Blueprints from plane to space: outlook of next‐generation three‐dimensional histopathology

Abstract

Here, we summarize the literature relevant to recent advances in three‐dimensional (3D) histopathology in relation to clinical oncology, highlighting serial sectioning, tissue clearing, light‐sheet microscopy, and digital image analysis with artificial intelligence. We look forward to a future where 3D histopathology expands our understanding of human pathophysiology and improves patient care through cross‐disciplinary collaboration and innovation.

This article reviews recent advancements in three‐dimensional histopathology in relation to clinical oncology, featuring serial sectioning, tissue clearing, light‐sheet microscopy, and digital image analysis using artificial intelligence. 3D histopathology is expected to expand our understanding of human pathophysiology and improve patient care.

1. INTRODUCTION

The microstructures of organs, tissues, cells, and their disease‐related alterations have captivated the interest of biologists and physicians for centuries. To aid this investigation, pioneers in the field invented and adapted a wide range of methodologies, including tissue sampling devices, chemical fixation, sectioning, staining, and microscopic observation with light and electron microscopy. Clinical histopathology is the application of these methodologies to diagnose and study disease. Historical pathologists such as Rudolf Virchow and Ludwig Aschoff described the strong relationship between morphological changes and physiological abnormalities, ¹ building the foundations of this medical field. Traditionally, micrometer‐thick sections are prepared from a patient's lesion‐derived tissue and stained with hematoxylin and eosin (H&E) before being observed using an optical microscope. This method allows for the comprehensive structural examination of tissue sections in great detail.

Today, histopathology is a widely accepted, reliable clinical procedure due to its comprehensiveness, reproducibility, high‐throughput, and accumulated experiences in addition to its commercial support. However, the traditional approach of sectioning tissue limits the observation of the human body's innately three‐dimensional (3D) structures to individual two‐dimensional (2D) planes. This poses a dilemma; there is an intrinsic risk that, while the objective interpretation is warranted within a 2D plane, perception outside of the section is dependent on the experience and subjectivity of a histopathologist. In this regard, it may not be an exaggeration to say that some of these contemporary techniques are just waiting to be updated.

A number of experts in basic science and clinical medicine have become aware of these limitations and have started various attempts to explore the 3D microstructure of histopathological specimens. This review briefly summarizes the relevant literature and future perspectives of 3D histopathology, especially in relation to oncology.

2. ISSUES IN CURRENT 2D‐BASED HISTOPATHOLOGY

In general, the diagnostic process of tumorous diseases in histopathology consists of three main domains: lesion detection, structural assessment, and spatial quantification. Lesion detection is the initial step, where a microscopist examines a specimen to identify and determine the location of any pathological abnormalities. The target lesion can be large enough to be easily recognizable during macroscopic examination or by low‐power observation with a microscope. An issue arises when the lesions are so small and spatially dispersed within a tissue that they are unlikely to be present in thin sections, especially if the number of slices examined is insufficient. In such cases, a microscopist must carefully inspect wide areas of serial sections to avoid underdiagnosis. More specifically, underdiagnosis may occur when concomitant higher‐grade components and surgical margin involvement are overlooked. Pathological guidelines and recommendations advise extensive block sampling of specimens for the diagnosis of some peculiar diseases or situations, including follicular thyroid carcinoma, signet‐ring cell carcinoma of the stomach, breast carcinomas, ovarian and fallopian tumors, cervical intraepithelial neoplasia of the uterus, prostatic adenocarcinoma, and germ cell tumors of the testis. ² , ³ , ⁴ However, due to the increased work associated with block sampling, this process is problematic for pathologists in hospitals and institutes with a large caseload.

After the detection of lesions, a structural assessment follows as the next crucial step. It is generally accepted that tumors grow in characteristic patterns, depending on their origin, subtype, differentiation, and aggressiveness. In addition, some uncommon tumors display a noticeable pathognomonic motif, leading to a straightforward diagnosis. However, the typical thickness of histologic sections is thinner than the diameter of cells, making it insufficient for the evaluation of cell assemblies, such as organ units and tumor clusters. In addition, there is no clear distinction between transverse, longitudinal, or oblique sections. As a result, the description of structural characteristics tends to be rhetorical and subjective, depending on the experience and impression of the examiner. Non‐mucinous adenocarcinoma, a prevalent lung cancer, is a good example that is challenging to evaluate in 2D observation. This type of tumor requires subcategorization into five components (lepidic, papillary, acinar, micropapillary, and solid) which have different degrees of aggressiveness and greatly impact the patients' prognosis. However, the reproducibility of 2D‐based subtyping is low even among experts, ⁵ indicating the risk of variations in the estimated size of a lesion and its T stage.

Spatial quantification is used to measure the extent of tumor growth, such as the size and the number of foci, which are generally correlated to its progression and aggressiveness, and are helpful to predict the patients' prognosis and treatment response. In most cases of tumor diagnosis, the diameter of the largest focus is considered representative and is often expressed with T and N stages of the TNM classification issued from the Union for the International Cancer Controls and others. ⁶ However, in conventional 2D‐based methods, these measurements rely on the human eye's perception of tumor size and number as displayed on the surface of grossly sliced tissue or micrometer‐thick sections on a glass slide. Identifying the optimal angle and location to expose the most representative cross‐section of the tumor is challenging. Moreover, when a tumor has an irregular shape, the largest diameter might not represent the best indicator. As a result of the aforementioned limitations, there is a possibility that the accumulated data based on these traditional methods may contain errors.

Conventional histopathologic methods are currently struggling to overcome these intrinsic limitations and meet advanced demands. From a data science point of view, such as the Nyquist–Shannon sampling theorem, ⁷ many difficulties in conventional 2D histopathology are attributable to the low sampling rate (Figure 1). In typical procedures for routine diagnosis, only a limited number of tissue blocks and thin sections are extracted from a centimeter‐sized resected specimen or organ, and these contain only a small fraction of overall information. A multisectioning approach is a common solution to deal with this issue, where the tissue blocks are sliced in an intermittent or sequential manner to sample multiple planes. However, this process is inevitably destructive and results in the loss of structural continuity. Moreover, even though it is used by histopathologists when necessary, this approach is not suitable for routine use as it multiplies the workload of technologists and pathologists, causing delays in turnaround time.

FIGURE 1.

Issues in the two‐dimensional (2D) approach to histopathology. Histopathological evaluation encompasses (1) lesion detection, (2) structural assessments, and (3) spatial quantification. The conventional thin section, either single or multiple, extracts just a small fraction of the lesion in three‐dimensional (3D) space and cannot fully describe all of its characteristics. Because of this, experience has been needed to compensate for and predict the “invisible gaps” between these sections. This intrinsic issue can be resolved by obtaining 3D voxel data across the whole specimen. The 3D histopathological approach involves whole‐sample serial sectioning or tissue clearing with light‐sheet fluorescence microscopy (LSFM), enabling the detection, assessment, and quantification of all lesions in a specimen without requiring structural predictions by well‐trained experts.

3. 3D HISTOPATHOLOGY: EXPANDING THE LIMITATION OF CONVENTIONAL 2D HISTOPATHOLOGY

3D histopathological approaches encompass specialized tissue processing and imaging technologies that can be used to obtain continuous spatial information of an entire specimen (Figure 1). Conceptually, the data obtained by 3D histopathological approaches is expected to overcome the limitations of 2D observation due to its capability of revealing true, quantitative features of whole 3D biological tissues. In the context of tumor diagnosis, these 3D data enable pathologists to find complex or rare structures inside the legions, such as tumorlets, microinvasion, and micrometastasis, that are challenging to detect and categorize with a single thin section (Figure 1). Any metrics of a tumor, including its diameter, surface area, volume, number of separated foci and their distribution can be inclusively and accurately accessed by using the 3D data.

Technical developments supporting the practice of 3D histopathology are emerging. The procedure can be implemented by continuous serial sectioning ⁸ or tissue clearing (Figure 2A). ⁹ , ¹⁰ The former approach is more familiar to pathologists because it often employs traditional thin sections with H&E staining. However, since it requires the highest level of precision in sectioning and staining of the specimen, the implementation of an automated system is usually required. Moreover, this sectioning process is inherently irreversible, physically dividing the sample into independent slices. To reconstruct 3D data, it is necessary to perform accurate slice‐to‐slice registration on the serialized sections, for which advanced algorithms have been developed (Figure 2B). ¹¹ , ¹² Nevertheless, this approach has begun to gather interest for its potential in 3D histopathology. In a representative study, researchers reconstructed continuous 3D H&E‐stained histological sections and, using machine learning segmentation algorithms, elucidated previously unobserved spatial relationships between neoplastic pancreatic cells and adjacent normal tissues (Figure 3A). ⁸ A robust 3D registration algorithm that is compatible with both bright‐field and fluorescence microscopy has also been developed (Figure 3B). ¹³

FIGURE 2.

Current three‐dimensional (3D) histopathology techniques. (A) Representative workflow of current 3D histopathology. Upper panel: workflow for serial sectioning‐based 3D histopathology: The fixed surgical specimen undergoes serial sectioning, followed by preparation into histopathological examination‐compatible tissue slides. Post hematoxylin and eosin (H&E) staining, the serially sectioned slides are scanned using an automated slide scanner. The acquired images are then reconstructed into a 3D volume following a multilayered pre‐processing, registration, and segmentation process. Lower panel: workflow of tissue clearing‐based 3D histopathology: The fixed surgical specimen initially undergoes delipidation and decoloration (a procedure to enhance the clearing efficiency), followed by 3D staining utilizing dyes or antibodies. The 3D‐stained specimen is subsequently processed with a refractive index matching solution (the final clearing procedure), after which the cleared sample is imaged using light‐sheet fluorescence microscopy (LSFM). The obtained images are directly reconstructed into 3D volumetric data. (B) Comparison of serial sectioning and tissue clearing workflows: serial sectioning typically employs bright field imaging, whereas cleared tissues undergo fluorescence imaging. The image resolution is typically 0.25 × 0.25 × 3 μm for slide scanner images and 0.45 × 0.45 × 2.9 μm for cutting‐edge light‐sheet microscopy. ³³ The 3D reconstruction of images acquired via serial sectioning demands substantial effort due to the potential misalignments between serial sections. Conversely, images obtained with light‐sheet microscopy are more straightforward to reconstruct, given the continuous imaging of an intact specimen. The estimated image processing time ranges from hours to days per sample for both techniques. (C) Relationship between the sample size, the resolution, and the magnitude of the acquired data (denoted by the diameter and color of the circle). A substantial increase in data size occurs as the resolution or sample size is expanded. Obtaining submicrometer‐order resolution when imaging a centimeter‐order sample can generate data volumes in the hundreds of terabytes. When the resolution is on the order of a few micrometers, however, imaging a sample of centimeter order can be accomplished in several hundred gigabytes. See also Table S1. GB, gigabyte; IHC, immunohistochemistry; N/A, not applicable; RI, refractive index; TB,terabyte.

FIGURE 3.

Three‐dimensional (3D) histopathology applications. (A, B) Serial sectioning‐based cancer tissue examinations. (A) CODA. ⁸ Acquired serial sectioned images undergo registration, followed by 3D reconstruction and machine learning‐assisted segmentation, culminating in a comprehensive 3D analysis. The resulting 3D data enables highly flexible and precise analysis. (B) VALIS ¹³ enabled highly precise registration and subsequent 3D reconstruction. (C, D) Clear, unobstructed brain/body imaging cocktails and computational analysis (CUBIC) tissue clearing‐based examinations of cancer metastasis in patient‐derived lymph nodes. (C) An example of a colorectal cancer case. ¹⁸ The workflow includes conventional two‐dimensional (2D) evaluation at the maximum plane and whole 3D staining and imaging followed by post hoc 2D evaluation. The 3D evaluation successfully detected micrometastases that had been missed by the conventional 2D evaluation. (D) An example of a breast cancer case. ¹⁶ The acquired grayscale fluorescence images were post‐processed to match the color of hematoxylin and eosin (H&E)‐stained samples for visualization. Image quality was largely equivalent to conventionally acquired H&E images. (E) Multispectral large‐scale single‐cell resolution 3D (mLSR‐3D) imaging following FUnGI tissue clearing and “on‐the‐fly” linear unmixing, enabling up to eight fluorophores during a single acquisition. ²⁹ The resulting 3D volume of the low‐grade glioma successfully facilitated detailed analysis, such as nuclei counting and volume estimation. (F) DIIFCO enabled in situ RNA detection with tissue clearing. ³¹ A human breast cancer organoid was whole‐mount stained with a probe for MKI67 mRNA and the nuclear stain (YO‐PRO‐1) is shown. VALIS, Virtual Alignment of pathoLogy Image Series; mLSR‐3D, multispectral large‐scale single‐cell resolution 3D; FUnGI, fructose, urea and glycerol clearing solution for imaging; DIIFCO, for diagnosing in situ immunofluorescence‐labelled cleared oncosamples.

Tissue clearing‐enabled 3D histopathology relies on the recent development of tissue processing technologies and protocols, as well as novel 3D imaging techniques (Figure 2A). Light‐sheet fluorescence microscopy (LSFM) is one such imaging technique which enables the acquisition of a series of planar images of optically cleared tissues by utilizing a thin sheet of excitation light. ¹⁴ This configuration enables the rapid acquisition of continuous optical sections without the requirement to physically section the specimen. The continuity of the 3D voxel data allows observation of 2D sections from any arbitrary direction. While the method is restricted to the detection of fluorescence signals, it is possible to generate fluorescence images that are analogous to the appearance of a H&E stain (Figure 2B). ¹⁵ , ¹⁶ , ¹⁷ The cleared sample is also compatible with post hoc paraffin embedding and sectioning. ¹⁸ The challenge of achieving efficient penetration of fluorescent probes in 3D samples is ongoing. However, significant progress has been made in addressing this challenge through recent advancements in the understanding of the chemical principles affecting probe penetration and the establishment of practical protocols. ¹⁹ , ²⁰ , ²¹ , ²²

Tissue clearing methods have assisted several proof‐of‐concept studies of 3D histopathology applied to cancer diagnosis, demonstrating that it has a potential to yield better clinical outcomes in comparison with conventional 2D histopathologic methods (Figure 3). Independent studies from multiple groups have demonstrated that 3D imaging data of cancer patient‐derived lymph nodes can enhance the diagnostic accuracy for detecting micrometastases by 14.8% ¹⁸ or 19% (Figure 3C,D). ¹⁶ Other attempts to examine prostate cancer biopsy specimens in three dimensions have demonstrated an evolving compatibility with the 2D pathological frameworks as well as improved concordance in Gleason score among pathologists. ²³ , ²⁴ , ²⁵ The same group also reported improvement in the evaluation of the glandular morphologies of Barrett esophagus specimens. ²⁶ Besides the compatibility with conventional H&E staining, attempts to obtain 3D images with label‐free, multiplex multicolor staining or mRNA labeling have been made (Figure 3E,F), further broadening the accessible data and improving our understanding of molecular mechanisms driving cancer heterogeneity. ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ Larger cohort studies and clinical trials are still awaited to assess how 3D histopathology affects diagnostic accuracy in relation to the patients' prognosis.

Overall, the attempts to date have demonstrated that 3D histopathology approaches have the ability to overcome the limitations associated with undersampling and the assessment of a limited number of thin 2D sections. In both multisectioning and tissue clearing methods, however, the data size and processing time are dependent on the sample size and resolution. Managing large datasets necessitates corresponding computational resources, and appropriate consideration must be given according to the purpose (Figure 2C and Table S1).

4. FORWARD‐LOOKING PERSPECTIVE OF 3D HISTOPATHOLOGY

We foresee a future where 3D histopathology will play a key role in translational medicine, enhancing the diagnostic capabilities of pathology and clinical oncology (Figure 4A). 3D histopathology can be seamlessly integrated into a conventional pathology workflow, aiding its adoption in the clinic. An optimized sample processing protocol will enable the reciprocal use of cleared tissues for LSFM imaging and formalin‐fixed paraffin‐embedded tissues for current routine pathological examinations, as well as further clinical molecular and genetic testing. ¹⁸ , ³² This present framework serves as a bridge between human and machine vision, enabling the translation of histopathological observation into a mutually comprehensible data format, especially H&E staining images. Machine vision is expected to provide the additional histopathological insights when curated by human vision. Moreover, the state of human histopathology itself stands to evolve, as machine vision begins to perceive data that was hitherto invisible to the human eye.

FIGURE 4.

Future perspective of histopathology. (A) Three‐dimensional (3D) histopathology extends the present histopathological framework. Current 3D histopathology focuses primarily on supplementing conventional two‐dimensional (2D) histopathology. It also aims to utilize data formats understandable for human vision such as hematoxylin and eosin (H&E) staining ⁴⁴ and apply machine vision for digital pathology, but is practically cumbersome in the digital scanner‐reliant workflow. In this regard, 3D histopathology methods can directly provide digital images whilst preserving sample architecture and morphology. (B) A future perspective for 3D histopathology. By leveraging the inherent properties of 3D data, pathologists will be able to access optical sections of volumetric H&E images from multiple orientations, enabling more sensitive diagnostics with human vision. Simultaneously, 3D data can provide more artificial intelligence‐friendly information modalities, such as 3D cell point clouds, which enable an ideal data analysis workflow for machine vision. (C) Limitations of current 3D histopathology in practice. There are a number of issues that must be addressed before a realistic framework for clinical 3D histopathology can be implemented. IHC, immunohistochemistry; LSFM, light‐sheet fluorescence microscopy.

Future inventions of devices to implement practical 3D histopathology workflow will further accelerate its distribution to the clinical field. Various groups around the world are currently developing different types of imaging devices that are both affordable and optimized for clinical use, such as the open‐top light sheet microscope ³³ and desktop‐equipped selective plane illumination microscopy. ³⁴ Full automation of the workflow from tissue processing to image acquisition and data processing is a plausible means to decrease the technical burden and aid uptake of the technology in the clinic. From the optics and engineering perspective, 3D histopathology now faces the challenge to achieve comparable optical resolution to traditional widefield microscopes without sacrificing image acquisition speed. Achieving adequate contrast and resolution is crucial for diagnostic precision. Possible solutions include the development of optical systems with higher numerical aperture, cameras with larger sensors and smaller pixels, and machine learning‐aided image processing. ³⁵ , ³⁶

Aside from the fruitful combinations with conventional 2D pathology enhancing human vision, 3D histopathology can also foster a stronger synergy with the latest digital/artificial intelligence (AI) pathology and image informatics trends, ³⁷ , ³⁸ facilitating machine vision. The advent of AI technologies aids in the identification of subtle or global features that may be difficult to see for the human eye. In the current state of 2D histopathology, however, research laboratories frequently encounter difficulties in digitally acquiring and integrating thin sections into data formats that are compatible with computational analysis. This is primarily due to the whole‐slide scanning process, which involves using commercially and clinically employed virtual slide scanners to obtain digitized images. ³⁹ , ⁴⁰ Because of the inherent analogue nature of the 2D preparation, the digitized whole‐slide images do not possess the original information that pathologists “see” via microscopic observation: it is like observing paintings in museums directly by eye rather than digitized photographs. Conversely, 3D histopathology methods have the potential to directly provide digital images of entire samples, negating the burdensome section scanning workflow whilst obtaining data that are suitable for computational analysis. The rich information contained within 3D data can readily provide digital sections from multiple orientations, enabling highly sensitive AI algorithms to be trained with only a few (e.g., dozens) samples. ⁴¹ These digital‐native characteristics and profusion of 3D information can enhance the precision and efficacy of pathological diagnosis.

The integration of pathological diagnosis with information science and AI presents many challenges, specifically in combining the expertise of pathologists with the analytical capacity of AI algorithms. Pathologists have extensive expertise in diagnosing diseases based on histopathological images (especially H&E slide images), whereas AI algorithms can take advantage of data in different modalities. This relationship has the potential to be enhanced by 3D histopathology. 3D histopathology data offer 3D serial section images as a diagnostic modality for pathologists using sections with H&E staining or its equivalent. ¹⁵ , ¹⁶ , ¹⁷ , ²⁵ , ²⁶ , ³⁴ Simultaneously, modalities other than microscopy images, such as 3D cell point clouds, can link complex 3D spatial information of tissues to AI information science. ⁴² 3D point clouds have extensively been used in engineering in combination with information science tools like machine learning. Due to the complex 3D structures of biological tissues, the 3D point cloud data of all cell positions in a volumetric image contains a wealth of 3D spatial context information. Thus, data‐mining algorithms based on a 3D cloud of histopathological points can extract meaningful information for precise diagnostics. In addition, managing biological data as 3D point clouds can resolve current issues in AI pathology, such as increased computational capacity, storage space, and associated costs resulting from larger image datasets. ⁴³ The representation of point clouds can reduce data size by orders of magnitude. In addition, the point cloud modality mitigates the issue of the diminished generalizability of AI diagnostic tools due to variations in the color and quality of H&E‐stained slides between institutions and facilities. The provision of data modalities that can be diagnosed by both pathologists and AI will therefore facilitate the integration of current pathological operations with the most recent developments in information science.

It is important to note that current 3D histopathology has limitations compared to traditional methods (Figure 4B). The most significant issue is the trade‐off between time and cost versus image quality and quantity. Most reported protocols necessitate a week‐long procedure from specimen preparation to the final image evaluation. Some protocols have successfully obtained turnaround times comparable to intraoperative frozen sections, ²⁴ but their applicability is still limited. Although there have been several attempts at clearing and imaging large tissues and whole organs, the currently available tissue clearing methods cannot render all samples sufficiently transparent. In addition, the processing and storage of volumetric imaging data requires large computer resources due to the considerable data sizes involved. Consequently, the protocols for tissue processing, devices for image acquisition, and pipelines for image analysis will remain under development until they are practical, efficient, and rapid enough for clinical use.

As discussed, we do not anticipate that the emerging trend of 3D histopathology will supplant the role of conventional histopathology in medical practice and oncological research. The current body of knowledge concerning human anatomy, histology, and their pathological alterations is based on hundreds of years of experiences, rigorous clinical evaluation, and accumulated evidence. In addition, histopathology is one of the best‐established medical diagnostic techniques, despite the many obstacles remaining in the field. We look forward to a future where 3D histopathology expands our understanding of human pathophysiology and improves patient care through cross‐disciplinary collaboration and innovation.

The clinical application of 3D histopathology is on the horizon. It is highly anticipated that this new modality will join the medical field as an indispensable procedure, contributing to our understanding of diseases and improving patient care. However, extensive trials, technical innovation, and collaboration across multiple areas of expertise are necessary to realize this vision.

AUTHOR CONTRIBUTIONS

Akira Leon Yoshikawa: Conceptualization; resources; writing – original draft; writing – review and editing. Takaki Omura: Conceptualization; resources; writing – original draft; writing – review and editing. Atsushi Takahashi‐Kanemitsu: Data curation; supervision; writing – review and editing. Etsuo A. Susaki: Conceptualization; funding acquisition; project administration; supervision; writing – review and editing.

FUNDING INFORMATION

This study was supported by Japan agency for medical research and development (JP20gm6210027, JP22ama221517, and JP21ak0101181 to E.A.S.), Japan society for the promotion of science KAKENHI (22H02824 to E.A.S.), Grants‐in‐Aid from the Nakatani Foundation for Advancement of Measuring Technologies in Biomedical Engineering, and the Mochida Memorial Foundation for Medical and Pharmaceutical Research (to E.A.S.).

CONFLICT OF INTEREST STATEMENT

E.A.S. is a co‐inventor on patents owned by RIKEN and CUBICStars, Inc. covering the clear, unobstructed brain/body imaging cocktails and computational analysis (CUBIC) tissue clearing and CUBIC‐HistoVIsion 3D tissue staining reagents, respectively. E.A.S. is also employed by CUBICStars Inc., which offers services based on CUBIC technology. E.A.S. also received lecture fees from Miltenyi Biotec, collaboration funding from Kantum Ushikata Co., Ltd, and research grants from the Takeda Science Foundation, the Nakatani Foundation for Advancement of Measuring Technologies in Biomedical Engineering, and the Mochida Memorial Foundation for Medical and Pharmaceutical Research. Other authors have no conflict of interest.

ETHICS STATEMENT

Approval of the research protocol by an Institutional Reviewer Board: N/A.

Informed Consent: N/A.

Registry and the Registration No. of the study/trial: N/A.

Animal Studies: N/A.

Supporting information

TABLE S1

Yoshikawa AL, Omura T, Takahashi‐Kanemitsu A, Susaki EA. Blueprints from plane to space: outlook of next‐generation three‐dimensional histopathology. Cancer Sci. 2024;115:1029‐1038. doi: 10.1111/cas.16095