Quantitative Phase Imaging for Label-Free Analysis of Cancer Cells — Focus on Digital Holographic Microscopy

To understand complex biological processes, scientists must gain insight into the function of individual living cells. In contrast to the imaging of fixed cells, where a single snapshot of the cell’s life is retrieved, live-cell imaging allows investigation of the dynamic processes underlying the function and morphology of cells. Label-free imaging of living cells is advantageous since it is used without fluorescent probes and maintains an appropriate environment for cellular behavior, otherwise leading to phototoxicity and photo bleaching. Quantitative phase imaging (QPI) is an ideal method for studying live cell dynamics by providing data from noninvasive monitoring over arbitrary time scales. The effect of drugs on migration, proliferation, and apoptosis of cancer cells are emerging fields suitable for QPI analysis. In this review, we provide a current insight into QPI applied to cancer research.


Introduction
A number of techniques are known to qualitatively convert phase changes to observable amplitude variations.Developed in the 1930s, the Phase Contrast (PC) microscope [1], uses a spatial filter and a phase plate to translate phase into intensity modulation.Since then, PC microscopy and its derivatives, Differential Interference Contrast (DIC) microscopy and Hoffman Modulation Contrast (HMC) microscopy, have become widely adopted as techniques for cellular observation.DIC uses two polarized light beams, which take slightly different paths through the microscopic sample.As their optical path lengths differ and as the beams are recombined, they interfere and thereby create a shadow effect at refractive index gradients [2].HMC uses an off-axis slit plate and a spatial modulator to alternately attenuate or enhance scattered light, thus creating a bright and dark shadowing effect at refractive index gradients, whereas image areas of constant refractive index appear grey [3].However, the images are non-quantitative, neither optimized for visualizing cells nor for performing measurements.
Quantitative phase imaging (QPI) methods use interference techniques to convert the phase information into pixel intensity, and in that way directly record the quantitative phase delay.Using QPI, studies associated with both thickness and refractive index fluctuations can be performed [4].Previous work has been reviewed by Kemper et al., who described new ways of monitoring the cellular morphology changes in response to drugs [5], and by Kim et al. who presented a review of digital holography (DH), with emphasis on microscopy techniques and applications [6].
There are a number of different technical implementations of QPI; and, by convention, a variety of names have been used to describe what is here collectively referred to as QPI.Terms differ even in some cases for the same commercial product (Table 1).In this review, we use the term QPI to comprise all, and we let other notations denote specific variations of hardware configuration.Optical methods are well suited for studying growth of both adherent single cells and cell clusters [15].The study of single cells over longer periods of time requires non-disturbing measurement techniques, such as QPI [51].QPI provides a quantitative measurement of intracellular structures without the need for artificial dyes or markers.By using only low intensity illumination, it is possible to perform imaging continuously and repeatedly at high capture frequency over the course of days or even weeks.Consequently, biological processes and live cell dynamics, which may occur infrequently, can still be captured at high sampling rate.If wide-field imaging methods are used, it is also possible to capture data of a large number of cells simultaneously.When combined with the ability to segment cell areas and perform temporal profiling of the image information of each cell, it is possible to monitor a large number of living cells over time.Indeed, in vitro studies monitoring the effects of new drugs on proliferation and migration of cancer cells are performed on a cell population level.

Technology
A complete description of the details of all possible technical setups of QPI is beyond the scope of this review; see Popescu et al. [4], Lee et al. [67], and Majeed et al. [68] for an overview of the details of commonly used technical setups, and Holden et al. [69] for an overview of available commercial systems.Important to note is that the different setups come with slightly different benefits and limitations, making them inherently well suited to a variety of purposes: High speed acquisition for high temporal resolution of flow imaging [70], in situ imaging of growing adherent cells [21,37,56], live cell tomography [71,72], or incubator monitoring [30,73].
Two of the most commonly used hardware configurations are DH in a separate path off-axis Mach-Zender configuration [4] and Spatial Light Interference Microscopy (SLIM) [74].In this review, we refer to them as DH microscopy and SLIM, respectively.

Principles of QPI
Apart from a few early studies of biological imaging mounted on reflective surfaces, all QPI set-ups for bio-imaging examine the sample using transmitted light (Table 1).Transmission QPI is based on the interference between the light entering and exiting the sample.A separation between refracted light of the sample (the object beam) and the reference light (the reference beam) can be achieved using either spatial or temporal modulation.There are two types of spatial modulation: In-line holography and off-axis methods.The first type is based on interference of a weakly scattered beam from the object and an un-scattered incident beam [74].For this optical instrument, the use of heavy algorithms is necessary for image reconstruction, because interferometric information overlaps with unwanted information.
In off-axis holography, the optical axis of the reference beam is slightly tilted with respect to the object beam.Thus, unwanted information can be spatially separated [75].DH is a common off-axis holographic technique for the QPI of cells [31].Here, a laser beam is spatially separated by a dichroic mirror.The object beam is diffracted by the cells and interferes with the reference beam at a slight angle at the recording sensor, where a fringe pattern characteristic of the object is created.The intensity distribution depends on the phase difference ∆ϕ [4].Temporal modulation, or phase shifting interferometry, involves changing the phase of the reference beam with respect to the object beam.An example of temporal modulation is SLIM [76].
The reconstruction of the phase and amplitude images from the recorded interferometric pattern varies with the hardware set-up.Typically, it consists of one or several reconstruction stages using Fourier Transforms and a focal selection stage, followed by one or more post-processing calculations where modular phase recordings are unwrapped and edge effects removed [4].The risk of introducing image artefacts in the unwrapping stage is reduced if the sample phase shift is less than one wavelength, or it can be avoided completely by imaging at several wavelengths, thus also allowing for imaging thicker specimens [73,77,78].All image reconstruction and manipulation stages can be performed after capture, as the complete three-dimensional information of the sample is stored in the interferometric recording.The result is an image (Figure 1b) where pixel intensity is directly proportional to the phase shift ∆Φ (Equation (1), with wavelength λ and refractive index in optical path direction n(z) as parameters.The image is well suited for automated image analysis for two reasons.Firstly, the image background can be used to offset the phase shift and allow comparison between images.Secondly, the images themselves display intensity gradients at cellular boundaries, making it possible to segment the image and extract information on each imaged cell separately (Figure 1c).using either spatial or temporal modulation.There are two types of spatial modulation: In-line holography and off-axis methods.The first type is based on interference of a weakly scattered beam from the object and an un-scattered incident beam [74].For this optical instrument, the use of heavy algorithms is necessary for image reconstruction, because interferometric information overlaps with unwanted information.
In off-axis holography, the optical axis of the reference beam is slightly tilted with respect to the object beam.Thus, unwanted information can be spatially separated [75].DH is a common off-axis holographic technique for the QPI of cells [31].Here, a laser beam is spatially separated by a dichroic mirror.The object beam is diffracted by the cells and interferes with the reference beam at a slight angle at the recording sensor, where a fringe pattern characteristic of the object is created.The intensity distribution depends on the phase difference Δϕ [4].Temporal modulation, or phase shifting interferometry, involves changing the phase of the reference beam with respect to the object beam.An example of temporal modulation is SLIM [76].The reconstruction of the phase and amplitude images from the recorded interferometric pattern varies with the hardware set-up.Typically, it consists of one or several reconstruction stages using Fourier Transforms and a focal selection stage, followed by one or more post-processing calculations

∆Φ
All image reconstruction and manipulation stages can be performed after capture, as the complete three-dimensional information of the sample is stored in the interferometric recording.Despite these appealing traits, several challenges persist.Thick cellular coverage may make it difficult to establish image background, and the choice not to use dyes will also reduce the ability to artificially enhance image contrast or specific image content, such as highlighting a specific molecule or intracellular structure.As a consequence, QPI images of very thin cells or cell structures may be undetectable due to noise.The phase delay measured is directly related to cellular dry mass and is the only physical property measured.Using image analysis, it is possible to extract further information using the recorded two-dimensional mass distribution within each cell.As the phase delay is a property not always translatable to known biological properties, it can be difficult to compare QPI directly to well-known legacy methods.Instead, QPI images provides a new way of looking at cells with its own distinctive characteristics.
The phase delay measured is directly related to cellular dry mass (Equation ( 1) and is the only physical property measured.However, using image analysis, it is possible to extract further information using the recorded two-dimensional mass distribution within each cell.A number of studies have compared the results of QPI analyses with other methods: QPI for tissue imaging and histological analysis have been investigated using tissue from breast [68] and prostate [76,79].The majority of QPI for studies of cell cultures have focused on cell count and/or cellular morphology (Table 1).As dry mass is independent of the intracellular water content, it can be a good indicator for cell growth.The ability to measure cell count was validated for the first time in 2008 by Mölder et al. [21], using human breast and prostate cancer cell lines, as well as a mouse fibroblast cell line.Measurement of cell growth has been well documented for both adherent and non-adherent cells [8,9,15,[25][26][27][28]47,49,56,59,61].By segmentation of cellular outlines, morphology and motility can also be studied [7,50,51,57,80].Furthermore, differences in cell types can be detected if they display distinctly different morphologies or migration patterns in response to cytotoxic drugs [30,81,82].

Cell Death Studies with QPI
The ability of QPI to differentiate between viable and non-viable cells without the use of staining has been extensively studied [21,83].The first DH microscopy images of tumor spheroids were published in 2003 [84], and the first images of single cells one year after that [85].Indeed, DH microscopy has turned out to be a very useful technique for cell death studies [21,83].Cells that go into programmed cell death, or apoptosis, show a decreased cell phase shift as the first indication of early apoptosis, but several morphological alterations take place, such as cell shrinkage, formation of small blebs, nuclear fragmentation, chromatin condensation, chromosomal DNA fragmentation, and breakdown into several apoptotic bodies [86].Pavillon et al. showed that dead cells could be identified within minutes through their DH phase signal, whereas with conventional trypan blue staining the identification took several hours [87].Also, DH microscopy was compared with electronic cell sizing and atomic force microscopy [88].DH was shown to be advantageous in terms of the noninvasive labeling, the time resolution, and the possibility of measuring both single cells and cell populations.Moreover, in a study by El-Schich et al., DH measurements revealed significant differences in the average cell number, the confluence, cell volume, and cell area when comparing untreated and etoposide-treated cells [50].We have cultured T leukemia Jurkat cells in ibidi chambers (ibidi GmbH, Martinsried, Germany) and treated the cells for 24 h with the cell death-inducing agent etoposide, or left untreated as a negative control.After the incubation, the cells were analyzed with DH microscopy.3D holograms are shown for untreated cells (Figure 2A) and for the etoposide-treated cells (Figure 2B).
Appl.Sci.2018, 8, x FOR PEER REVIEW 9 of 17 electronic cell sizing and atomic force microscopy [88].DH was shown to be advantageous in terms of the noninvasive labeling, the time resolution, and the possibility of measuring both single cells and cell populations.Moreover, in a study by El-Schich et al., DH measurements revealed significant differences in the average cell number, the confluence, cell volume, and cell area when comparing untreated and etoposide-treated cells [50].We have cultured T leukemia Jurkat cells in ibidi chambers (ibidi GmbH, Martinsried, Germany) and treated the cells for 24 h with the cell death-inducing agent etoposide, or left untreated as a negative control.After the incubation, the cells were analyzed with DH microscopy.3D holograms are shown for untreated cells (Figure 2A) and for the etoposidetreated cells (Figure 2B).Interestingly, several DH parameters that were measured, including cell number, area, thickness, and volume, were decreased after Etoposide treatment of the Jurkat cells (Table 2).[81].Interestingly, results presented on the individual cellular level of prostate cancer cells and mouse fibroblast cells treated with cell death-inducing drugs showed that these cell types behaved very differently [82].Using rapid profiling with high-speed live-cell interferometry, Huang and co-workers recently showed that the proto-oncogene B-Raf (BRAF)-inhibitor sensitive melanoma cell lines differed in measurements of optical cell biomass [62].Interestingly, within 24 h, biomass kinetic signatures were obtained for three Interestingly, several DH parameters that were measured, including cell number, area, thickness, and volume, were decreased after Etoposide treatment of the Jurkat cells (Table 2).[81].Interestingly, results presented on the individual cellular level of prostate cancer cells and mouse fibroblast cells treated with cell death-inducing drugs showed that these cell types behaved very differently [82].Using rapid profiling with high-speed live-cell interferometry, Huang and co-workers recently showed that the proto-oncogene B-Raf (BRAF)-inhibitor sensitive melanoma cell lines differed in measurements of optical cell biomass [62].Interestingly, within 24 h, biomass kinetic signatures were obtained for three pairs of sensitive or resistant melanoma cell lines, which showed a discrimination.
Cell death and cell proliferation are stringently controlled by the cell cycle, and studies have shown the ability of QPI to identify the different phases of the cell life cycle, i.e., specific changes in cell phase correlating to either a G1 or a G2/M arrest [13,25,38,44,57,89].Kamlund et al. constructed individual cell trees showing the way in which the drug salinomycin affects individual cells cultured in normoxia or hypoxia, by monitoring the times to first and last division [13].In this way, sensitivity to different conditions could be analyzed.Moreover, several studies of cell cycle utilize the fact that cells in mitotic arrest, G2/M, increase in cellular thickness and cell volume [44,89].

Choice of Cell Culture and Imaging Vessel
QPI measures the path-length shifts associated with the dry mass of the specimen [4- 6,67,90,91].The elimination of the contrast agent removes one cause of sample variability as the sample is observed directly.However, there is no longer a possibility of artificially highlighting specific properties in situ, a situation which introduces new challenges in evaluating the relationship between imaging signal and cell pathology.As with all interferometric methods, QPI is highly sensitive to noise and contamination in the light path.Consequently, the choice of imaging chamber, cell type and physical growth conditions is of the utmost importance.QPI has optimal function on phase objects, thin objects with little light attenuation, and small differences between the refractive index of sample and background.The majority of studies have so far used a few common cell lines well suited to produce QPI signal (Table 1).The choice of cell culture chamber is more varied, with both commercial glass and plastic containers being used, as well as "homemade" chambers.Since the image is calculated from an interference pattern in a Fourier domain, any contamination which can cause the light to refract (other than the sample itself) will appear in the final image as an overall increase in noise levels, and not only as a singular point of contamination.In particular, a separate path configuration, e.g., the Mach-Zender, is especially vulnerable, as contamination can corrupt each path separately.
The main benefit of QPI is its non-invasive characteristics, thus making it possible to perform studies over long periods of time.Typically, some further restrictions are placed on the culture chambers used, as normal cell culture at these time scales requires regular change of the cell culture medium.Consequently, long term studies have used various kinds of perfusion chambers to be able to replace the medium without disturbing the imaging view [30,32,39,43,49,55,61].Several studies have also used larger culture chambers with large enough volume to avoid changing the medium entirely, such as by using a T25 cell culture flask (Table 1).Glass and plastic have been used in the majority of cases, but studies have shown that thin films mounted on glass are acceptable from the perspective of optical quality [92].

Automated Analysis and QPI
A promising potential for cellular imaging is the ability to use QPI to combine the non-invasive full-field imaging at short imaging intervals with automated analysis of spatio-temporal cell signatures, which enables the gathering of data from a large number of individual cells for long periods of time, typically several cell cycles [93].Given a stable segmentation of cell area and tracking of cell position, it is possible to monitor morphological cell parameters over time, while at the same time recording cell position and movement.When phase shift images are segmented, a large number of cellular features related to morphology, density, texture, and motility can be calculated for each individual cell.The highly stochastic nature of cell phenotypes results in high variability between systems to distinguish between different cell types and cell stages.Comparisons are often only possible using population averages or large sample sets.Consequently, a successful analysis of individual cells will require some degree of automation.Cellular and tissue dynamics were measured nearly ten years ago by studying speckle fluctuations caused by cellular and sub-cellular motions [94,95], and several studies have since investigated the characterization and separation of blood cells [96,97].Abnormalities of erythrocytes are important markers for disease.For example, Moon et al. used QPI obtained by off-axis DH microscopy to automatically calculate many useful parameters of red blood cells, using algorithms to provide joint statistical distributions of the cell properties [98].Simple and fast methods are needed urgently because morphology and functionality can differ quickly over time in a blood sample.Go et al. used digital in-line holographic microscopy combined with machine learning algorithms to classify different types of erythrocytes [51].
Recent advances in image analysis and machine learning classification algorithms applied to various types of medical imaging [99] have only just started to be applied to QPI.A typical approach for automated analysis consists of two major steps: The extraction of features from images, and the classification of cells into groups on the basis of the features.Studies have shown the ability of QPI to distinguish between cancer cell types in blood using relatively few basic features such as maximum and average intensity, and cell diameter [20,100].However, the majority of studies use more features.The inline-DH microscopy method used by Singh et al. [19,20] to achieve large-scale fingerprinting capabilities was useful for characterizing circulating tumor cells (CTC) in bulk flow.The authors were able to characterize tumor cell lines with different metastatic potential, and to distinguish drug resistant tumor cells from their normal counterparts.
The first studies of classification of adherent cells were published in 2017 for applications of drug exposure [51] and cell cycle state [57], both relying on a few dozen simple features relating to size and shape (e.g., area, eccentricity, average, and maximum phase shift) and texture (e.g., energy, homogeneity, and entropy).Assigning a direct biological interpretation to QPI-derived features can be difficult as intracellular fluctuations in phase shift could be due to variations in cell thickness, as well as density or changes in internal organelle arrangement.Moreover, several features can be expected to show high correlation because they are all collected from the phase shift profile, thus providing redundant information.The clear benefit of studying adherent cells instead of cells in flow is the ability to simultaneously gather information on cellular morphology, migration, and motion over time.The time spent for each analysis will be limited by the expected speed of the analyzed process.
Hejna et al. [57] recently identified 26 digital features referred to as quantitative digital holographic cytometry (DHC).When trained on these features, machine-learning algorithms achieve blind single cell classification with up to 95% accuracy.Indeed, the choice of classification strategy is of vital importance to the outcome of the analysis, as each strategy balances a tolerance for false positives and negatives with computational complexity and speed.Classifiers can be binary, sorting cells into distinct classes, or proportional, assigning a probability to a cell to belong to a specific class.
In supervised machine learning, a model classification algorithm is trained using a training sample set consisting of training data labeled, classified, or annotated manually to provide a ground truth.When training is complete, the obtained model is validated on test data of unknown class.The more representative the training data is to the expected test data, the more accurate the classifier In many real-world scenarios, test data may differ significantly from the training set.In the case of building a model for clinical use, this means that each model must be possible to train on samples from one set of patients with known cancer types, while still retaining its classification accuracy on samples from other patients with unknown diagnosis.Achieving this robustness is crucial for any use of machine learning outside of the lab, and although several studies have shown the potential of various classification approaches to distinguish between cells of known types when trained on cells from the same individual [54] or on known cell lines [51], few studies so far have explicitly addressed this problem.Changes detected by image analysis were already visible on day 1 for 0.25 µM etoposide, whereas effects on viability, as detected with spectrophotometry assays, were detected only on day 3 for 5 µM etoposide concentration, leading to the conclusion that the morphological changes observed occur before and at lower concentrations than the reduction seen in cell metabolic activity or viability assays [51].Moreover, Roitshtain et al. used a low-coherence off-axis interferometric phase microscopy setup, which allows a single-exposure acquisition mode, and thus is suitable for quantitative imaging of dynamic cells during flow.After acquisition, the optical path delay maps of the cells were extracted and then used to calculate 15 parameters derived from the cellular 3D morphology and texture [54].In a very recent study, activated macrophages were analyzed and information of both cellular morphology and molecular content were collected with the help of the combination of Quantitative Phase Microscopy (QPM), Raman spectroscopy, and auto fluorescence imaging [101].Macrophages are immune cells, which are mostly analyzed with immunological methods.However, here the activation state of macrophages could be detected, in comparison with the cytokine secretion and intracellular expression of molecules related to the immune response.

Discussion
QPI applications including cell counting, migration, and morphology assays have become increasingly popular, but several challenges still persist.The morphological label-free analysis ability of QPI is a fast, automatic, and cost-efficient evaluation tool for analyzing quantitative parameters, including cell area, thickness, volume, population confluence, and cell count.The need for QPI applications in clinical cancer diagnostics and treatments is emerging.There is a demand for tools to classify cells, and to determine cell morphology, differentiation, proliferation, morphological changes of cells transfected with DNA or siRNA, cell death, and effects on cell movement-all in a high-throughput manner.Since QPI is performed on live cells without any labeling, the cells can be investigated with other methods-or the cells can be cultured for longer periods after the analysis.However, extensive image analysis is required due to the richness of information in each image.Analyzing blood samples or tumor biopsies containing small subpopulations of drug-resistant cancer cells is also emerging with the increased need for individualized treatment.Interestingly, by using QPI in combination with machine-learning, the features provide biologically independent information across a variety of mammalian cell state transitions that can be used to standardize holograms for the purpose of kinetic single cell cytometry.The training speed and computer resources can be limitations in automated QPI using machine-learning.Also, it will be important in the near future to share data and to develop algorithms.

Figure 1 .
Figure 1.Example of QPI image: (a) Interference pattern from digital holography (DH) recording showing cells in transmission mode as seen by the sensor.An interference pattern between object and reference wave has been disturbed by the refraction of cells in the object light path; (b) Reconstructed phase image using a Fresnel approximation, numeric refocusing, and unwrapping.Image intensity is proportional to optical path length through the cells; (c) Reconstructed phase image with segmented cell areas.Cell regions (white) have been separated from background (grey) using a simple thresholding.Scale: 200 × 175 µ m 2 , human prostate DU145 cells in a T25 flask.

Figure 1 .
Figure 1.Example of QPI image: (a) Interference pattern from digital holography (DH) recording showing cells in transmission mode as seen by the sensor.An interference pattern between object and reference wave has been disturbed by the refraction of cells in the object light path; (b) Reconstructed phase image using a Fresnel approximation, numeric refocusing, and unwrapping.Image intensity is proportional to optical path length through the cells; (c) Reconstructed phase image with segmented cell areas.Cell regions (white) have been separated from background (grey) using a simple thresholding.Scale: 200 × 175 µm 2 , human prostate DU145 cells in a T25 flask.

Table 2 .
Quantitative DH measurements of Jurkat cells treated with 250 µ M etoposide for 24 h.
Colomb et al. have shown details of the apoptotic process, where the apoptotic blebbing in prostate cancer cells was clearly visualized by DH microscopy

Table 2 .
Quantitative DH measurements of Jurkat cells treated with 250 µM etoposide for 24 h.

µm 2 ) Av. Cell Optical Thickness Avg (µm 2 ) Av. Cell Optical Volume (µm 3 )
Colomb et al. have shown details of the apoptotic process, where the apoptotic blebbing in prostate cancer cells was clearly visualized by DH microscopy