Three-dimensional histopathological reconstruction of bladder tumours

Histopathological analysis is the cornerstone in bladder cancer (BCa) diagnosis. The histological grade and stage of BCa are important prognostic parameters for therapeutic decision-making [1, 2]. Patients with muscle-invasive bladder cancer have a higher risk of cancer-specific mortality compared to patients without muscle invasion [2]. Initial treatment is generally performed by transurethral resection of the bladder tumour (TURB), allowing tissue removal regardless the location or size. Histopathologic analysis of TURB specimens is limited since it is difficult to determine the degree of invasion [3]. Important landmarks for staging, such as the lamina propria or the muscularis propria (MP), are often difficult to recognize in the small TURB fragments. Moreover, these fragments are often subject to thermal artifacts or tangential sectioning [4]. Both factors result in a poor interobserver agreement of BCa staging [5]. Furthermore, BCa is often understaged at the initial resection because of the absence of MP in the resection specimen [6]. As a result, upstaging to a muscle-invasive tumour after a second TURB occurs in 10–20% [2] of BCa. To improve staging, en-bloc resections, in which the bladder tumour is resected as a whole are increasingly performed [3]. Previous studies including en-bloc resections have reported to contain MP in 96–100% of cases [3, 7, 8]. In en-bloc resections, a better impression of the spatial arrangement of the tissue can be obtained, assisting the pathologists with the orientation [3]. In general, pathologists examine an average number of 1–2 slides per tissue block, depending on the type of tissue [9]. These sections are only a small fraction of a the three-dimensional (3D) resected specimen. As only a fraction of heterogeneous tissue is examined, undersampling of the BCa specimen is likely to occur. To enable a pathologist to study a complete tissue sample comprehensively, it is necessary to cut and stain multiple consecutive tissue slides. However, it is impractical and time consuming to visualize and analyze a large number of tissue slides per sample by conventional microscopy.

Whole slide image (WSI) systems provide pathologists histopathologic glass slides as digital high resolution images with magnifications similar to conventional microscopy. WSI offers the possibility to create a digital 3D reconstruction based on stacks of two-dimensional (2D) WSIs. 3D reconstructed histology images hold the potential to assist pathologists in visualizing the 3D spatial arrangements of structures. One of the major problems encountered in 3D image reconstruction of histology slides is tissue deformation due to fixation, cutting, and mounting of the specimen during sample preparation [10‐14]. Several methods have been proposed to align multiple 2D WSIs to create a 3D image [10‐16]. To our knowledge, the 3D digital reconstruction of BCa tumours from histological slides has not been described before.

In this study, we present a method of preparing, reconstructing and visualizing 3D images reconstructed from multiple consecutive 2D WSIs of BCa en-bloc specimens. Moreover, we introduce a technology that enables interactive manipulation of 3D histological volumes. Additionally, we describe a method to visualize the tumour and MP, without the submucosal layer, allowing to review the spatial relationship of these structures.

We have demonstrated the work-up of high quality 3D reconstructions of en-bloc TURB specimen from sequential 2D WSIs. With the provided segmentations of BCa tumour and MP, the spatial relationship between these structures could be visualized in 3D. By providing better insight in the growth pattern of a tumour, this method has the potential to assist the pathologist in accurate staging of BCa.

To our knowledge, this is the first study focusing on 3D reconstructing BCa tumours out of multiple 2D histological slides. By evaluating a 3D volume instead of the conventional 2D images, we assume that more information can be visualized. We created segmentations of the tumour and the MP and faded out the area in between, to emphasize the distance between these two layers and clarify the changes that occur throughout the different slides. Booth et al. used 3D segmentations to visualize the mammarian lobule in relation to ductal carcinoma in situ and noticed a complexity that was of higher order than was apparent from the 2D slides [13]. A small number of previous studies have created 3D reconstructions out of histological slides, using multiple sections [10‐15]. Booth et al. discussed the depth of the stack for the 3D reconstruction of breast cancer and found 30 slides, with a thickness of 4 μm, to be an insufficient number for a spatial reconstruction [13]. In future, we aim to increase our reconstructions to determine the optimal number of sections to create a 3D volume. Moreover, we would like to investigate if it is necessary to use consecutive slides. Due to the anisotropic voxel (volumetric pixel) size of the 3D image, important information could be missed.

We have shown good qualitative results in the registration of multiple WSIs. In previous studies that focused on the 3D reconstruction of histological slides, various registration methods have been used [10‐15]. In our study, we have only used existing open-source tools, being easily accessible. Registration of the WSIs is a challenging task since the volume of the slides within a specimen can differ enormously due to deformation of the tissue. Deformation occurs as a result of the fixation with formalin, cutting and mounting of the specimen. In our protocol, we tried to minimize deformation by adjusting the protocol and keeping the slides as horizontal as possible. Some groups used artificial fiducial markers as a reference point for the registration [14, 22]. These fiducial markers are placed alongside the specimen during embedding. However, during cutting and mounting of the specimen, non-linear deformation occurs throughout the whole specimen, and therefore external fiducial markers cannot predict this deformation for the whole tissue [15, 22]. Other groups have used structures that are visible through several planes as a reference point [10, 13]. However, especially in tumourous tissue, it is difficult to use this method for the basis of alignment because of the heterogeneity of the tumour. Another obstacle for the visualization is the staining inconsistency, which can even be noted within sections, several groups have tried to make up for the staining differences by colour normalization [23].

In our study, we only mentioned a qualitative measure for the accuracy of registration. Quantitative measures for registration are difficult to determine. Some groups have used Hausdorff evaluation based on manual delineations and concluded only minor registration errors [12, 14]. However, in our study we did not use Hausdorff evaluation due to the lack of anatomic markers in the tissue, in specific within the tumour. Thereby are these anatomic markers mostly based on manual segmentation of the marker, thereby introducing a dependency.

During the sample workup the major reason of slide loss were exportation problems from the Philips server due to the size of the images. Therefore, local storage solutions as well as network speed and infrastructures must be adapted. The FDA recently gave approval for the use of digital slides scanners in diagnostics [24]. We expect that this recent development may give a boost to the use and therefore the evolution of digital slide scanners and subsequently data storage more reliable.

A drawback of the presented methodology is the time consuming process of tissue preparation, as also reported in other studies [12, 13, 25, 26]. Sectioning and mounting of the tissue requires human labour. Onozato et al. have showed a method for automated sectioning with positive results [16, 26]. As for now, this is not common practice in pathology laboratories, since these machines still take more time than manual sectioning [26]. Furthermore, the manually delineation of the tissue structures was a time consuming process. Eventually, automatic tissue segmentation can be implemented [10]. Furthermore, we have used only a small portion of the tumors to create a 3D reconstruction since were not allowed to section entire tumor blocks. In future work we would like to use tumor specimen that we can section completely.

3D pathology has the potential to improve insight in the growth pattern of BCa and thus refine the staging. Although pathologist are trained in visualizing the spatial arrangement of a tumour based on the 2D representation of a single slide, 3D reconstructions could support when orientation of specimen is difficult to determine. Moreover, 3D reconstructions could give better insight in the suggested sub-staging of non-muscle invasive T1 BCa. In T1 BCa the 5-year progression rates vary between 21 and 50% [1]. The two most studied systems, stratification of the muscularis mucosa invasion (T1a versus T1b) and assessing the amount of invasion into the lamina propria (T1e and m), show an increased prognostic value compared to normal T1 staging [27]. The main objection of not implementing the T1a/b sub-staging system in clinical practice is the discontinuous nature of the muscularis mucosae and interobserver variability [28]. The T1e/m staging system however, proved to be a more consistent method [29, 30]. We believe by implementing the 3D reconstructions of en-bloc bladder resection specimen into clinical practice, the visualization of the possible tumour ingrowth can be simplified. In the future, we will focus on the clinical relevance of creating 3D reconstructions for routine diagnosis. Eventually, we need to assess the added value of 3D reconstructions, together with optimizing visualization techniques. Moreover, this could be of use in other fields, such as follicular thyroid neoplasms, to determine invasion depth.

In conclusion, we have created a 3D reconstruction of a BCa tumour out of multiple 2D WSIs. 3D reconstructions show added value of tumour configurations and may improve insight into the architectural changes and refine the classification of BCa by providing detailed spatial and structural information to the assessment by light microscopy.