Enhanced resolution 3D digital cytology and pathology with dual-view inverted selective plane illumination microscopy

: The current gold-standard histopathology for tissue analysis is destructive, time consuming, and limited to 2D slices. Light sheet microscopy has emerged as a powerful tool for 3D imaging of tissue biospecimens with its fast speed and low photo-damage, but usually with worse axial resolution and complicated configuration for sample imaging. Here, we utilized inverted selective plane illumination microscopy for easy sample mounting and imaging, and dual-view imaging and deconvolution to overcome the anisotropic resolution. We have rendered 3D images of fresh cytology cell blocks and millimeter- to centimeter-sized fixed tissue samples with high resolution in both lateral and axial directions. More accurate cellular quantification, higher image sharpness, and more image details have been achieved with the dual-view method compared with single-view imaging. deconvolution.


Introduction
For decades, histopathology has provided the reference standard for cancer diagnosis, prognosis prediction, and treatment decisions. In most cases, a biopsy or surgery is performed to extract cells or tissues for observation and analysis under a microscope. Usually, tissue biopsies are formalin-fixed and paraffin-embedded (FFPE), sectioned into thin slices, stained with dyes, and mounted on glass slides for examination by pathologists. This process has been the gold standard for tissue diagnosis for over a century. However, it is time-consuming, laborious, and destructive of tissue. Moreover, because of the preparation of the tissue and cellular material as thin sections or smears placed onto glass slides, the images are limited to 2D slices, which do not faithfully represent the true 3D architecture of the cellular or tissue material, and thus the original environment of the tumor.
The native 3D microarchitecture of tumors has been increasingly explored over the last two decades for better understanding of tumor structure. Many of these past studies have focused on prostate cancer, due to the specific importance of microarchitecture in its diagnosis and grading. For example, 3D prostate adenocarcinoma was first showed by A.H. Boag et al. [1] in 2001. They cut serial sections from archived paraffin-embedded prostate specimens of varying Gleason patterns (GPs) and Gleason scores (GSs) which were immunostained using DAB as the chromogen, digitized with a brightfield microscope, and reconstructed into a 3D volume. They found that instead of the separate glands seen in 2D, a complex array of interconnecting tubules was formed in GP 3, and a glandular fusion in GP 4. This was an excellent early attempt to analyze 3D prostate structures. However, because of the limited digital image resolution and computing resources at that time, resolving adjacent cells and precise edges of glands was a challenge. A similar study was performed by Tolkach et al. in 2018 [2], and important 3D architectural features were found to distinguish patterns GP 3 and GP 4, which is critical for prostate cancer treatment and prognosis. Although both studies have demonstrated that 3D imaging and reconstruction can be more informative than traditional 2D histology, the serial physical slicing approach used is laborious and sensitive to tissue deformations between consecutive slices. In 2016, Royen et al. [3] combined tissue clearing and confocal fluorescence microscopy to investigate the clinical and biological prostate cancer (PCa) features in the actual 3D context from intact tissue, and showed that the glandular system in GP 3 and GP 4 are highly interconnected. In 2017, Glaser et al. [4] used a variant of light sheet fluorescence microscopy (LSFM) to image an optically cleared prostate biopsy, and found that by showing pathologists serial optical sections derived from the 3D volume of prostate images, tangential sectioning artifacts can be distinguished from poorly formed glands, and inter-observer agreement of GS can be improved. Based on these findings, it is becoming increasingly evident that the 3D structure of tumors is important for cancer diagnosis, grading, and overall understanding of tumor microarchitecture, and thus could be a vital new addition for patients.
In addition to confocal [3,5,6] and LSFM mentioned above, multiple other imaging methods have also been utilized for non-destructive imaging of tissues, including optical coherence tomography (OCT) [7], microscopy with UV surface excitation (MUSE) [8,9], structured illumination microscopy (SIM) [10], and nonlinear microscopy [11,12]. While these methods may have significant cost, complexity, and speed advantages for 2D imaging compared to LSFM, for 3D tissue imaging they are limited by varying combinations of low resolution, low penetration depth, or a relatively slow imaging speed. In addition, all of the above-mentioned methods suffer from anisotropic resolution, which could distort 3D tissue architecture renderings and thus hinder new work to analyze and quantify 3D tissue microarchitecture. Therefore, there is a clear need for a non-destructive imaging tool that can accurately represent the 3D structures of the tissue or cellular architecture, with subcellular resolution in any arbitrary direction in the image volume, and with image contrast similar to traditional histopathology.
In this work, we use dual-view inverted selective plane illumination microscopy (diSPIM), one variant of LFSM, for its unique capabilities of imaging large samples with high resolution, low photo-damage [13,14] and isotropic resolution [15]. Optical sectioning ability is obtained by scanning a laser beam to create a light sheet into the sample, and the illuminated area is captured by another orthogonal de-coupled imaging path in a wide-field manner. Together, they provide non-destructive, fast 3D imaging with high contrast and resolution, providing a potentially useful tool for clinical applications. In most cases, LSFM is applied to small, translucent model organisms for study of embryogenesis [13,16,17], neuroscience [18], or optically cleared tissues, such as mouse brain [19,20] or small tissue blocks [6], entailing complicated sample preparation methods or imaging configurations.
Similar to the open-top light-sheet microscope (OTLS) [4,21] or light sheet theta microscopy (LSTM) [22], the design of diSPIM enables high-speed centimeter-size sample imaging with easy sample mounting. In OTLS, the objectives are placed below the sample holder in a lens immersion medium, and one of several customized imaging plates between the objective lens and sample is selected to provide a precise match of refractive index between the sample, the imaging plate, and the objective lens immersion medium [23]. In contrast, in the diSPIM architecture the multi-immersion objectives are immersed in the same bath of immersion medium along with the sample, with no additional interfaces introduced between the objective lenses and the sample. Currently, both OTLS and LSTM are designed for singleview imaging. In contrast, a specific feature of the diSPIM design [24,25] is that the illumination and detection paths can switch roles and capture orthogonal images of the sample by turn, contributing to more isotropic resolution in 3D by fusing and deconvolving the two views. Though multi-view deconvolution has been used before to obtain isotropic resolution, it was either applied to small and transparent samples (micrometer size) [15,23,[26][27][28]; large, but sparsely labeled structures [29]; required beads embedded as registration fiduciaries [20]; or required complicated imaging configuration and sample preparation for sample rotation and multi-view acquisition [30]. Here, with no need of rotating samples, we applied dual-view imaging and deconvolution for large (millimeter to centimeter size) 3D virtual histology of conventionally prepared densely-labeled samples to achieve more isotropic resolution, specifically DRAQ5 and eosin (D&E) stained cytology and tissue biopsy samples [31]. Nuclear signals from the DRAQ5 channel were used for dualview registration, with no need to embed extra beads into the samples. We first tested our dual-view imaging and deconvolution on a phantom of 500 nm fluorescent beads. Then, a cell sample stained with D&E was used to validate the dual-view dual-color imaging and data processing, for the benefits of its relatively smaller size and less scattering property than tissue specimens, and for its potential relevance in 3D cytopathology. Finally, we applied the methods to an optically-cleared prostate biopsy, to evaluate its potential for accurate 3D virtual histology.

Phantom preparation
Five (5) μL of 1:200 diluted 500 nm sub-diffraction YG (yellow-green) fluorescent beads (Fluoresbrite, Polysciences) were added to 1 g of 3% gelatin solution. Several drops of the mixed solution were placed onto a cover slide. After curing at 4°C for 5 minutes, the cover slide was mounted into a custom imaging chamber filled with water and imaged. 10X water objectives (Nikon, 0.3 NA) were used for this sample.

Cells and tissue acquisitions and preparation
Buccal cells were obtained by gently rubbing and rotating a cotton swab along the inside of the cheek of a human volunteer for 5 seconds. After removing from the cheek, the swab was immersed in phosphate-buffered saline (PBS) solution immediately and gently agitated. Several cheek swabs were made to collect more cells. The cell suspension was then centrifuged at 180 × g for 5 minutes. After discarding the supernatant, the cells were stained with a 1:100 w/v solution of eosin Y (Sigma-Aldrich) in 80% ethanol for 1 minute and centrifuged again at 180 × g for 5 minutes. The process was repeated with a 1:100 v/v dilution of 5 mM DRAQ5 (Biostatus, Ltd) in PBS for 1 minute before the cells were resuspended in a 3% gelatin block. Finally, the gelatin block was immersed in PBS and imaged with the 10X 0.3 NA water objectives.
A frozen human prostate biopsy was collected from the Louisiana Cancer Research Consortium Biospecimen Core Laboratory under a Tulane University Biomedical Institutional Review Board approved protocol. The frozen biopsy was thawed, fixed with formalin, and then thoroughly rinsed with PBS. Although diSPIM is tolerant to irregularity of the tissue surface, the biopsy was cut flat with a vibratome for more efficient correlation of optical sections and subsequent FFPE sections. After cutting, the biopsy was about 14.8 mm × 2.7 mm × 0.4 mm in size. It was then optically cleared using the X-CLARITY protocol (X-CLARITY, Logos Biosystems). Next, the cleared sample was stained with DRAQ5 overnight with gentle agitation followed by PBS rinsing for 3 times, and then stained with eosin for 30 minutes. The concentrations of DRAQ5 and eosin (D&E) used were the same as in the cell staining protocol. After removal from eosin solution, the stained sample was rinsed again with deionized water and PBS successively, and then immersed in two changes of X-CLARITY mounting solution for RI-matching. Finally, the sample was fixed in the imaging chamber and immersed in X-CLARITY mounting solution for dual-view imaging with the ASI/Special Optics Cleared Tissue Objective (CTO) objective lenses.
Traditional H&E-stained sections of the same biopsy were prepared after diSPIM imaging by the Tulane Histology Research Laboratory for further comparison and analysis.

diSPIM imaging
The diSPIM system, originally developed by Hari Shroff and ASI and commercially available from ASI [15], has been described in our previous paper [24]. Since then, several adjustments to our system have been made (Fig. 1). Firstly, the scanner tube lenses have been replaced with adjustable tube lenses to keep 4-f spacing for easily switching between water objectives (Nikon, 10X, NA: 0.3) and cleared tissue objectives (CTO, ASI/Special Optics, 15.3X-17.9 X, NA: 0.37-0.43 over RI range of 1.33 to 1.55, working distance: 12 mm) for dual-view multi-immersion imaging. For imaging the cleared prostate, we used the CTOs to minimize spherical aberrations. Secondly, we switched our SSD array from RAID 10 mode to RAID 0 mode for higher sequential writing speed. Finally, we changed our data acquisition card to PCI 6733 (National Instruments) to obtain more analog outputs for controlling of different illumination modes and laser lines. A 488 nm laser (Omicron, PhoxX) was used for imaging the eosin channel, and 647 nm laser (Omicron, LuxX + ) for the DRAQ5 channel. The beam waist was adjusted to ~12 μm when using 10X objectives and ~7 μm using CTOs at 488 nm wavelength. Volumetric images of one z layer were obtained by moving the sample with respect to the objectives using the XY motorized stage by raster scanning in the x-y direction, with multiple y strips acquired with about 20% overlap between two adjacent strips to ensure accuracy and uniformity of stitched images. For each y strip, the XY stage moved in the x direction with a constant speed, and the Tiger controller (ASI) was used to provide a trigger output once the stage arrived at a pre-determined position to synchronize the stage motion with the camera trigger. Dual-color imaging was achieved by automatically turning on and off the two lasers alternatively step-by-step along each y strip. The field-of-view (FOV) with the 10X objectives is about 1.33 mm × 1.33 mm, with the theoretical lateral resolution of 1.3 μm, and the FOV of the CTO varied from 0.75 mm × 0.75 mm to 0.87 mm × 0.87 mm, with the theoretical lateral resolution of 0.73 μm to 0.84 μm, over its RI range. For samples thicker than one 2 FOV (considering the 45° angle of each FOV), the objectives can be moved vertically relative to the sample to capture more depth information, with about 20% overlap between two adjacent z layers. As shown in the system schematic ( Fig. 1), the diSPIM necessitates a set of adjusted coordinates (prime coordinates) because of the 45° angle of the objectives compared to the traditional xyz coordinates. For SPIM-A (left illumination + right camera detection), x' A y'z A ' is used ( Fig. 1(B)). X' A is along the axis of the illuminating beam that travels through the left objective, y' is the direction that the beam scans, and z' A is the detection axis of view A. Similarly, x' B y'z B ' is used for SPIM-B (Right illumination + left camera detection. See the red labels in Fig. 1(B)). Fig. 1. diSPIM schematic. A virtual light sheet generated by one of the micro-mirror scanners that scans in the y(y') direction rapidly illuminates the sample through the illumination objective at an oblique 45° angle from one objective. The emission signal from the sample is then collected by the detection objective lens positioned perpendicular to the illumination objective and imaged with the dedicated camera for that side. The stage moves in the x direction to create a 3D stack. The objectives move vertically to achieve extended z depth (A). The two light paths can then switch roles for illumination and detection alternatively for dualview imaging. Compared to the traditional xyz coordinates, a new set of coordinates of x' A / B y'z' A / B for diSPIM is used. X' A / B axis is along the illumination direction of left/right light path, z' A / B is the corresponding detection axis of view A/B (B), and y' is the same as y (C). BS: beam splitter. MM: micro-mirror scanner. TL: adjustable tube lens. DM + F: dichroic mirror + emission filter. Ill obj: illumination objective. Det obj: detection objective.

Post image processing
After data sets corresponding to the two views for each color channel had been captured and stored digitally, the post image processing was run on a Dell Precision T7910 work station with Xeon CPU E5-2699 v3 processor and 128 GB RAM. Firstly, customized MATLAB code was used for shifting and transformation. Shifting is to restore the ± 45° angles of view A and view B, and transformation is to align the two views to the same orientation -either by resampling one of the views in 90°, or both views 45° clockwise / anticlockwise, for the subsequent dual-view registration step [32]. Secondly, we cropped a sub-volume with overlapped region from each single-view volume to do dual-view registration based on cell nuclei signals, and dual-view deconvolution using the Multiview Reconstruction plugin in Fiji [32,33]. The theoretical point spread function (PSF) was used in the deconvolution procedure. The theoretical resolution was utilized to generate a three-dimensional Gaussian kernel for use as the PSF, where full-width-at-half-maximum (FWHM) of major axis is equal to the axial resolution, and FWHMs of the other two directions are equal to the lateral resolution. The direction of major axis is parallel to the detection direction (orthogonal to the light sheet). The deconvolution iteration number was set to be 10. For D&E-stained samples, after shifting and transformation, we first remapped the DRAQ5 and eosin images into a composite RGB stack to simulate the traditional H&E images using the virtual transmission microscopy (VTM) method introduced by Giacomelli et al [31,34]. Next, we separated the RGB images into R, G, and B channels, and did the registration and deconvolution channel-by-channel based on the same inverted nuclei information in all the three channels. Then we fused the three deconvolved channels and inverted to form RGB stacks again. Stitching of different y strips was done with the Grid/Collection stitching plugin in Fiji [35]. Finally, the processed data was loaded into Amira (Thermo Fisher Scientific) for 3D visualization. Note that the eosin and DRAQ5 channels were not registered separately before being combined and remapped into the RGB colorspace. This is because these two dyes stain tissue very differently -DRAQ5 stains nuclei sparsely which can be readily used as fiduciaries for dualview registration, but eosin staining is dense and less featured, which could make registration difficult. Therefore, utilizing the same nuclear signals in the remapped R, G and B channels for registration could avoid possible mismatches between separately registered DRAQ5 and eosin channels.
To compare dual-view deconvolution with single-view deconvolution, we also loaded the D&E images from SPIM-A into Amira and used the theoretical PSF to do non-blind deconvolution. Figure 2 shows a comparison of maximum intensity projections in the x'y', x'z' and y'z' planes between one of the single-view data sets and the dual-view deconvolution results of 500 nm fluorescent beads phantom using the 10X 0.3NA water immersion objectives. Most of the beads are preserved after deconvolution, with more isotropic resolution in 3D. Apparent improvement in resolution by dual-view deconvolution can be clearly seen in the images of the second row and the Gaussian fitted intensity plots, especially along the z' axis. Based on measurements from 10 beads in each field, the full width at half maximum (FWHM) for single-view x', y' and z' are 2.72 ± 0.17 μm, 1.48 ± 0.19 μm, and 4.73 ± 0.26 μm, and for dual-view deconvolved are 1.93 ± 0.14 μm, 1.12 ± 0.21 μm, and 1.54 ± 0.08 μm. Thus, 1.4fold, 1.3-fold, and 3.1-fold reductions in the PSF size have been achieved in the x', y' and z' directions, respectively, by dual-view imaging and deconvolution.  Figure 3(A) shows a deconvolved volume of a buccal cell sample rendered in H&E pseudocolor. Buccal cells characterized by nuclei surrounded by pouches of cytoplasm with various shapes can be clearly seen. In Fig. 3(B-J), comparisons of minimum intensity projections in three orientations between single-view from SPIM-A and dual-view deconvolution results are shown. Note that the resolution in y'z' and x'z' has been improved significantly by dual-view fusion and deconvolution (D-G, I-J) eliminating distortions in cell and nucleus shape seen in the single-view. The green arrows and blue circles in A, B and C indicate the same region of the sample. Note that in 2D slices, some cells overlap with each other and are difficult to differentiate, as indicated by the green arrows in Fig. 3(B) and (C). But in 3D, rotating the cells to an appropriate orientation can easily separate them (indicated by the green arrow in Fig. 3(A), which shows the same cell in B and C).    quantitatively, same volume , Decon, and H zed through on ). Statistical an cted cell numb antly more cell mber from Dec registration be he H&E image e groups were antly higher (p on to enhance YZ planes from were chosen to co ensity projection econ views. Inver PIM-B and Decon a preliminary (1,560 μm × 3 H&E slices of ne-way analysi nalysis reveals bers of SPIMls (p < 0.0001) con is still sign etween the D& es compared w e also calculat (p < 0.0001) t image sharpne SPIM-A, SPIMompare high resol of 11 sequential i rted intensity prof n, indicated by lin y experiment w 390 μm × 4 μm f the same sam is of variance ( that there is n -A and SPIM-) than the prev nificantly diffe &E and H&E v with SPIM-A ted similarly. than either of ess, resulting in

Discussion
In this study, we have demonstrated that dual-view imaging and deconvolution can improve resolution three dimensionally, with 1.4-fold, 1.3-fold, and 3.1-fold improvements in resolution achieved in x', y' and z' (the detection axis for one of the single-views) directions respectively, compared with single-view using sub-diffraction beads and 10X water objectives. This method improves the imaging resolution of diSPIM using these objectives to less than 2 μm in all three directions, which is sufficient to resolve human cell nuclei in any arbitrary orientation. Typically, multi-view registration and deconvolution is applied to micrometer-size, translucent samples, sparsely labeled structures, or requires beads embedded with sample as fiduciaries, or sample rotation during imaging. The calculation is challenging and complicated when encountering large, opaque samples with dense labeling and different optical paths inside the scattering sample. Sample rotation and bead embedding further complicate sample preparation, and embedded beads may be incompatible with optical clearing. Here, we demonstrate that based on fluorescently-labeled nuclear signals present in densely-labeled tissues for dual-view registration, dual-view deconvolution can successfully be applied to cytological samples and millimeter-to centimeter-sized optically-cleared biopsy specimens, verifying the potential for accurate 3D cytology and histology. Furthermore, we demonstrate a method for dual-view registration and deconvolution of images after remapping into an inverted RGB colorspace to simulate the appearance of brightfield H&E images. It should be mentioned that isotropic resolution can also be achieved by axially swept light sheet microscopy (ASLM) without the need of deconvolution [37,38]. ASLM leverages axial sweeping of a scanned light sheet, with only signals excited by the diffraction-limited beam waist area being collected via an electronic confocal slit detection scheme, and with the axial resolution being mainly dependent on how thin the illumination beam waist is. The method may avoid some of the limitations of the dual view approach (i.e. lower data collection, processing, and storage requirements), but it may increase imaging time compared to traditional LSFM, and care must be taken to ensure that beam distortions within the sample due to scattering or aberration does not broaden the beam waist above the diffraction limit thereby decreasing axial resolution. There are several situations in which 3D imaging with isotropic resolution could be useful in virtual histology or cytology imaging. Firstly, as demonstrated here, it could enable more accurate quantitative measurements of nuclear and cytoplasmic volume to calculate the nuclear-to-cytoplasmic ratio, which is useful in cancer diagnosis, and would provide a more accurate representation of true cellular shape, compared to traditional 2D cytology smears, or 3D microscopy images with low axial resolution. In addition, it could provide an enhanced ability to separate otherwise overlapped features in 2D optical slices (for instance, cell nuclei comprising glands) compared to other 3D imaging methods with low axial resolution. This could especially be helpful when studying and quantifying three-dimensional patterns of tumor growth. Accurate three-dimensional representation of tumor microarchitecture could also benefit cancer diagnosis. It has been demonstrated that 3D cytology could reduce the false negative rate for adenocarcinoma detection [39], and viewing 3D histology slices in a single orientation could improve intra-observer agreement [4]. Though further evidence in a larger data set is required to determine whether more isotropic resolution could further improve diagnostic accuracy over other single-view 3D imaging methods, the method presented here could be helpful by providing pathologists a tool for evaluating tissue histology in any desired orientation without fear of distortions introduced by the imaging method itself. Lastly, with better 3D resolution and higher imaging sharpness, the method presented here could build a solid foundation for classification and quantification of cells or tumor architecture, for instance expanding our ongoing work in topological analysis of tumor microarchitecture for quantitative tumor grading [40] from 2D to 3D. Other information, such as nuclear shape, gland size, gland shape, glandular connectedness, distance between adjacent glands, etc. can also be more accurately analyzed in 3D with more spatially isotropic resolution, for instance to supplement prostate cancer histology as one pertinent example of interest.

Conclusion
The gold standard of histopathology for tissue or cytology specimens is destructive, time consuming, and limited to 2D slices or smears. 3D imaging of cytology and histology samples is important for more comprehensive understanding and diagnosis of diseases, by observing and differentiating signals from different observation angles, and providing accurate 3D renderings of tumors and their microenvironment. However, current 3D imaging tools do not fulfill all the requirements of fast, high-resolution, and isotropic imaging for volumetric cytological samples or large, scattering tissue biopsies. With diSPIM imaging and dual-view deconvolution of virtual H&E images, more accurate determination of 3D architecture on those samples has been obtained. Enhanced, more isotropic 3D resolution, higher levels of image detail and higher image sharpness have been proven qualitatively and statistically in the dual-view deconvolved results compared with single-view imaging. The method described here should be further investigated in larger sample sets to determine the impact of improved and isotropic resolution on clinical diagnostic endpoints. Processing on centimetersize samples is achievable by separating the data set into several small blocks and stitching together after deconvolution, although at the expense of longer processing times. More timeefficient image processing methods deserve further investigation and effort.

Funding
National Institutes of Health; National Cancer Institute; Innovative Molecular Analysis Technologies (IMAT) (1R33CA196457).