healthy fibroblast cells 142BR and the melanoma cell lines COLO679 and A375; (A) 2D feature space; (B) receiver operating characteristics (ROC) and area under the curve (AUC) for classifier evaluation; (C) box plot of feature ID 1 to 2 (FID 1-2) (***p < 0.0001).

F I G U R E 2
Feature analysis without pre-processing multispectral imaging of healthy fibroblast cells 142BR and the melanoma cell lines COLO679 and A375; (A) 3D feature space; (B) receiver operating characteristics (ROC) and area under the curve (AUC) for classifier evaluation; (C-E) box plot of feature ID 1 to 3 (FID 1-3) (***p < 0.0001).
Overall, this work shows that label-free non-invasive multispectral imaging of cell autofluorescence could greatly contribute to early-stage melanoma diagnosis by combining regular imaging with insights into cell metabolism.expertise.This work was partially supported by the Australian Research Council Centre of Excellence for Nanoscale Biophotonics CE14010003.Aline Knab, Shannon Handley, and Abhilash Goud Marupally acknowledge the PhD scholarship support from University of New South Wales.Dr. Habibalahi is supported by Cancer Institute NSW early career fellowship (2021/ECF1291).Open access publishing facilitated by University of New South Wales, as part of the Wiley -University of New South Wales agreement via the Council of Australian University Librarians.