2.1.

First-Generation PWS Instrumentation for Quantitative Analysis of Intracellular Disorder ($L_d$)

This initial design of the PWS microscope was based on several key components including a critical illumination system via a fixed low-numerical-aperture ($20\times$, $\mathrm{NA}=0.4$) objective lens (38-339, Edmund Optics, Barrington, New Jersey), a manual sample stage, and a spectral filtering system using a scanned 10-μm slit spectrometer (SP-2150i, Acton Research, Acton, Massachusetts).^16,21 With this system, collimated broadband light from a Xenon lamp (66902 100W, Oriel Instruments, Stratford, Connecticut) was focused on the sample by the low-numerical-aperture objective lens. Backscattered light was collected by the same objective lens, and the magnified image was focused on the slit of the spectrometer, which was mounted on an automated linear stage (T-LA60A, Zaber Technologies, Vancouver, British Columbia), to allow the slit to be scanned across the entire illuminated field ($\sim 200\mu \text{m}$). Images were collected with a CCD camera (Coolsnap HQ, Photometrics, Tucson, Arizona) with 6.45-μm pixels from the output of the spectrometer at each 10-μm step of the automated stage. Each image collected by the camera yielded an image with the $y$-axis representing the spatial $y$-dimension in the image plane and the $x$-axis representing the wavelength of light. By combining all the images collected at each $x$-spatial position, a three-dimensional data cube ($x,y,\lambda$) was formed representing the diffraction-limited spatial data as well as the sub-diffraction spectral data.

All PWS measurements yield backscattered intensity spectra corresponding to each spatial pixel in an image of a cell, which can be quantitatively analyzed to find a diagnostic biomarker for cancer called the disorder strength or $L_d$. $L_d$ quantifies the spatial variation of refractive index in the cell, which increases linearly as function of both the length scale and density of the intracellular refractive index fluctuations. Thus, increased $L_d$ values correlate with local macromolecular condensation events such as heterochromatin and euchromatin condensation in the nucleus.^22,23 The computation of disorder strength ($L_d$) is given by $L_d(x,y)\approx {\sigma }_n{l}_c$, in which $\langle l_c\rangle$ is the spatial ($x,y$) correlation length of the refractive index fluctuations, and $\langle {\sigma }_n\rangle$ is the standard deviation of the refractive index fluctuations.^15,16 In physical terms, the spatial correlation length, $\langle l_c\rangle$, corresponds to the size of intracellular structures causing refractive index fluctuations, and the standard deviation, $\langle {\sigma }_n\rangle$, is proportional to the density of the intracellular structures. This calculation generates an image of the $L_d$ values at each spatial position (pixel) in the sample. The total magnification of the system is such that the pixel size at the object plane is less than the diffraction-limited spot size, so $L_d$ values and their spatial distributions are independent of the hardware.

For the experiments described here, $L_d$ was calculated on the spectra between 500 and 700 nm. Prior to the $L_d$ calculation, each signal is normalized by the input illumination spectrum acquired from a mirror measurement, and a sixth-order low-pass Butterworth filter is applied to remove noise. A low-order polynomial subtraction was used to remove low-order slopes in the signal that result from differences in the lamp spectrum, sample roughness, and instrument artifacts. HTPWS $L_d$ values had to be scaled down by a factor of 3.37 to match $L_d$ values from the first-generation system to account for the reduction in spectral sampling frequency. For diagnostic purposes, cells were characterized by their mean intracellular $L_d$ and patients by the mean of all intracellular $L_d$ values from each measured cell. Statistical evaluation of diagnostic performance was performed in Excel (Microsoft, Redmond, Washington) and MATLAB® (MathWorks, Natick, Massachusetts). Mean $L_d$ values between different sample groups were compared using the two-sided Student’s $t$-test, and effect size was computed as Cohen’s $d$.

2.2.

Goals of the HTPWS Instrumentation

While the first-generation PWS microscope effectively demonstrated the modality’s potential as a clinical diagnostic tool, its design limited its effectiveness for large-scale high-throughput clinical studies and research. As described, the first-generation PWS system collects the backscattered image by linearly scanning the entire cell using the combination of the slit spectrometer and scanning stage, a process which takes approximately $3\min /\mathrm{cell}$. Measurements for an entire slide are slowed further, since the field of view is limited to approximately 120 μm requiring at least 2 to 3 min to manually find and focus on the cell. As a result, the total time to measure approximately 30 cells per patient is $\sim 4$ to 5 h. To improve upon the performance of the first-generation instrument, a new HTPWS system has been developed, as shown in Fig. 1, which reduces the time to measure patient samples and automates the data acquisition process. More specifically, to automate the process of finding and selecting cells for measurement that previously took more than an hour per slide, the HTPWS system used automated hardware to scan an entire slide at low magnification and to analyze the collected images to identify and locate the positions of cells. In addition, the slit-spectrometer approach to spectral filtering was replaced by an acousto-optic tunable filter (AOTF) to decrease data collection times and to eliminate the requirement of spatially scanning each image, significantly reducing data collection times from $3\min /\mathrm{cell}$ to less than $5\mathrm{s}/\mathrm{cell}$. While the HTPWS system fundamentally operates on the same principles as the first-generation device, the changes are detailed here in the sections pertaining to the tunable illumination hardware and optics, high-speed fully automated hardware, and custom software interfaces and algorithms used to achieve fully automated data acquisition.

Fig. 1

Schematic of the high-throughput partial wave spectroscopy (HTPWS) instrument. Tunable illumination is incident on the sample from an acousto-optic tunable filter (AOTF) with the illumination numerical aperture set by an electronic aperture. Backscattered reflectance is collected through a second electronic aperture that sets the collection numerical aperture with a high-speed CMOS camera. All data collection is automated via the acquisition GUI. Transmission bright-field image collection is also possible with the fiber-coupled LED source.

2.3.

Illumination Design

The HTPWS microscope uses a Köhler illumination alignment with tunable illumination to increase both uniformity of illumination and spectral sampling speed. Light from a Xenon lamp is focused through an AOTF (HSI-300, Gooch & Housego, Orlando, Florida). The AOTF has a minimum switching speed of 50 μs, bandwidth of 3 nm, and a spectral range from 450 to 800 nm. The tuned light exiting the AOTF is focused through an electronic motorized aperture (62281, Newport Corporation, Irvine, California) that sets the illumination numerical aperture. Light exiting the aperture is collimated and passed through the field aperture, after which it is focused onto the back focal plane of the objective lens ($40\times$, $\mathrm{NA}=0.6$ LUCPlanFL N, Olympus, Center Valley, Pennsylvania). This new illumination scheme achieves uniform intensity across the sample plane due to the Köhler alignment, and wavelength switching is less than 100 μs. In addition, because the incident illumination is tuned to a single monochromatic wavelength, this illumination system will allow future multimodal experiments combining fluorescence with HTPWS. These HTPWS experiments can be performed with molecularly specific dyes, which have been shown to enhance the refractive index of cellular organelles within the sample.²⁴

2.4.

High-Speed Automated Hardware

To shift from an entirely manual and user-intensive measurement process to automated high-throughput measurements, high-speed automated hardware was added to the system to replace each manual operation performed by the user. The new automated hardware is controlled via custom algorithms written by the authors in MATLAB®. First, the sample stage was upgraded with automated, encoded, linear stages (A-LSQ075B-E01, Zaber Technologies) for the $x$- and $y$-axes as well as an automated linear stage for the $z$-axis (T-LS28, Zaber Technologies). An automated objective turret was added to switch between the high-magnification ($40\times$) and low-magnification ($10\times$, UPlanFL N, Olympus) objectives. The angle of the collected backscattered light was controlled using a second motorized aperture before being focused on an ultra-fast CMOS camera (Hamamatsu ORCA-Flash 2.8, Bridgewater, New Jersey) with 3.63-μm pixels binned $2\times 2$. For maximum speed, synchronization of wavelength tuning and image capturing was done via hardware triggers between the AOTF and the camera; $2\times 2$ binning was enabled with the camera to maximize sensitivity and minimize exposure time for the fastest acquisition.

In addition to the PWS illumination, a transmission illumination arm was added to allow collection of bright-field images alongside PWS measurements. A white-light emitting diode [light emitting diode (LED)] fiber-coupled source (LE-1W-CE, WT&T, Lachine, Quebec) was connected to a fiber collimator, and the output beam was passed through a diffuser. This transmitted light was collected via a high-resolution scientific color CMOS USB camera (DCC1645C, Thorlabs, Newton, New Jersey) that was added to the system for this purpose. This camera was also used for rapid collection of low-magnification/low-resolution transmission images for slide mapping. A flipper mirror was used to switch between the transmission collection camera and the camera used for PWS measurements.

2.5.

Automated Slide Mapping

The first task performed during a HTPWS measurement is generating a large low-magnification (typically $10\times$ to $20\times$) image of the slide. This is accomplished using an algorithm that rapidly collects many low-resolution images and tiles them together to create the full image of the slide. A user defines the bounds of the region to be mapped by specifying the positions of two diagonal corners. The algorithm then calculates the number of images required to map the entire region specified based on a pixels-to-micron conversion factor specific to the objective and imaging sensor used. The region is then raster scanned, and an image is acquired at each $x$ and $y$ position in order to make a complete image of the region without gaps or overlaps. Finally, all the images are tiled together to form a complete image of the entire region. To maximize the speed of the algorithm, autofocusing (using a custom autofocus algorithm—discussed in Sec. 2.7) is performed only on the first image, and predictive autofocusing is used for all subsequent images. The focus is rechecked and corrected every 10 images as necessary throughout the scan. Images are stored in memory until the end of each scanning line, and then saved in order to prevent the algorithm from overloading the computer’s physical memory.

2.6.

Semi-Automated Cell Selection

Cell selection is performed using the large low-magnification image of the slide generated by the slide-mapping algorithm. This can be accomplished in one of two ways: a user can manually select cell positions using the mouse and saving them to a list for measurements, or image-segmentation algorithms unique for the cell type on the slide can be employed to automatically generate a list of candidate cells which are then shown to the user for approval. Manual selection allows the user to zoom-in and explore any region of a slide image and select positions anywhere on the image with a crosshair. Each selected position is saved to a positions list that is exported to the acquisition graphical user interface (GUI). In contrast, when an automated segmentation-based selection algorithm is running, the user is immediately prompted with a list of potential cell positions. Images of each potential cell and its local surroundings are displayed to the user, and a prompt asks whether to keep or reject the position. This semi-automated form of position selection has the advantage of being much faster, but it is limited to cell types for which an algorithm has been developed. We have already developed one such algorithm for stained buccal cells.

2.7.

Autofocus Algorithm

In order to perform both automated slide mapping and automated PWS measurements, a rapid autofocus algorithm was developed to accurately identify the correct focus plane at any position on the slide. Due to the unique needs of slide mapping and HTPWS measurements, the former requiring higher speed and the latter requiring greater accuracy, a two-step algorithm was developed. The first step is a predictive autofocus algorithm based on the equation of a plane in three dimensions. Three in-focus $x$, $y$, $z$ positions on the slide are collected to generate an equation for a plane that predicts the in-focus position anywhere on the slide. Between points where autofocusing is actually performed, this algorithm is used to predict the in-focus $z$-position during slide mapping. Thus, based on the equation of a plane, the predicted $z$-position is given by $z=\frac{d-ax-by}{c}$, where $a$, $b$, $c$, and $d$ are constants defining the equation of the plane in three-dimensional space, and $x$, $y$, and $z$ are spatial coordinates.

To more accurately get an in-focus image at a cell position for HTPWS measurements, an algorithm based on edge detection of the field aperture is used. The best focus is determined by the highest number of edges detected at the edge of the field. In this manner, focus consistency, which is critical to prevent variability in quantitative HTPWS $L_d$ analysis, is obtained by focusing on a fixed object that is always in the same position. The approach is similar to one proposed by Hughlett and Kaiser with the exception being that we use the edge of the field aperture instead of a shadow projection wire to quantify focus via an edge-detection technique.²⁵ Edges corresponding to the aperture are isolated by segmenting the field-of-view and by applying a black/white threshold to a Sobel gradient magnitude image of the field. A slight erosion of the segmented field leaves a mask that can be applied to images to obtain only edges corresponding to the field aperture.

To find an in-focus image, the algorithm must search for the $z$-position that corresponds to the maximum number of edges from the field aperture. It does this by scanning a user-determined range around the predicted focus position with large incremental steps ($\sim 5\mu \text{m}$). At each position, the number of edges is found using a Sobel edge detector. Because the number of edges forms a Gaussian curve with the in-focus position corresponding to the center of the peak, the algorithm detects when the number of edges switches from increasing to decreasing and stops scanning. The stage then backtracks in fine increments ($\sim 0.3\mu \text{m}$) to find the maximum number of edges corresponding to the aperture at the current $x$, $y$ coordinate. Figure 2 shows the difference in edges detected at the field aperture outside the mask for in-focus and out-of-focus images.

Fig. 2

From left to right: (a) In-focus image of a buccal cell and (b) corresponding edge map for the in-focus image showing the field aperture edges visible outside the border of the mask applied to remove the field. (c) Out-of-focus image of the same buccal cell and (d) corresponding edge map for the out-of-focus image showing no edges detected outside the border of the mask.

2.8.

Automated Spectral Measurements

Spectral measurements are completed on the HTPWS microscope via the automated measurement and analysis interface. The list of selected positions is automatically loaded, and the user sets the parameters for the scan including the wavelengths to scan, the illumination bandwidth, exposure time, input NA, and collection NA. Typical settings for a high-throughput scan include a spectral range of 450 to 700 nm with a step size of 1 nm. Input NA is typically set with the input aperture at 10%, approximating plane-wave illumination without sacrificing more light than necessary. Output NA is typically not constrained by the electronic aperture and is instead determined by the objective lens ($\mathrm{NA}=0.6$). When the scan is running, the system automatically moves to each position on the stored list, autofocuses, and spectrally scans the sample, collecting an image at each illumination wavelength. The result is a three-dimensional data cube ($x,y,\lambda$) identical to that provided by the first-generation system, but with the benefits of being fully automated and requiring much less time. Figure 3 illustrates the process required to complete an HTPWS measurement.

Fig. 3

Flow chart of the HTPWS measurement process including slide mapping, segmenting cell positions, collecting the spectral image cube, and generating an image showing the $L_d$ distribution within a cell.

3.1.

HTPWS System Performance

Performance with the new HTPWS microscope is dramatically improved compared with the first-generation system. A single HTPWS measurement can be completed in less than 5 s compared with 3 to 4 min for the first-generation system, an approximately 42-fold improvement. For a typical measurement of 30 cells, slide mapping is completed in 10 min. Manual cell selection adds another 10 min, while semi-automated cell selection using a segmentation algorithm can be achieved in less than 5 min. Finally, the spectral measurements are completed in 20 min. Thus, for a typical measurement, the entire process on a single slide for approximately 30 cells completes in approximately 40 min compared with 4 to 5 h for the first-generation system. The increases in performance and automation allow approximately seven patients to be measured on the HTPWS system for one on the first-generation system.

3.2.

Validation Experiments

In order to characterize the new HTPWS system, a series of experiments were performed to validate that its performance correlated with the previous data collected on the first-generation system. In particular, it was necessary to demonstrate that the new system measured the same spectral information as the first-generation system and that the HTPWS instrument was sensitive to the nanoscale properties of the samples measured. Validation of the spectral data was performed using a uniform 1-μm ${\mathrm{SiO}}_2$ thin-film reference standard (Filmetrics, San Diego, California). Measurements were taken on the same region of the thin film using both the first-generation spectrometer-based PWS system and the new HTPWS system. To account for the use of detectors with different sensitivities on each system, the exposures for each separate detector were set to give the same signal-to-noise ratios for the two systems. The spectra at specific pixels within the field-of-view from both measurements were then compared directly for a match within the same wavelength ranges used for PWS measurements (500 to 700 nm). Further spectral validation was also performed using more complex nonuniform samples. Spectra from the same individual latex microspheres of varying sizes (4.3 and 11 μm) that were allowed to dry out of solution on a glass slide (Thermo Scientific, Waltham, Massachusetts) were measured and compared for a direct match between individual pixels in the same locations. Finally, spectra from identical cells [buccal and HT29 (colon cancer) cell types] measured on both systems were plotted for a match at the same pixels and regions.

Spectral comparison between the HTPWS system and the spectrometer-based first-generation system showed consistently similar results from identical samples. Figure 4 shows the spectra generated from the same location on the ${\mathrm{SiO}}_2$ thin-film reference. These spectra were generated by averaging the spectra over a diffraction-limited area at the same location on the sample for both systems. A sixth-order Butterworth low-pass filter was applied to the signals from both systems to remove noise after normalizing by the spectra collected from a mirror. This result was compared with the spectra provided by the manufacturer of the thin-film reference, and there was an excellent agreement in terms of amplitude, slope, oscillation frequency, and phase. Figure 4 also shows a comparison of spectra averaged from pixels that make up the same diffraction-limited spot in an HT29 colon cancer cell. A match can clearly be observed between the spectra from the two systems at this location in the cell. In this case, perfect matching is much more difficult to achieve due to inhomogeneity of the sample, nonuniform sample topography, and differences in detector pixel sizes between the two systems. With perfectly uniform samples, such as the thin-film reference, the different pixel sizes of the detectors are not an issue, because the signal is the same at every point on the sample. On random samples, such as cells, the signal can vary significantly from pixel to pixel, making the difference in pixel size much more significant when matching spectra. Averaging spectra from the pixels that make up a diffraction-limited spot in each system helps to account for this and allows decent spectral matches to be obtained when the same location in a cell is analyzed.

Fig. 4

(a) Normalized and filtered 1-μm thick ${\mathrm{SiO}}_2$ thin-film reference spectra plotted from both the HTPWS system and spectrometer-based first-generation PWS instrument. The spectra are averaged over a diffraction-limited spot in the same location on the sample for both instruments, and a low-pass Butterworth filter is applied to remove noise. Root-mean-square error between the spectra was 0.01. (b) Normalized and filtered spectra comparison averaged at the same diffraction-limited spots in an HT29 colon cancer cell with both the HTPWS and first-generation PWS instrument.

To demonstrate and compare the nanoscale sensitivities of the first-generation PWS instrument with the new HTPWS instrument, nanoscale phantoms were created and measured. The phantoms were constructed using solutions of polymer nanospheres (Thermo Scientific). Separate phantoms were created corresponding to different length scales using spheres with diameters of 20, 40, 60, 80, 100, and 125 nm. Each phantom was made by applying a single droplet of the sphere solution to a glass slide and letting the spheres randomly assemble on the slide as the solution dried. This left behind a thin-layered structure of spheres that could be used to represent a random assortment of particles at a specific length scale. HTPWS measurements were taken from each phantom at the same locations on both systems. Twenty-five measurements were acquired from each phantom at different positions to allow for statistical comparison of the data. In order to compare phantoms comprised of spheres of different diameters, measurements were acquired in each phantom from regions of similar thickness based on the number of spectral oscillations (5 to 7 oscillations or 2.5 to 3.5 μm). For each phantom, 25 regions of interest were selected, and $L_d$ analysis was performed on the pixels in these regions. These $L_d$ measurements were then plotted as a function of the phantom nanosphere size to demonstrate sensitivity of $L_d$ to nanoscale length scales.

The measurements of the nanoscale phantoms performed on both systems were analyzed to calculate the $L_d$ value of each phantom. Figure 5 summarizes the results of this analysis, showing the sensitivity of $L_d$ to nanoscale length scales. The length-scale dependence of $L_d$ can clearly be observed as $L_d$ values show a steadily increasing trend with increasing diameter of the nanospheres making up the phantoms. Correlation between the length-scale of phantom spheres and $L_d$ is linear with an $R^2$ value of 0.93.

Fig. 5

Dependence of mean phantom $L_d$ as a function of the diameter of spheres making up the phantom. $L_d$ increases linearly with the size of the spheres making up the phantom. Error bars display the standard error of the mean $L_d$ value for each phantom.

3.3.

Diagnostic Performance Experiments

Two experiments were performed to test the diagnostic performance of the new HTPWS system. First, HT29 colon cancer cell lines were used to model a clinical diagnostic test. The experiment consisted of two groups, control vector HT29 (CV) cells and epidermal growth factor receptor (EGFR) knockdown HT29 cells, a less aggressive genetic variant. The HT29 control vector and EGFR knockdown cells were first collected in centrifuge tubes and centrifuged for 5 min at 1000 rpm. The supernatant was then removed, and the cells were plated on a glass chamber slide. The slides were checked to ensure that they contained at least 20,000 cells. Two milliliters of fresh cell culture medium was added to each chamber slide, which was then incubated at 37°C for at least 5 to 6 h. After incubation, the medium was completely removed from the chamber slides, and the slides were washed with 70% ethanol to remove any traces of the medium. Following this, the slides were immediately fixed using 70% ethanol and kept in a 4°C refrigerator until PWS measurements. Using this protocol, one slide each was prepared of control vector HT29 cells and one of EGFR knockdown HT29 cells. The two slides were measured unstained back-to-back on the first-generation spectrometer-based PWS system and the HTPWS instrument. The same 25 cells from each cell line were measured to allow for statistical comparison of the data. Previous PWS measurements of the CV and EGFR knockdown HT29 cells had shown significantly lower $L_d$ values associated with the EGFR knockdown cells compared with CV.¹⁶ Performing a second experiment with these HT29 cell lines on both the HTPWS system and the first-generation system yielded the same result. Figure 6 shows distribution, average $L_d$ values, and the corresponding effect size for the CV and the EGFR knockdown cells. Comparison of the results from both PWS systems shows similar effect sizes for the differences between the mean $L_d$ values for the CV and EGFR cell types, 1.16 for the HTPWS system and 1.23 for the spectrometer, respectively. $P$ values were also comparable with 0.0007 for the HTPWS instrument and 0.0002 for the spectrometer-based PWS instrument.

Fig. 6

$L_d$ measurements on CV and EGFR knockdown HT29 cell lines. The distribution of the cell $L_d$ values (a) and the mean $L_d$ values are shown (b). The CV cell line measured higher average values for $L_d$ compared with EGFR knockdown cells with $p=0.0007$ and an effect size of 1.16.

Clinical lung cancer diagnostic performance with the HTPWS system was also evaluated in a small experiment including 23 patients, consisting of 9 patients with cancer and 14 smokers. This human study was performed in accordance with the Institutional Review Board at NorthShore University HealthSystem. Cells were brushed from each patient’s cheek and smeared onto a glass slide before being fixed in 95% ethanol and stained using Papanicolaou stain just prior to measurement. For each patient, approximately 30 cells were measured and used to determine mean $L_d$ values for the individual patients as well as for each diagnostic category. Measurements were also performed on the first-generation system to correlate $L_d$ measurements between the two systems. Diagnostic performance of the system was represented by quantifying the difference in the mean $L_d$’s of the cancer and smoker patient groups using the data collected from all 23 patients. Average $L_d$ measurements were computed for each patient and for two groups, patients with cancer and smokers. Figure 7 shows the diagnostic results for the smoker and cancer groups. The cancer group had a significantly higher average $L_d$ compared with the smoker group as measured with the HTPWS instrument, $p=0.02$ and effect $\text{size}=1.00$.

Fig. 7

HTPWS lung cancer patient $L_d$ results. The distribution of the cell $L_d$ values is shown for the median patients in each group (a), and the mean patient $L_d$ values are plotted (b). Cancer patients had significantly higher average $L_d$ values compared with smokers, $p=0.02$ and effect $\mathrm{size}=1.00$.

Similar results to those in Fig. 7 were achieved with the first-generation PWS instrument. Cancer patients had significantly higher $L_d$ values than smokers with $p=0.03$ and effect $\text{size}=0.90$. To verify consistent results between the two systems, correlation between individual cell $L_d$ values and patient $L_d$ values was plotted for the two systems. Figure 8 shows the correlation between patient $L_d$ values for both systems. The correlation between the patient $L_d$ values for the two systems yielded $R^2=0.93$. For individual cell $L_d$ values, the correlation was $R^2=0.92$. The greater variance in the $L_d$ values observed for both groups in this study can be attributed to some focus error with this initial version of the automated measurement system and variability in slide quality due to sample collection, smearing, and storage protocols.

Fig. 8

Correlation of patient $L_d$ values between the HTPWS system and first-generation spectrometer-based PWS system ($R^2=0.93$).