Tissue clearing and 3D reconstruction of digitized serially sectioned slides provide novel insights into pancreatic cancer

Abstract

Pancreatic cancer is currently the third leading cause of cancer death in the United States. The clinical hallmarks of this disease include abdominal pain that radiates to the back, the presence of a hypoenhancing intrapancreatic lesion on imaging, and widespread liver metastases. Technologies such as tissue clearing and three-dimensional (3D) reconstruction of digitized serially sectioned hematoxylin and eosin stained slides, can be used to visualize large (up to two to three-centimeter cube) tissues at cellular resolution. When applied to human pancreatic cancers, these 3D visualization techniques have provided novel insights into the basis of a number of the clinical characteristics of this disease. Here we describe the clinical features of pancreatic cancer, review techniques for clearing and the 3D reconstruction of digitized microscope slides, and provide examples that illustrate how 3D visualization of human pancreatic cancer at the microscopic level has revealed features not apparent in 2D microscopy, and in so doing has closed the gap between bench and bedside. Compared to animal models and 2D microscopy, studies of human tissues in 3D can reveal the difference between what can happen and what does happen in human cancers.

eTOC blurb

Kiemen et al. provide a comprehensive overview of two techniques for the 3D visualization of pancreatic cancer: tissue clearing and 3D reconstruction of serially-sectioned tissues. The authors explain the methodological differences as well as the advantages of each technique while putting the benefits of 3D analysis in a clinical context.

Introduction

Pancreatic cancer robs those diagnosed of hope ¹. It has been estimated that globally 495,773 people were diagnosed with pancreatic cancer in 2020, and that 466,003 died from the disease ². As bad as these statistics are, they appear to be getting worse. Rahib and colleagues have predicted that deaths from pancreatic cancer will become the second leading cause of cancer-related death in the United States by the year 2030 ^3–5. New approaches to understanding, diagnosing and treating this disease are needed.

Historically pancreatic cancer has been diagnosed and studied using two-dimensional (2D) microscopy ^6,7. These studies in 2D significantly under sample the cancers. If the greatest dimension of the average surgically resected pancreatic cancer is 3cm, and if the pathologist examines one 5 micron thick hematoxylin and eosin (H&E) stained slide section for every 1cm of tumor, then pathologists examine only 0.05% of the average surgically resected pancreatic cancer. This incomplete view has several implications. First, the connectivity of cancer cells is lost. Cancers that appear to be separate islands of neoplastic ducts embedded in stroma in 2D, are, in fact, often contiguous branching 3D tubes of cancer cells ⁸. Second, the prevalence of a feature, such as vascular invasion, will be underestimated as focal microscopic features are unlikely to be captured in the minuscule fraction of the cancer sampled. Third, relationships among cells are lost in thin sections unless the two cells coincidentally fall in the same 5 micron plane of section. For example, the relationship between an immune cell and a neoplastic cell cannot be appreciated unless the two cells are in the same plane of section ⁹. Fourth, the exact point at which an event occurs is not identifiable in 2D slides. In 2D we can observe cancer cells in a nerve, but we don’t know at what point along the full length of that nerve the cells invaded the nerve. Recent technological advances in the microscopic visualization of human tissues in 3D make the visualization of even large volumes at the cellular level possible, and in so doing overcome many of the limitations of 2D microscopy.

Here we provide a brief clinical overview of pancreatic cancer, we review two broad technologies for 3D visualization of cancers at the microscopic level, and we then show how the 3D visualization of human pancreatic cancer has provided insight into clinically important issues. Looking forward, we envision that true 3D multi-omic analyses of cancers on the subcellular level will soon revolutionize our understanding of cancer growth and metastasis.

Clinical

Signs and Symptoms

Intractable abdominal pain, particularly severe pain that radiates to the back, is one of the most dominant symptoms experienced by patients with pancreatic cancer ^1,10,11. This pain has been linked to cancer cells invading nerves (perineural invasion) and, when carefully examined, perineural invasion can be found in virtually all pancreatic cancers ^10,12–15. Perineural invasion, in turn, has been associated with poor prognosis in patients with pancreatic cancer ^10,16. Many of the nerves in the pancreas arise in either the celiac plexus or the superior mesenteric plexus, and pancreatic cancers can grow along nerves into the connective tissues around the celiac and superior mesenteric arteries ^12,16,17. Local recurrence, often in the area of the celiac and superior mesenteric plexuses, occurs in 15–25% of patients with pancreatic cancer who undergo surgical resection, and this local recurrence has been correlated with perineural invasion ^18–20. Local recurrence secondary to the involvement of nerves is such a significant problem, that some have suggested that the standard surgery (pancreatoduodenectomy) should be extended to include the resection of these nerve plexuses ²¹.

Signs and symptoms related to the obstruction of the pancreatic and bile ducts can also dominate the clinical picture for patients with pancreatic cancer. The main pancreatic duct and the distal common bile duct both pass through the head of the pancreas before they drain into the duodenum, and both ducts are often narrowed by cancers in the head of the gland ²². Obstruction of the bile duct leads to jaundice, while obstruction of the pancreatic duct causes upstream chronic pancreatitis with loss of both exocrine and endocrine cells ^23,24.

Other signs and symptoms of pancreatic cancer include deep venous thromboses, pulmonary emboli, migratory venous thrombophlebitis, weight loss and nausea ^25–27.

These signs and symptoms are non-specific, and most patients with pancreatic cancer are not diagnosed until after the disease has spread to other organs. The liver is the most common site of metastasis, and close to 80% of patients develop liver metastases ^28–30. The impact of this is devastating as the five-year survival rate for patients with liver metastases is only 3% ³¹

Diagnosis

Computed tomography (CT) can be performed to confirm the presence of a pancreatic lesion ³². When intravenous contrast is given, the majority of pancreatic cancers are visualized as ill-defined hypoenhancing (relative to the adjacent normal pancreas) intrapancreatic masses ^32,33. Other imaging modalities include magnetic resonance imaging (MRI) and endoscopic ultrasound (EUS) ³². EUS has the advantages that a biliary stent can be placed in patients with obstructive jaundice, and that biopsies for pathological examination and for molecular analyses can be obtained ^34,35. As a variety of neoplastic and non-neoplastic entities can form mass lesions in the pancreas, pathology is the “gold standard” in establishing the diagnosis ^6,7.

Pathology

Pancreatic cancers grossly form firm, white, ill-defined masses that replace pancreatic parenchyma ^6,7. At the microscopic level, when viewed in 2D, pancreatic cancers appear as individual glands haphazardly arranged in dense desmoplastic stroma ^6,7,36. One manifestation of this haphazard growth is the diagnostically helpful finding of a neoplastic gland immediately adjacent to a muscular vessel ³⁷. When examined with 2D microscopy, vascular invasion, including the invasion of lymphatic spaces as well as the invasion of veins, is identified in 65% of surgically resected pancreatic cancers ^6,7,38,39. As noted earlier, perineural invasion is seen microscopically in virtually all pancreatic cancers ¹⁶.

Compared to other cancer types, the stroma is particularly dense in pancreatic cancers ^6,7,36. This stroma is rich with activated fibroblasts, dense collagen, and immune cells ^36,40.

Treatment

The optimal treatment for patients with pancreatic cancer depends on the stage of the disease. Surgery, the best hope for a cure, is indicated for patients with low-stage disease ⁴¹. Close to 50% of surgically resected patients with stage T1 or T2 disease are alive five years after surgery ⁴². Radiation therapy is often also given, particularly in the United States, to achieve better local control of the disease, and for patients with locally advanced pancreatic cancer to shrink the cancer to the point where surgery is an option ⁴³.

Chemotherapy can be given before (neoadjuvant) and after (adjuvant) surgery, or as a primary therapy in patients with advanced disease ⁴⁴. Chemotherapy is the mainstay of treatment for patients with advanced disease, and several drug combinations, including FOLFIRINOX (fluorouracil, irinotecan, leucovorin, oxaliplatin), gemcitabine/nab-paclitaxel, and nanoliposomal irinotecan/fluorouracil, have been shown to improve survival by 2 to 6 months in patients with metastatic pancreatic cancer compared to single-agent gemcitabine ⁴⁴. Several investigators have suggested that inadequate drug delivery to the cancers accounts for the poor response of pancreatic cancer to systemic chemotherapy^36,45,46.

Despite improvements in survival for other cancer types over the past decade, the five-year survival rate for patients diagnosed with pancreatic cancer in the United States is an abysmal 11% ^31,47.

Recent advances in 3D microscopy reveal features of pancreatic cancer that are not apparent in 2D microscopy ⁴⁸. These features provide clinically relevant insights into the dominant characteristics of pancreatic cancer, and it is hoped that these insights will translate to improved patient outcome.

Techniques for 3D Microscopy

Improvements in tissue clearing, artificial intelligence (AI) and microscopy now make 3D visualization of large volumes of tissue at the cellular level possible ^49,50. Here we present two techniques, clearing and the 3D reconstruction of serially sectioned and digitized microscope slides, each with strengths and weaknesses, that have been applied to the 3D visualization of human pancreatic cancer (Figure 1) ^48,50–58.

Figure 1:

Two approaches to 3D histology include 3D reconstruction of serially sectioned slides (left side) and tissue clearing (right side). 3D reconstruction of serial sections includes fixing the tissue, embedding it (usually in paraffin), serially sectioning the block and then digitizing the resultant slides. The digital images can then be registered and cell types identified using deep learning techniques. Clearing includes fixation, permeabilizing the tissue to allow for the penetration of antibodies, labeling with antibodies, clearing and then imaging using either light sheet or confocal microscopy. Both procedures can produce 3D datasets amendable to microscopic visualization and quantification of histology not apparent in 2D (bottom third). (Copyright Johns Hopkins University Department of Art as Applied to Medicine).

Tissue clearing

The prerequisite to visualizing intact thick tissues is transparency, as light must be able to pass through the sample. While some specimens are near-transparent and may be successfully imaged without clearing, many human and animal tissues, and especially dense, fibrous pancreatic cancers, need to be rendered transparent to enable optical sectioning using modern microscopy techniques. We therefore use the term “tissue clearing” when referring to 3D immunolabeling and optical visualization in this review.

Two distinct procedures are necessary for 3D microscopy using tissue clearing. First, cells and tissues of interest are labeled. The principles follow those of 2D immunofluorescence, with the additional complexity of needing to label antigens deeper in the tissues. To accomplish this deeper labeling, the extracellular matrix has to be permeabilized and cell membranes partially delipidated (Figure 1) ^50,59. Human pancreatic cancer is especially challenging, as it is densely fibrotic ^36,60. Permeabilization can be increased through a combination of detergents, collagenases, acids, denaturants, and temperature changes ^48,61–63. The cells are then labeled using long, up to weeks, antibody incubation times (Figure 1). As the size of antibodies determines their ability to penetrate tissues, nanobodies, which are small (~15kDa versus monoclonal antibodies which are ~150kDa) single domain antibodies found in Camelidae (Alpaca, Llama), are emerging as a versatile option ^64–67. In addition to immunolabeling, the autofluorescence properties of specific tissues and cells can be exploited to visualize structures without labeling ^48,68. Sources contributing to autofluorescence in the pancreas include nicotinamide adenine dinucleotide phosphate (NADPH), laminins, collagens, elastin, porphyrins and lipofuscins ^68,69. Collagens and elastin account for the most intense autofluorescent signal in the pancreas, making autofluorescence a useful approach to visualizing vessels (as seen in Figure 2A and Supplemental Figure 1). Furthermore, arteries can be distinguished from veins based on the characteristic intense autofluorescence of the elastic lamina in arteries (Supplemental Figure 1). In the pancreas, excitation using a 488nm light-source can be used to highlight the autofluorescence, while autofluorescence is generally lower at higher excitation wave lengths (>640nm).

Figure 2:

A: Cleared human pancreatic cancer with surface rendering. Note the neoplastic cells (green) growing parallel to a vein (red, running down the center of the image from 12 o’clock to 6 o’clock) and its associated thicker artery (red, left side of the image). The neoplastic cells grow around a nerve (blue) and parallel to the vein. Supplemental Figures 1A and 1B are paired videos. B: Cleared human pancreatic cancer reveals that isolated glands in 2D (B1) are actually often longer tubes of cancer cells in 3D (B2). Supplemental Figures 2A and 2B are paired videos. C: Tissue cleared human pancreatic cancer. Selected images at serial levels illustrate that vascular invasion (C2) can be missed (C1) in 2D. In C1, cancer cells (green) are seen growing next to, not in, the vessel (red). C2 shows the point at which the cancer invades the vessel, and C3 shows the same cancer growing inside the vessel. Supplemental Figure 4 is a paired video that highlights the focality of vascular invasion. (A:Light sheet microscopy, 3.2x. Red = autofluorescence, green = carcinoembryonic antigen labeling, blue is S-100 protein labeling; B: Light sheet microscopy, 6.4x. Red = autofluorescence, green = labeling for cytokeratin 19; C: Light sheet microscopy, 10x. Red = autofluorescence, green = immunolabeling for carcinoembryonic antigen. Z-stack distance between left and middle image is 36μm; distance between middle and right image is 57μm).

The next step is to overcome the opaqueness of tissues so that the excitation and emission light can pass through the tissues.

Tissue clearing overcomes the opaqueness of tissues by creating tissues with a uniform refractive index (Figure 1) ^49,50. Living organisms are composed of cells and tissues with varying refractive indexes. Water has a refractive index of 1.33, while the refractive index of lipids and proteins can range from 1.4 to 1.6 ^49,70,71. Light bends as it passes from one medium into a medium with a different refractive index. As a result, complex heterogenous tissues become opaque. Tissue clearing overcomes this by reducing the heterogeneity of refractive indices in the tissues being studied ^49,70,71. This transparency can be achieved through aqueous based and solvent based approaches ^62,72. For the human pancreas, the combination of dichloromethane and dibenzyl ether have proven to be highly effective ^48,59,62,73.

The final step is microscopy. The most commonly used techniques are light sheet fluorescence microscopy (LSFM), optical projection tomography (OPT) and, for higher magnification, laser-scanning confocal microscopy or two-photon microscopy ^{59,61–63,69,74}. Each of these techniques has strengths and weaknesses.

Light-sheet fluorescent microscopy (Figure 2, and Supplementary Figure 1) uses laser beams to illuminate a plane of adjustable thickness within the tissue ⁷⁵. A synonym for light sheet microscopy is therefore selective plane illumination microscopy (SPIM). The objective capturing light emitted from excited fluorophores within the tissue is placed orthogonal to the laser beams and has a high numerical aperture ⁷⁵. Filters are selected to match the emission spectrum of the fluorophores and each excitation channel (wavelength of laser light) is recorded separately. Exposure times will vary depending on the thickness of the tissue, fluorophore signal intensity, and laser intensity, but it generally takes between 50ms and 220ms to record a single image (times are doubled when using a two-sided laser, which results in better light penetration in larger tissues). The sample can then be optically sectioned by moving the tissue through the plane of light. Depending on the size of the sample and the number of channels recorded, the acquisition time for an entire specimen can easily reach 2 to 5 hours. Long imaging times are possible as the low laser intensities used with LSFM typically do not produce significant photobleaching (the permanent alteration of a fluorophore, leading to loss of fluorescence signal). A downside of LSFM is that the axial resolution is limited to ~1μm, which is worse than the lateral resolution ⁷⁵.

Optical projection tomography (OPT) has also been widely used to visualize labelled and cleared samples. Technically, OPT is comparable to medical computed tomography, only OPT uses ultraviolet, visible, and near-infrared photons instead of x-ray photons ⁷⁶. Image contrast is achieved through absorption/scattering in transmission OPT, while it is achieved through fluorescence in emission OPT. The sample is usually rotated, allowing for the acquisition of multiple 2D images from different angles (2D projections). These 2D images are then used to reconstruct 3D volumes, however, the reconstruction algorithms add complexity to OPT 3D microscopy ⁷⁷. OPT and LSFM can often complement each other. In general, OPT is more suitable for larger samples recorded at lower magnification, while LSFM, although limited in axial resolution, can achieve higher resolution images of smaller samples.

Laser scanning confocal microscopy employs an array of pinholes to focus both the excitation (point illumination) light and the emission light ⁷⁸. The field-of-view is then scanned point-by-point ⁷⁸. Adaptation of the point of focus allows for optical sectioning and the subsequent creation of 3D volumes at extremely high (150X) magnifications. The penetration of excitation light and detection of emitted light is, however, limited, making laser scanning confocal microscopy not suitable for thicker tissues ⁷⁸. Two-photon microscopy is another point-scanning fluorescent microscopy technique, that reduces photobleaching by using two photons of higher wavelength instead of a single photon of lower wavelength ⁷⁹. Photons are emitted almost synchronously, only around 50 femtoseconds apart, which results in the same excitation signal but with significantly less photobleaching and reduced scattering in deeper tissue sections. These scanning fluorescent microscopy techniques come with a trade-off, as although they can generate extremely high-resolutions images, the field of view is generally limited compared to LSFM.

The data sets produced by these various microscopy techniques can be further processed using surface or spot rendering, enabling shape and co-localization analyses. Pattern recognition algorithms can also be applied to trace complex neuronal or vascular networks and to identify connections that may not be apparent to the human eye ⁸⁰.

3D reconstruction of digitized serial microscope slides

Several methods have been developed to create 3D reconstructions of microscopic slides from serially sectioned samples ^52–57. Tissues are typically serially sectioned, the slides digitized, and then the digital images are registered to create digital tissue volumes (Figure 1) ^52–57. Manual segmentation or deep learning can then be used to label components in the digitized images (Figure 1) ^52,81. In the pancreas as many as ten structures (normal ductal epithelium, pancreatic precancers, pancreatic cancer, islets of Langerhans, vessels, nerves, acini, stroma, lymph nodes, and fat) have been labeled to a resolution of 1–2 microns ⁵². Unused intervening sections can be used for immunohistochemistry, imaging mass cytometry, somatic sequencing and for spatial transcriptomics, creating 3D integrations of histologic, gene, protein and immune features ^52,82–88.

In this review, when describing the 3D reconstructions of microscopic slides, we will focus on the approach called CODA, as CODA has been applied extensively to human pancreatic tissues ⁵². CODA relies on four major steps: image registration, cell detection, tissue segmentation, and data visualization (Figure 1). The registration step maximizes the 2D cross correlation of pixel intensity between pairs of images to correct for the rotation, translation, folding, splitting, and stretching of tissues that occurs during microtomy. The cell detection step uses color deconvolution to isolate the hematoxylin channel, and 2D peaks in this channel are then used to define nuclear coordinates. The tissue identification step then uses a small dataset of manual annotations to train a deep learning semantic segmentation algorithm. Once trained, the algorithm can efficiently label, to a resolution of 1 micron, microanatomical structures in the pancreas such as ductal epithelium, acinar tissue, cancer, and vessels. Finally, the image registration, cell detection, and tissue segmentation are integrated to create 3D renderings of pancreas microanatomy at large scale (up to cm³) while maintaining single cell resolution.

Clearing and 3D reconstruction of digitized serial microscope slides are not mutually exclusive, as, after imaging, cleared tissues can be washed, formalin-fixed and paraffin-embedded and then serially sectioned for 3D reconstruction ^48,69,89. Although the resultant images may have retraction artifacts, they can none the less be useful in validating observations made in cleared tissues ⁴⁸.

Strengths and weaknesses

The primary differences between clearing and 3D reconstruction of serial microscope slides relate to the size of tissues, and the number of markers, that can be examined (Table 1).

Table 1.

A comparison of two methods for visualizing human pancreatic cancer in 3D at the microscopic level

Feature	Clearing	3D Reconstruction of serial microscope slides
Size of sample that can be studied	Entire human brains have been studied.	Determined by the size of the microscope slides (x,y plane) and by number of sections cut (z plane) .
Application to densely fibrotic pancreatic cancer	Limited by poor antibody penetration.	Not limited by antibody penetration. Uniform resolution even deep in tissues.
Number of tissue types that can be studied in human pancreatic cancer	Usually three or four fluorophores (antibodies) plus autofluorescence of vessels	10+ distinct tissues labeled with CODA
Application of other technologies such as sequencing	Limited by the harsh chemicals used.	Limited only to technologies applicable to formalin-fixed paraffin-embedded tissues.
Cost	$	$ $ $ (proportionate to number of slides sectioned)
Time to process one sample	4–50 days (depending on protocol used, use of unconjugated or conjugated antibodies/nanobodies).	Sectioning / scanning: 1 week Computer processing: 1 week
	Upscaling is easily possible, as samples can be run in parallel.
Quality of images	Visually stunning,	More easily related to standard hematoxylin and eosin stained sections.
Nuclear detail	Often lost. Propidium iodide can be used to detect nuclei in labelled cells, but details are often lost.	Retained.

Clearing can be used to visualize whole organs, including entire human brains ^59,90. However, clearing relies on antibody penetration, and antibodies do not penetrate deeply into densely fibrotic tissues such as human pancreatic cancer ^48,62. By contrast, the potential size of the tissue that can be reconstructed using 3D reconstruction of microscope slides is limited only by the size of the slides (x,y-direction) and by the number of sections cut (axial, or z-direction) ⁵². Additionally, since the tissue is sectioned in the generation of serial microscope slides, all tissue is visualized homogeneously; there are no “dark” regions as happens when there is poor antibody penetration using clearing. Several approaches have been developed to overcome the limitations of antibody penetration, including first sectioning larger samples into smaller slabs, clearing and labeling the smaller slabs individually, and then stitching the visualized slabs together using imaging software ⁶³. At the other end of the scale, tissue clearing can also be applied to organoids alone or combined with other tissue engineering approaches, including tissue engineered microvessels ^74,91. The sizes of tissue that can be examined and resolutions that can be achieved by the two approaches are summarized in Table 2.

Table2.

Comparison of resolutions and sample sizes for human pancreatic cancer

	Light sheet microscopy	Serial sections tissue
lateral resolution	>0.3 μm	~0.25 (40x) or ~0.5 μm (20x)
axial resolution	>1 μm	>4 mm (tissue section thickness)
Sample size size limit in x-, y-dir.	~15 X 15 mm2 (LSFM)	~ 20 × 50 mm2
Sample size size limit in x-, y-dir.	< 10 mm	unlimited
z - continuity	good	moderate

LSFM= light sheet fluorescence microscopy

In 3D reconstruction of serial microscope slides, artificial intelligence-based segmentation allows rapid identification of multiple tissue components that are distinguishable in hematoxylin and eosin (H&E) stained sections (in the human pancreas, this includes 10 labels with CODA) ^52,81. By contrast, only four or five components are typically identified with clearing; three or four using antibodies with different fluorophores, and one taking advantage of distinct patterns of autofluorescence ^48,52,69.

The main advantage of 3D reconstruction of serial microscope slides is that, because unstained intervening slides are available, this approach is readily integrable with any technique that can be applied to formalin-fixed, paraffin-embedded tissues, such as: immunohistochemistry, image mass cytometry, somatic sequencing, and spatial transcriptomics ^52,86,87. This level of integration is generally not possible with clearing because of the harsh chemicals employed.

Insights from 3D

The 3D visualization of human pancreatic tissues has provided clinically important insights into pancreas pathology that are not apparent in 2D (Figure 3 and Table 3). In standard 2D H&E stained microscope slides, pancreatic cancer appears as isolated glands haphazardly scattered in desmoplastic stroma (Figures 2B1 and 3E, and Supplementary Figure 2A). When visualized in 3D, it is clear that invasive pancreatic cancer cells, in fact, can grow as continuous branching tubes (Figures 2B2 and 3E, and Supplementary Figure 2B) ⁸. These tubes do not grow randomly, but sometimes grow along tissue planes that are not appreciated in 2D sections. While the collagen fibers in the stroma of the pancreas appear disorganized when viewed in 2D, in the plethora of planes visualizable in 3D, these fibers can be shown to be well-oriented, particularly around pancreatic ducts and vessels (Figures 3A and 4A, and Supplementary Figure 3A) ^{48,52,92–96}. Benias and colleagues recently described previously unrecognized interstitial spaces lined by cells expressing endothelial markers between well-oriented collagen fibers around vessels and ducts in the biliary tree, and they hypothesized that these spaces are present widely in the body and that these spaces are a pathway for cancer spread ⁹⁷. The patterns of growth observed in 3D suggest that pancreatic cancer may invade using the microchannels identified by Benias. Indeed, well-oriented collagen fibers and interstitial spaces in the connective tissue surrounding veins in the pancreas may explain the propensity of pancreatic cancer to grow parallel to veins (Figures 2A and 3C, and Supplementary Figure 1) ⁴⁸. What appears as random in 2D, when studied in 3D, is found to be a propensity for growth parallel to vessels ⁴⁸.

Figure 3:

Illustration highlighting some features identifiable in 3D microscopy, that would be missed/incorrectly interpreted in 2D histology (blue slices). A: Collagen fibers that appear haphazardly aligned in 2D, can be seen to have an organization in 3D. B: Counts of immune cells in in 2D only sample a small portion of what is shown to be a heterogenous immune infiltrate in 3D. C: Venous invasion is almost ubiquitous in 3D, however it can easily be missed in 2D sections. D: The extent of perineural invasion visualized in 3D can be underestimated in 2D microscopy. E: Invasive carcinoma appears as isolated glands in 2D is often shown, in fact, to be contiguous branching tubules when viewed in 3D. (Copyright Johns Hopkins University Department of Art as Applied to Medicine).

Table 3.

Comparison of Clinical Correlates of 2D and 3D microscopy

Clinical feature	2D microscopy	3D microscopy
Pain	Perineural invasion-visualized for a short distance.	Perineural invasion-visualized for a longer distances (> 4mm).
Hypoenhacement on imaging	Desmoplastic stroma.	Venous invasion resulting in reduced outflow which, in turn, diminishes inflow.
Local dissemination of disease	Perineural invasion.	Perineural invasion and continuous and discontinuous intravenous spread.
Propensity for metastases to the liver	Unclear	Venous invasion is ubiquitous.
Pathology diagnosis	Gland next to vessel is a diagnostic feature, likely secondary to haphazard growth of the cancer cells.	Carcinoma specifically grows in well-oriented collagen planes and parallel to veins.
Screening/early detection	Single cross section of pancreatic intraepithelial neoplasia lesions.	Size of precursor lesions can be determined when contained in a tissue block, and they can be studied using a variety of ancillary technologies.

Figure 4:

Three-dimensional reconstructions of serially sectioned human pancreas using CODA. In panel A, note how the collagen fibers in one plane appear poorly aligned, and in another appear well-aligned. B: The precursor lesion (red) measures 1.2cm in one plane and 2.6mm in another. If measured as 2.6mm the lesion would be classified as pancreatic intraepithelial neoplasia (PanIN), but if measured as 1.2cm it would be classified as an intraductal papillary mucinous neoplasm. In panel C the intervening unstained sections of a human pancreatic intraepithelial lesion were immunolabeled for CD45. In 3D, the intensity of the inflammatory infiltrate can be seen to range from minimal (blue box) to intense (red box). (A and B, H&E, C, immunolabeling for CD45) Supplemental Figures 3A and 3B are paired videos.

The growth of pancreatic cancer parallel to vessels may, in turn, explain the high prevalence of venous invasion in pancreatic cancer. Venous invasion is observed in 60–65% of surgically resected pancreatic cancers studied in 2D, but when studied in 3D venous invasion can be identified in nearly 100% of the cancers (Figures 2C and 3C, and Supplementary Figures 1 and 4) ^38,39,51. The ubiquitous presence of venous invasion in even low-stage pancreatic cancers has important clinical implications. First, since the veins of the pancreas drain into the liver, venous invasion may explain the high prevalence of liver metastases and, ultimately, why pancreatic cancer is so deadly ⁹⁸. Second, because veins thrombose in areas of venous invasion, venous outflow from pancreatic cancers will be reduced ³⁸. With diminished outflow, the inflow of blood into the tumor will be reduced. Venous invasion may therefore also contribute to the hypoenhancement on imaging and the poor delivery of systemic chemotherapies to pancreatic cancers ^38,98.

The high prevalence of perineural invasion, noted in 2D studies, has also been observed in 3D studies. In 3D one can appreciate better the length of growth of the cancer along nerves ^10,52. Perineural invasion of 2500 microns has been reported in 2D histologic slides, while in 3D distances as great as 4 mm are readily visualizable (Figures 3D and 5A, and Supplementary Figure 5) ^10,99. These distances highlight the role of perineural invasion in local spread of disease and therefore in local recurrence ^18,19,99. Similarly, 3D examination of foci of venous invasion have identified foci in which cancer cells grow in continuity inside the vein for 2 mm before growing out into the stroma (Figure 5B) ⁵². These findings highlight venous invasion as a mechanism of intraparenchymal spread of pancreatic cancer.

Figure 5:

Three-dimensional reconstructions of serially sectioned human pancreas using CODA. Panel A highlights invasive carcinoma (yellow) invading a nerve (brown) and then extending for a distance of 4mm along the nerve. Note that the “moment” the cancer invades the nerve can be identified in this 3D visualization. In panel B the invasive carcinoma (yellow) extensively grows within a vein (green) for a distance of at least 2mm. Seven points can be identified where neoplastic cells cross the media of the vessel. Supplemental Figure 5 is a paired video.

Studies using 3D microscopy, because they can be used to identify the exact point at which a process occurs, can also help advance our understanding of disease mechanisms (Figure 5, and Supplementary Figure 5) ⁵². For example, studies of the “moment of venous invasion” using tissue clearing have shown that epithelial to mesenchymal transition is only transient during vascular invasion⁵¹.

3D analyses have also uncovered flaws in the histologic classification of precursor lesions. In the current classification system, based on 2D microscopy, pancreatic intraepithelial neoplasia (PanIN) lesions are distinguished from intraductal papillary mucinous neoplasms (IPMNs) based on the size of the lesion ^100,101. PanINs are, by definition <0.5cm, while IPMNs are, by definition, ≥1.0cm ^100,101. However, some lesions that meet diagnostic criteria for a PanIN in 2D (they are <0.5cm in greatest dimension in the plane that they happened to be sectioned), when visualized in 3D, are found to involve >1.0cm of duct, and thus meet the size threshold for an IPMN (Figure 4B) ⁵². A classification system based on volumes, as is done with CT, would allow for lesion classification independent of the plane of section.

When lesions are entirely confined to a block of tissue, 3D analyses can also be to quantify the number of cells in that lesion ⁵². Since PanINs and other precursors are a target for early detection efforts, understanding the exact size of PanINs and the number of cells in each PanIN will inform efforts to develop early detection tests.

The most significant advantage of CODA is that only every third slide has to be used to generate the 3D images ⁵². The two intervening slides can be used for molecular or immunolabeling studies ⁸⁶. A lesion can be identified in 3D and then the intervening unstained slides of that lesion can be laser capture microdissected and genetically sequenced, providing a 3D map of somatic mutations in the lesion. An understanding of the 3D architecture of somatic genetic alterations in complex branching and bending structures is needed to define molecular heterogeneity.

Similarly, intervening unstained sections can be used for immunolabeling, providing a 3D architecture of the immune reaction associated with a specific lesion (Figures 3B and 4C, and Supplementary Figure 3B) ^9,86.

Summary and conclusions

Visualizing human pancreatic cancers in three dimensions at the microscopic level has provided insights into the biology and clinical behavior human pancreatic cancer ^48,51,52. We envision two advances that will transform this field even further. First, as spatial transcriptomics technologies are optimized for formalin-fixed paraffin-embedded tissues, detailed 3D spatial transcriptomics will soon be possible ^102,103. Second, as the resolution of clinical imaging techniques such as computed tomography and magnetic resonance imaging, improve, we foresee the direct, one-to-one correlation of 3D clinical imaging with 3D microscopy ^104,105. AI will only increase the impact and speed the rate of advance of these analyses ¹⁰⁶.

The earth is not flat. Neither are human tissues.

Supplementary Material

1

Supplemental Figure 1: Video of cleared human pancreatic cancer highlighting the neoplastic cells growing parallel to a vein. A: Serial images show the cancer (green) growing along a nerve (blue) and parallel to the vein (red). The wavy autofluoresence of the artery (red) can also be appreciated. B: 3D rendered video of the cancer shown in A. The neoplastic cells (green) grow parallel to a vein (red, running down the center of the image from 12 o’clock to 6 o’clock) and its associated thicker artery (red, left side of the image). (Light sheet microscopy 3–2x. Red=autofluorescence, green= labeling for carcinoembryonic antigen, blue= labeling for S100).

3

Supplemental Figure 2: Video of cleared human pancreatic cancer (see Figure 2B1 and 2B2) highlighting that the individual ducts of cancer (green) visualized in 2D (video A) actually connect in 3D (video B). (Light sheet microscopy, 6.4x. Red = autofluorescence, green = labeling for cytokeratin 19).

5

Supplemental Figure 3: A) Video of three-dimensional rendering of serial sections using CODA shows periductal and intralobular collagen sheets (pink) surrounding a pancreatic duct (blue). B) Video of pancreatic precursor with overlaid local immune cell density shows clear regions of low inflammation (grey) near the margins of the tumor and high inflammation (red) near the central duct of the tumor.

7

Supplemental Figure 4: Video of 3D clearing showing venous invasion in certain sections only, see also Figure 2C. Note how early in the video the cancer cells (green) do not appear to involve the vessels, while clear venous invasion can be seen at 7seconds. (serial images, light sheet microscopy, 10x, red = vascular autofluorescence, green = labeling for carcinoembryonic antigen).

8

Supplemental Figure 5: Video of perineural invasion generated from serial sections using CODA shows cancer (yellow) wrapping around and extending along a nerve (brown) for >4mm.