Non-destructive 3D pathology with light-sheet fluorescence microscopy for translational research and clinical assays

Abstract

In recent years, there has been a revived appreciation for the importance of spatial context and morphological phenotypes for both understanding disease progression and guiding treatment decisions. Compared with conventional 2D histopathology, which is the current “gold-standard” of medical diagnostics, non-destructive 3D pathology offers researchers and clinicians the ability to visualize orders of magnitude more tissue within their natural volumetric context. This has been enabled by rapid advances in tissue-preparation methods, high-throughput 3D microscopy instrumentation, and computational tools for processing these massive feature-rich datasets. Here, we provide a brief overview of many of these technical advances along with remaining challenges to be overcome. We also speculate on the future of 3D pathology as applied in translational investigations, preclinical drug development, and clinical decision-support assays.

Motivation for 3D pathology

For over a century, the gold standard for diagnostic medicine has been based on histopathology, which is the examination of thin tissue sections mounted and stained on glass slides and visualized with an analog brightfield microscope. Conventional histology involves preserving tissue specimens in formalin-based fixatives that degrade nucleic acids, followed by dehydration in chemicals such as xylene and ethanol such that the tissues may be embedded in paraffin wax and thinly sectioned (4 – 5 microns thick) onto glass slides. Once sectioned, the tissues are most-often stained with chromophores such as hematoxylin and eosin (H&E), and occasionally with targeted probes such as antibodies or oligonucleotides (for in situ hybridization). Some negative aspects of this process, as illustrated in Fig. 1, include: (1) destructive sectioning of tissues, which can impede downstream molecular assays that require ample tissue material, (2) severe sampling limitations, in which 1% or less of a tissue specimen is typically sectioned and visualized by pathologists, and (3) the lack of volumetric information, which can result in investigational and diagnostic ambiguities. The digitization of histology slides, known as whole slide imaging (WSI), has recently been approved by the US Food & Drug Administration (FDA). While this facilitates computational analysis (i.e., artificial intelligence and machine learning), it requires an additional step of scanning slides that does not mitigate any of the challenges of slide-based histology listed above.

Figure 1. Key advantages of 3D vs. 2D pathology.

(1) Destructive sectioning of tissues in conventional 2D histology methods can impede downstream molecular assays that require ample tissue material, unlike non-destructive 3D pathology methods. (2) Due to the destructiveness and time-consuming nature of physically sectioning tissues onto glass slides, there are severe sampling limitations for conventional 2D histology. With non-destructive 3D pathology, large specimens can often be imaged in toto, which enables improved diagnostic sensitivity, the ability to identity “rare events” that are often missed with 2D sections, and most importantly: (3) the ability to accurately characterize complex 3D structures and cell distributions, which can improve investigational and diagnostic certainty.

Recent technical advances now offer the potential to improve upon the long-standing “gold standard” of histopathology by imaging thick tissue specimens non-destructively such that 100% of the specimen is maintained for downstream assays. With this technological paradigm, volumetric information about the tissue microarchitecture and molecular constituents can be obtained throughout the entire specimen, which is orders-of-magnitude more tissue than is currently sampled/visualized via conventional 2D histology sections. The main innovations that have enabled this new field of “non-destructive 3D pathology,” which will be surveyed within this article, include thick-tissue clearing and labeling strategies to generate fluorescent samples that are highly transparent to light, along with high-throughput user-friendly 3D microscopy systems in conjunction with advanced data-science methods and instrumentation for handling the massive datasets that are generated.

Three-dimensional (3D) microscopy methods have been available to life scientists for several decades, most ubiquitously in the form of confocal and multiphoton microscopes(1–5), and occasionally through reconstructions of serially sectioned tissue blocks(6–8). Major advances in biological understanding have been gained by such workhorse microscopy techniques in research settings. However, such systems are limited in speed because they are fundamentally laser-scanned imaging techniques in which images are generated in a point-by-point fashion over time. While various approaches have been developed to improve imaging speed, such as the use of spinning disks for confocal microscopes(9, 10), and temporal focusing or multi-focal methods for multiphoton microscopy(11–13), imaging speeds have still precluded the mainstream application of such techniques for high-throughput preclinical or clinical applications.

One area in which conventional optical-sectioning microscopy approaches (e.g., confocal and multiphoton microscopy) offer great advantages is for imaging highly scattering uncleared tissues, as is necessary for in vivo / intravital imaging applications. This is because of the background-rejection and background-suppression abilities of confocal and multiphoton microscopy, respectively, which are sacrificed with faster camera-based 3D microscopy techniques such as light-sheet and light-field microscopy. Sophisticated miniaturized forms of confocal and multiphoton microscopes have even been developed to enable endoscopic microscopy within human patients – for surgical guidance or early detection of malignancies – or for long-term neuroimaging in the brains of awake freely-moving rodents, amongst many other applications(14–19). In this review, we will focus on the imaging of ex vivo specimens, in which a myriad of clearing/labeling approaches are readily available, and where device miniaturization is not of the highest priority.

We will attempt to provide a high-level survey of the following topics pertaining to non-destructive 3D pathology of excised tissues for translational research and clinical diagnostics: (1) We will first provide a brief history of light-sheet fluorescence microscopy, which has in recent years become recognized as a powerful technique for high-resolution evaluation of large preclinical and clinical specimens that are optically cleared. (2) We will provide a focused discussion on recent hardware advances and challenges for cleared-tissue light-sheet fluorescence microscopy (LSFM) systems. (3) We will provide a summary of recent advances in thick-tissue fluorescence labeling and optical clearing. (4) We will discuss advances and challenges in raw data processing and image formation. Note that we will not cover downstream image-interpretation and analysis methods, such as deep-learning-based denoising, super-resolution, and deconvolution methods, as well as machine-learning methods for image translation (style transfer), image segmentation, and classification. While these topics are highly important and rapidly evolving, they deserve (and have been the subject of) dedicated review articles(20–27). (5) Finally, we will provide a forward-looking perspective on future directions for the field of 3D pathology in translational research, preclinical drug development, and clinical-assay development.

Brief introduction to light sheet fluorescence microscopy

In its basic form, light-sheet microscopy uses spatially localized illumination of a specimen with a thin sheet of light that is designed to coincide with the focal plane of a camera-detection path. Selective illumination in this way is powerful because it enables recording of volumetric images of biological specimens with low background signal, high speed, and minimal photobleaching or phototoxicity(28, 29). In contrast, confocal and wide-field microscopes illuminate the entire thickness of a specimen and as a result face drawbacks in speed, background signal, or light dosage.

While light sheet microscopy was first used in the early 1900s to study colloids(30, 31), it experienced a boom in popularity starting about a century later when redeveloped for the imaging of fluorescent biological specimens as light sheet fluorescence microscopy (LSFM)(32, 33). In the past ~20 years, LSFM has been used to study important biological problems including embryonic development(34), cardiac hemodynamics(35), and neural dynamics(36). Many of these studies have been facilitated by highly innovative LSFM instruments, with key advances centered on the choice and arrangement of objective lenses relative to the specimen. These LSFM design innovations can yield significant advantages for specific applications in terms of speed, resolution, range, throughput, cost, and simplicity.

Recent hardware advances and challenges for cleared-tissue LSFM systems

In recent years, there has been considerable effort into improving the performance and utility of LSFM systems for cleared-tissue imaging. As with most technologies, pursuit of improved performance for one system specification often results in unavoidable trade-offs. However, state-of-the-art systems strive for an optimal balance in performance, guided by the recent needs of researchers and cleared-tissue imaging experiments (37).

Many of the key design choices for LSFM center on the arrangement of objective lenses, i.e., LSFM systems are often built around the specimen. The original cleared-tissue LSFM design uses two perpendicular objective lenses arranged around a specimen, with a detection objective above or below the specimen, an excitation objective positioned at the side, and the specimen held in place on a substrate, within a cuvette, or within a rotating capillary (Fig. 2a). While this geometry is relatively simple to implement, it places physical constraints on the lateral size of the specimen such that it is not possible to image large, centimeter-scale tissue slabs or multiple specimens mounted in standard holders (for example, well plates). Alternatively, the specimen may be held on a substrate with the perpendicular objective lenses each oriented at an angle relative to the substrate. Positioning both of the objective lenses on the sample side (Fig. 2b) minimizes optical aberrations but can constrain sample geometries (38–44). Positioning both of the objective lenses below the sample holder, called open top light sheet (OTLS) microscopy (Fig. 2c), requires some care with sample preparation to avoid optical aberrations but enables laterally unconstrained imaging that is particularly well suited to high-throughput imaging of specimens in a range of formats (40–46). A third group of LSFM designs use a ‘single-objective’ to both deliver the light sheet and collect the emitted fluorescence. In these systems, the illumination and collection beams share a single objective and are non-orthogonal to one another, requiring a method such as remote refocusing to rectify the imaging plane (Fig. 2d) (47–58). Orienting the objective in the vertical direction with respect to a horizontal specimen holder makes use of the objective’s full working distance and dramatically increases the system’s tolerance to refractive-index mismatch (relative to inverted and open-top LSFM systems). Most recently, the light-sheet theta microscope (LSTM) and hybrid OTLS system have been developed (Fig. 2e) (46, 59). In LSTM (Fig. 2e) and hybrid OTLS systems (Fig. 2f), a high-NA collection objective is oriented in the vertical direction (similar to single-objective systems), and a separate illumination objective(s) is used to deliver the light sheet. These designs possess all the advantages of a single-objective LSFM system, while the use of a separate illumination objective provides more degrees of freedom to optimize the system resolution for a desired imaging application.

Figure 2. Light-sheet fluorescence microscopy (LSFM) and data handling. (a-g) Architectures of traditional, inverted, open-top, single-objective, light-sheet theta, and hybrid light-sheet microscopy systems.

Relevant characteristics of each microscope architecture, including lateral constraints on specimen size and usable imaging depth are highlighted. (h) Typical data acquisitions associated with any light-sheet microscopy system are listed. This includes raw data generation from the microscope itself at 1+ GB/sec, storage of the data on high-speed drives, transfer of the datasets to network storage, and optional storage of these often multi-TB datasets in the cloud.

As with any microscopy method, higher spatial resolution is commonly achieved by using objectives with a higher numerical aperture (NA). Higher NA objectives typically have exceedingly short working distances, which for conventional microscopy methods (i.e., conventional widefield and confocal microscopes) limits the specimen size in only one dimension. However, the dual-objective architecture of LSFM systems results in more severe geometric constraints, where the specimen size is now limited in two dimensions. In addition, for the highest NA objectives, it may not be possible to position two objectives relative to one another for LSFM imaging due to the physical dimensions of the objectives themselves. Additional spatial-resolution improvements under exploration include incorporation of structured illumination, stimulated emission depletion, and deep learning-based techniques, although all these methods are yet to become mainstream in cleared-tissue LSFM systems (60–63).

While the previously mentioned approaches are primarily concerned with lateral resolution, there has been equal if not greater interest in improving the axial resolution of a LSFM system. Unlike other more conventional microscopy methods, the use of a dedicated excitation optical path provides more degrees of freedom which can be optimized to provide improved resolution in the axial dimension (i.e., engineering the light sheet). The majority of LSFM systems use Gaussian beam illumination, which introduces a fundamental trade-off. The Gaussian light sheet is either tightly focused (high axial resolution) yet short and unable to cover the field of view of the system, or the Gaussian light sheet is weakly focused (low axial resolution) but long, providing full coverage of the system’s field of view. To overcome this trade off, strategies have been developed for tiling or axially sweeping a tightly focused light sheet across the field of view (64, 65). Tiling methods discretely step the light sheet across the field of view, capture a separate image for each position, and computationally fusing the overlapping portions of each image together, yielding a single image with high axial resolution across the field of view. One drawback to this approach is decreased imaging speed (multiple image exposures are required for each position with a specimen). Axial sweeping methods overcome this drawback. Rather than discretely stepping the light sheet across the field of view, the light sheet is continuously moved, in sync with the rolling shutter of the camera, which acts as a moving slit, rejecting thick areas of the light sheet and exposing pixels to only the tightly focused portion of the light sheet. Although axial sweeping requires only one image exposure, the optomechanical elements currently used for axial sweeping can still limit imaging speeds. Additional approaches include using non-Gaussian light-sheets. Examples include Bessel, Airy, and optical lattices (66–69). While these illumination strategies may improve axial resolution, they can result in reduced image contrast, or require computationally expensive deconvolution methods to recover improved performance (70–72).

Although the highest spatial resolution is always desirable, it is worth noting that for any volumetric imaging modality, dataset sizes and imaging times typically scale with the cube of the spatial resolution (73). Therefore, large gains in efficiency can be achieved by employing multi-scale imaging strategies. For example, an efficient multi-scale imaging workflow can be achieved by first using a low magnification system to screen specimens and identify interesting regions of interest, followed by detailed inspection of small regions of interest using a higher-magnification system. Similar to other microscopy methods, the simplest multi-scale implementation is the use of a turret of multiple objectives with varying magnifications. This approach has been demonstrated with several of the previously mentioned LSFM architectures, including the original cleared-tissue design, as well as a recently published multi-resolution OTLS system (44). Alternatively, multiple LSFM systems with different magnification set points can be used. In a third approach, a single system can be designed with multiple imaging paths, such as the recently published hybrid OTLS system (46). Finally, there are approaches which can leverage a single lens to simultaneously provide high resolution and a large field of view (74). These lenses, with a so-called high space-bandwidth product, may be combined with high megapixel-count cameras to provide multi-scale imaging. Regardless of the imaging application or multi-scale imaging approach, the spatial resolution of a LSFM system is often chosen to be ‘just enough’ for the given scientific question and imaging experiment.

Recent advances and challenges in data and image processing

LSFM datasets are notoriously challenging to work with. The difficulties begin at the time of acquisition. Unlike many other microscopy methods, LSFM systems generate data at the maximum data rate of the camera (~1 GB/sec for state-of-the-art sCMOS cameras). For LSFM systems using multiple cameras, the aggregate acquisition rate can quickly reach multiple GB/sec (75, 76). To robustly capture this flood of information, specialized considerations must be made to the downstream hardware and software.

Datasets must be streamed to high-speed solid-state drives (SSDs) or traditional hard disk drives (HDD) that can be aggregated together as a redundant array of independent disks (RAID) to achieve high write speeds. However, the size of these arrays is often limited to 10’s of TB at most, which is often too little space to store LSFM datasets. Therefore, datasets must often be transferred from a local acquisition workstation to some form of larger networked storage, where the local SSDs act as a temporary cache. To prevent this local cache from overfilling, the network transfer speeds must exceed the LSFM system’s acquisition speed. High-speed networking at speeds of 10, 40, or even 100 Gb/sec may be necessary. This requires significant testing and investment from an institution to provide the proper IT infrastructure to support these high-speed data transfers (77). It is important to note that bandwidth requirements can be reduced by performing online compression. However, the compression algorithm’s speed must ideally outpace the LSFM system’s acquisition speed (78, 79).

Once an entire dataset has been successfully captured, the data must then undergo a number of processing steps. For cleared-tissue experiments, the first step is to register and align the many adjacent imaging tiles resulting from the serial acquisition of those volumetric tiles across a given specimen. Several commercial software packages (such as Imaris Stitcher) as well as popular open-source tools (such as TeraStitcher, BigStitcher, BigStream, and Stitching Spark) have been developed to address this challenge (80, 81). These alignment methods can be performed with varying levels of complexity, ranging from pure translation of tiles, to full affine transformations, to complete non-rigid deformation and alignment. Once the alignment of all tiles has been determined, the datasets are often ‘fused’, yielding a single contiguous volumetric dataset with blended seams between adjacent tiles. These operations are increasingly computationally expensive and should ideally be tailored to the requirements of a specific imaging dataset. These datasets may then be visualized and/or subjected to downstream post-processing routines. Regardless of the step, care must be taken to parallelize the task across ample computing resources, and to optimize the runtime speed to complete the computational task within a reasonable timeframe.

Recent advances and challenges in tissue labeling and clearing

In parallel with many of the hardware-based innovations described above, recent years have witnessed major breakthroughs in technologies for sample preparation and imaging workflows that enable the imaging of thicker intact specimens, with higher resolution and lower scattering/aberration, in rich molecular detail. We review several of these developments here, with an emphasis on areas that are particularly relevant for pathology samples.

Traditional workflows for optical microscopy of thick biological specimens have required physical sectioning of the specimen for several reasons. First, nearly all microscopy-based pathology workflows utilize 2D wide-field microscopy techniques in which thin sections (≤10 μm) are required to produce clean, blur-free snapshots. Second, even work that utilizes traditional 3D microscopes (e.g., confocal microscopy) rarely image beyond ~100 μm depth due to the scattering of light by the specimen. Third, the labeling of thick specimens by large exogenous probes (e.g., antibodies) can require weeks or longer for thicker specimens (>100 μm) due to the slow pace of passive diffusion of probes within a sample. While many powerful studies are performed within these constraints, they lack information that can most easily be gained by studying larger intact specimens such as the 3D structure of glandular networks (82). The past ~20 years has seen a flurry of developments that make imaging of thick specimens easier than ever and have helped set the stage for rapid progress in 3D pathology.

Tissue clearing and expansion

A plethora of “tissue clearing” techniques now exist for combating light scattering in tissues and enabling the imaging of specimens 1–10 mm in thickness (83–88). A first subset of techniques are hydrophobic, or “solvent-based” clearing methods that use dehydration and immersion of specimens in various organic solvents in order to reduce the variations in index of refraction that lead to scattering of light. A second subset of techniques uses infusion of aqueous or hydrophilic solutions containing solutes including various sugars or alcohols. Example solvent-based clearing techniques include the DISCO family of methods and ECi, while example aqueous-based clearing techniques include SeeDB, FRUIT, and CUBIC (84, 86, 89–92). A third and somewhat distinct group of clearing techniques, such as CLARITY, synthesizes within the specimen an acrylamide-based hydrogel polymer that serves as a scaffold to hold proteins and some other structures in place while allowing lipids to be removed and enabling infusion of an index-homogenizing solution (87). All these clearing procedures differ substantially in key properties including complexity, compatibility with different types of fluorescent labels, and the final index of refraction. For instance, Tanaka et al. (93) used iDISCO for tissue clearing of various formalin-fixed paraffin-embedded (FFPE) human tumors to study epithelial-to-mesenchyme transition and angiogenesis while Barner et al. 2022 used CUBIC-HV for tissue clearing of fresh axillary lymph nodes to assess the ability of 3D pathology to improve breast cancer staging (94). Impressively, even human organs that are multiple centimeters thick, such as whole kidneys, can be cleared and imaged by light-sheet microscopy, as shown by Zhao et al (95) using the rapid clearing method, SHANEL (Fig. 3A–C). SHANEL is based on the use of the detergent CHAPS, which forms small micelles that diffuse more rapidly into tissues than the conventional permeabilization detergents sodium dodecyl sulfate or Triton X-100, in conjunction with N-methyldiethanolamine for decolorization of colorful heme molecules from residual blood clots.

Figure 3. Clearing, expansion, and labeling of tissues.

A-D) Whole human kidney cleared via SHANEL, labeled with a dextran (magenta) and nuclear stain (green), and imaged by light sheet microscopy to reveal blood vessels and glomeruli. E-F) Human kidney sections imaged by confocal microscopy after hydrogel expansion and revealing differences between E) a healthy patient and F) a patient with minimal change disease. The specimens were antibody stained for vimentin (blue), actinin-4 (green), and collagen IV (red), and counterstained for nuclei (white). G-I) Cleared mouse kidney tissue that was labeled via FLARE (amines (red), carbohydrates (green), and nuclei (blue)) and imaged by open-top light-sheet microscopy. Panels A-D adapted from reference (95) (CC BY 4.0). Panels E-F adapted from reference (101) [**Note to Annual Reviews: Permission to be obtained prior to publication]. Panels G-I used from reference(113). [**Note to Annual Reviews: The panels were created by the authors and the publisher grants authors the right to reuse their own figures without permission.]

Tissue-hydrogel hybrids have been extensively used in recent years not just for clearing of specimens, but also for physical expansion to enable super-resolution expansion microscopy (ExM) at a resolution of 70 nm or better (96–98). This is achieved in most cases by synthesizing within the sample a polymer composed of sodium acrylate and acrylamide co-monomers that links to the sample (87, 99, 100). The resulting tissue-hydrogel hybrid, with its abundant carboxylate groups, has a high osmotic pressure and, when immersed in deionized water, will enlarge to several times its original size. The procedure can be performed with very low distortion and isotropic 3D expansion so that features too close to resolve in the original specimen can be resolved in the expanded state even with lower-resolution microscopes.

Most ExM work to date has been applied to model organisms, but there are notable examples in which it has been used to study human pathology specimens (101, 102). For example, Zhao et al. (101) used ExM to study human kidney biopsy specimens (Fig. 3 E–F), where they were able to accurately distinguish minimal change disease (MCD) from focal segmental glomerulosclerosis (FSGS); in current clinical practice, electron microscopy is required to distinguish MCD and FSGS by examining the nanoscale details of podocyte foot process effacement. ExM already has the potential to become a valuable tool for clinical research and pathology, and as the technology continues to advance at a rapid pace, it is likely to gain even more powerful capabilities in the years to come.

Tissue labeling

Broadly speaking, fluorescence microscopy uses a rich and ever-growing palette of labels that can provide valuable information about the specimen. Some of the most popular labels used in pathology contexts include antibodies for labeling of specific target proteins, lectins that bind specific carbohydrates, oligonucleotide probes for fluorescence in situ hybridization (FISH) detection of DNA or RNA, and small molecule labels that relatively nonspecifically bind lipids, DNA, carbohydrates, or proteins.

Although current clinical practice evaluates only thin sections of biopsy tissue (typically ≤10 μm), which requires little time to stain, the staining of thick specimens with relatively large antibodies (150 kDa, or ~15 nm across) can be prohibitively slow (e.g., weeks or longer to stain specimens >1 mm), and there has been considerable innovation in developing strategies to efficiently deliver antibodies to thick tissues for 3D imaging. Chung and coworkers, in particular, have developed several methods to speed up antibody (or other probe) labeling reactions in tissue-gel specimens, as reviewed in (87). SWITCH (103) used a two-step approach in which probes are first delivered in a buffer containing a detergent that prevents probe binding. The probes are then “activated” in a second buffer without detergent that enables binding. This approach enables decoupling of the rates of probe diffusion and binding and leads to more uniform staining of thick specimens. In a creative use of tissue-gel hybrids called ELAST, Ku et al. (104) synthesized a highly elastic hydrogel within tissue specimens so that antibodies and other probes, whose penetration times scale quadratically with sample thickness, could diffuse much more rapidly through the specimen when physically “thinned” via stretching. However, the increased labeling speed is offset in part by the relatively long time of ~20 days required to create the tissue-gel specimen in the first place. Finally, stochastic electrotransport uses electric fields to enhance diffusion of antibodies and other changed species through tissue-gel specimens (105).

In a different approach for thick-tissue labeling and clearing, Susaki et al. (106) developed procedures called CUBIC-HV for 1-step efficient and uniform staining of whole organs from model organisms by carefully tuning concentrations of detergent, antibody/probe, and additives such as quadrol, urea, and salt, while also controlling temperature. Indeed, the use of passive diffusion for chemical labeling can also be effective across whole human organs when implemented with procedures tailored for speed, such as SHANEL (Fig. 3D) for labeling of whole human organs (95).

Small-molecule labels are powerful tools for staining pathology specimens and are also highly effective in 3D light sheet microscopy. For some of our work, we have used noncovalent staining with common DNA-affinity dyes (e.g., DRAQ5, TO-PRO3) to label nuclei, in conjunction with the general physiology stain, eosin, which has an affinity for basic structures including proteins (42, 45, 46). These combinations of stains enable a rapid and simple fluorescent analog of the traditional H&E (hematoxylin and eosin) pathology stain when implementing some image processing to convert the fluorescence images (acquired with grayscale cameras) to look like chromogenic (absorption-based) H&E histology (Fig. 4B) (107–110). This moderately specific H&E-analog stain has even been used to train a deep learning-based algorithm to create “computational immunostains” based on training sets with specimens co-stained with the H&E analog and a highly specific antibody (Fig. 4C) (82). This, in turn, enables “image-translation assisted segmentation in 3D (ITAS3D)” for the 3D segmentation of diagnostically important tissue structures (e.g. prostate glands) without requiring tedious manual annotations or slow/expensive antibody staining(82).

Figure 4. Translational applications in pathology.

A) Human lymph node studied by IBEX at 64 channels and imaged by confocal microscopy (scale bars = 500 μm and 50 μm). B) Cleared human prostate biopsies that have been stained with a fluorescent H&E analog and imaged by open-top light-sheet (OTLS) microscopy (scale bars 1 mm, 25 μm, 10 μm). C) A 3D pathology dataset of a prostate biopsy stained with a fluorescent analogue of H&E (left). Deep learning-based image translation was used to convert the H&E dataset into a synthetic dataset that looks like it has been immunolabeled to highlight a cytokeratin biomarker (brown) that is expressed by the epithelial cells in all prostate glands. In turn, this synthetically immunolabeled dataset allows for accurate 3D segmentation of the prostate gland epithelium (yellow) and lumen spaces (red). Quantitative features derived from these segmented 3D structures are used to train a machine classifier to stratify between aggressive (recurrent) versus indolent (non-recurrent) cancer (82) D) Breast cancer tumor tissue section stained with H&E. Cancerous regions are marked in red, cancer-like region marked in black, and stromal tissue marked in yellow. E) Spatial transcriptomics data corresponding to D) with transcriptomic-based clustering indicated by color of dot. F) Heat map of highly differentially expressed genes for clusters in E). Panel A used from reference(117) with permission. Panel B from reference (45) (CC BY 4.0). Panels D-F adapted from reference (125) (CC BY 4.0).

We also recently developed a highly versatile, covalent labeling method called FLARE (Fluorescent Labeling of Abundant Reactive Entities) that can covalently label amines and carbohydrates together with noncovalent labeling of DNA (94, 111–113). The amine and DNA stains again produce a fluorescent analog of H&E, while the concurrent carbohydrate stain provides a fluorescent analog of the traditional PAS (periodic acid Schiff) histology stain widely used in pathology (107). The feature-rich tri-color stain (Fig. 3G–I) has several useful attributes. First, the covalent stains, which are not removable by washing, are compatible with a wide range of protocols for clearing and expansion. Second, FLARE is compatible with other labeling procedures such as antibody labeling of specific proteins or fluorescence in situ hybridization (FISH) labeling of nucleic acids, and the specific dyes used for FLARE can be selected from a wide range of commercial dyes available across the entire visible spectrum. These small molecule stains are generally bright and rapid owing to their small size and ability to label abundant general chemical targets on a sample.

In situ proteomics and genomics.

Standard fluorescence microscopes routinely image 3–5 channels, but there has been substantial interest in recent years in developing techniques that can image a larger number of channels in situ in order to concurrently study many proteins, mRNA, DNA loci, or other molecules of interest. Many of these emerging techniques have primarily been applied to cells or 2D tissue sections, but the tools, broadly speaking, are highly promising for their potential applications in 3D pathology.

One approach to boost the number of channels for protein detection in optical imaging is to use multiple rounds of imaging, with antibody removal or fluorophore bleaching in between, such as with methods termed MxIF, t-CyCIF, CODEX, or IBEX (114–117). These techniques have been demonstrated for the imaging of tens of channels (Fig. 4A), typically on thin specimens, although, for instance, Murray et al. demonstrated the use of SWITCH to perform up to ~20 rounds of sequential labeling and imaging in hydrogel-stabilized mouse brain tissues (103).

Rather than study protein abundance through highly multiplexed antibody labeling, some approaches have sought to study mRNA since mRNA abundance can be a good proxy for protein expression (118) and since mRNA can be readily targeted by fluorescence in situ hybridization (FISH) for nearly any gene. For example, MERFISH and seqFISH use encoded multi-round FISH techniques to detect hundreds to thousands of mRNA targets (119, 120), FISSEQ and STARmap utilize in situ sequencing for mRNA detection with workflows based on next-generation sequencing pipelines (121, 122), and spatial transcriptomics and DBiTseq use spatial barcoding by a substrate or microfluidics device, respectively, to tag mRNA from specific regions of a specimen with barcodes that can be decoded during sequencing (123, 124). These and other methods for spatially resolved detection of mRNA are rapidly evolving and hold great promise for the study of pathology samples at single-cell resolution, with high mRNA detection efficiency, and with genome-wide coverage. Several commercial efforts have gained traction in this area and have been used in studies of human pathology specimens, albeit typically in two dimensions. For example, a commercial implementation of spatial transcriptomics from Visium has been used to study breast cancer tissues (Fig. 4D–F), revealing details of the tumor microenvironment such as metabolic reprogramming and obtaining new findings that may be helpful for diagnosis or treatment (125). In addition, the GeoMx platform from NanoString, which uses spatially photocleavable DNA barcodes to target panels of mRNA (or proteins) (126), is used for targeted analysis of small regions of tissue including for the study of lung injury in humans resulting from COVID-19 infection (127).

Future directions

There are a wealth of opportunities for non-destructive 3D pathology to be applied in diverse areas that can impact patient care, from translational research and preclinical drug development to clinical assays. Each of these scenarios offers multiple ways in which complementary technologies and methods may be coupled with 3D pathology to provide novel capabilities, some of which are outlined below.

For translational investigations, 3D tissue models, such as spheroids and organoids, are increasingly valued over traditional 2D tissue cultures (i.e. cell monolayers in a dish), as shown in Fig. 5. Such constructs recapitulate the complex and heterogeneous microenvironment of tissues with high fidelity, and thus lead to insights that translate more readily to human biology. There is an obvious need and value for high-throughput volumetric imaging of such tissue constructs, including in vitro imaging of dynamic processes over time. Since living cells and tissues, even at the scale of a few hundreds of microns, contain refractive heterogeneities that lead to light scattering and aberrations, there is a technical challenge for volumetric microscopy approaches to mitigate these effects for deep imaging with high contrast and resolution. Examples include adaptive optics techniques to dynamically compensate for tissue-induced aberrations (128–131), and multi-view imaging approaches that can provide a larger volumetric field of view by viewing the specimen from different angles and/or directions (33, 132–134).

Figure 5. Microscopy from cells and organoids to organisms.

There are clear advantages in terms of biological realism and insights when imaging whole organisms or 3D cultures vs. traditional cell monolayers on petri dishes. However, there are exponentially greater challenges in terms of imaging speed/throughput and dataset sizes. With 3D pathology methods, the computational hurdles represent the next frontier for many biological and clinical applications.

For translational investigations and preclinical drug development, the use of animal models, and especially rodent models, will continue to be an important and necessary step towards gaining improved insights and for hypothesis testing in preparation for human studies. Here, a major challenge for 3D pathology is the size of the tissues that must often be interrogated, which can approximate the size of clinical lesions (i.e., millimeters to centimeters in extent). Not only does this introduce challenges in terms of tissue labeling and clearing, but also in terms of imaging times, data storage, and big-data analysis. Tissue labeling can be particularly challenging. While animal models can often be engineered to express fluorescent proteins, clever approaches are being devised (as mentioned previously) to facilitate rapid exogenous labeling with small and large molecules (e.g., antibodies), as well as rapid optical clearing of entire organs (e.g., mouse brains) and organisms (e.g., whole mice). Finally, with the popularity of approaches for massively multiplexed DNA and RNA in situ hybridization and whole-genome/transcriptome sequencing of 2D histology sections, there is a push towards extending these “spatial biology” and “spatial omics” techniques into 3D tissues, which will necessitate additional advances in automated fluidic buffer exchanges for cyclic barcoding and labeling rounds, and other innovative approaches for imaging highly multiplexed information from 3D tissue volumes labeled with diverse molecular probes.

In a previous section, we outlined some of the challenges and solutions being devised to handle the acquisition and immediate post-processing of large 3D pathology datasets. For example, multi-resolution imaging workflows will likely be essential for most applications of 3D pathology in the near future, other than large-scale efforts in neuroscience and other “big-science” endeavors supported by large philanthropy or governmental mandates. For downstream image-analysis, however, many more challenges exist. While a discussion on machine-learning and image-analysis methods is beyond the scope of this article, a major challenge for the field of machine learning will be to generate ground-truth annotations for 3D segmentation and classification(135, 136). Recent efforts in weakly supervised learning and “annotation-free” segmentation are attempting to overcome some of these barriers (82, 137–140). Another challenge is the memory constraints of current GPUs, which limits the ability to process high-resolution 3D pathology datasets over sufficiently large “chunk” volumes to capture the necessary spatial context for diagnostic determinations(141). Even if chunk sizes can be accommodated by GPUs, training and inference times can still be prohibitive without parallelization across large banks of GPU arrays. An interim solution may be to analyze 3D pathology datasets as stacks of individual 2D images, or with “2.5D” methods that incorporate the information from a few adjacent 2D levels to improve the analysis of each 2D cross section within a volumetric stack (82, 142, 143).

For preclinical drug development, the value proposition for 3D pathology includes the ability to identify mechanisms of action for lead candidates, and to identify toxicities at earlier stages. Both of these goals can translate into significant cost and time savings for researchers and pharma companies. It is likely that drug developers will play a significant role as early adopters of 3D pathology, helping to refine these methods and paving the way for clinical adoption in the future.

Clinical applications of 3D pathology can be largely grouped into (1) early and accurate diagnosis, (2) improved prognostication for clinical decision support, and (3) predictive assays to identify ideal patient candidates for specific treatments such as chemotherapy and immunotherapies. As detailed in a recent perspective article(144), 3D pathology has the potential to complement traditional 2D histology by offering orders-of-magnitude more tissue coverage with volumetric insights. This will enable more-accurate quantification of complex tissue structures, such as glandular network that are disrupted in a predictable way as a function of disease progression and aggressiveness, as well as improved quantification of complex cell distributions that are highly heterogeneous, such as the tumor-immune microenvironment that plays a critical role in patient prognosis and response to various drug therapies (and especially immunotherapies). Finally, there are many “rare events” that are difficult to spot on 2D histology sections due to the minimal amount of tissue that is sampled by such techniques, but which can be more-readily identified and quantified through 3D pathology of large tissue volumes. Examples of such rare events that hold prognostic importance include tumor cells that are showing signs of pre-metastastatic behavior (e.g., lymphvascular invasion and perineural invasion) and minimal-residual disease after radiation or chemotherapy.

To provide equitable and accurate clinical decision support, computational pathology (i.e., “pathomics”) should ideally be combined with other modalities such as radiomics and genomics, as well as clinical data in electronic health records. The advantage of 3D pathology is that it can enhance other “omics” technologies within the clinic. For example, non-destructive imaging will allow larger amounts of tissue to be available for downstream molecular analyses. In addition, scanning large tissue volumes vs. thin tissue sections will enable improved co-registration with radiology images for combined pathomic and radiomic analyses.

Summary

In this review, we have outlined the potential biomedical impact and technical challenges of non-destructive 3D pathology. In terms of the technical strategies behind various implementations of 3D pathology, Fig. 6 provides a list of some of the most-common steps that we have mentioned for tissue preparation, 3D imaging, and post-processing/analysis. Many future innovations will focus on improving the throughput of these various steps, ideally shifting the time scales from hour and days to minutes. As discussed in the previous section, exciting new applications of 3D pathology – ranging from translational research and preclinical drug development to clinical assays – will continue to emerge with the hopes that many of them will be rigorously validated and incorporated into standard preclinical and clinical workflows.

Figure 6. Examples of key considerations and steps in 3D pathology.

Various applications of 3D pathology will each have unique technical requirements and time scales for tissue preparation, high-throughput 3D microscopy, and data handling/analysis. A few examples of technical steps are listed.