Visual Assessment of 2-Dimensional Levels Within 3-Dimensional Pathology Data Sets of Prostate Needle Biopsies Reveals Substantial Spatial Heterogeneity

Abstract

Prostate cancer prognostication largely relies on visual assessment of a few thinly sectioned biopsy specimens under a microscope to assign a Gleason grade group (GG). Unfortunately, the assigned GG is not always associated with a patient’s outcome in part because of the limited sampling of spatially heterogeneous tumors achieved by 2-dimensional histopathology. In this study, open-top light-sheet microscopy was used to obtain 3-dimensional pathology data sets that were assessed by 4 human readers. Intrabiopsy variability was assessed by asking readers to perform Gleason grading of 5 different levels per biopsy for a total of 20 core needle biopsies (ie, 100 total images). Intrabiopsy variability (Cohen κ) was calculated as the worst pairwise agreement in GG between individual levels within each biopsy and found to be 0.34, 0.34, 0.38, and 0.43 for the 4 pathologists. These preliminary results reveal that even within a 1-mm-diameter needle core, GG based on 2-dimensional images can vary dramatically depending on the location within a biopsy being analyzed. We believe that morphologic assessment of whole biopsies in 3 dimension has the potential to enable more reliable and consistent tumor grading.

Introduction

Prostate cancer (PCa) is the second leading cause of cancer-related death in the United States and the fifth worldwide.¹ The worldwide incidence is 1.3 million with a mortality rate of 359,000 per year.² Surgical removal of the prostate, ie, radical prostatectomy (RP) and radiation therapy are standard treatments for localized PCa. However, despite such aggressive treatment approaches, 30% to 40% of patients still experience biochemical recurrence,^3,4 defined as a rise in the blood level of prostate-specific antigen in patients with PCa. Likewise, many patients with indolent PCa may be overtreated with RP and radiation therapy, which can lead to severe complications, such as urinary incontinence or erectile dysfunction.^5,6 This suggests that an improved indicator of tumor aggressiveness is needed, allowing for better personalized treatments.

PCa risk management largely relies on the visual assessment of biopsied tissues with optical microscopy (ie, histopathology) to assign a Gleason grade group (GG). Biopsies with lower scores more closely resemble normal tissue and are thus considered to be more indolent (nonlethal disease). According to the current guidelines, a single biopsy should be sectioned and visualized by pathologists at 3 closely spaced levels (typically ~20 μm apart)⁷ that collectively represent an ~1% of the whole biopsy.⁸ These “representative” sections from a part of the biopsy may not adequately capture the morphologic heterogeneity of the tumor⁹ and can also lead to ambiguities, such as tangential 2-dimensional (2D) sections of fully formed glands (Gleason pattern 3) being misinterpreted as poorly formed glands (Gleason pattern 4).^10,11 In terms of spatial heterogeneity, Reyes and Humphrey⁹ found that “representative” sections missed clinically important atypical structures that were apparent in exhaustively serially sectioned specimens. In another study,¹² it was shown that biopsies with atypical glandular proliferation might show focal carcinoma in “additional” sections, even if immunohistochemical analysis of “representative” sections did not identify malignancy. These findings suggest that the tumor grade, as assessed from different 2D slices within a 3-dimensional (3D) tumor volume, is likely to vary between slices (intrabiopsy variability). Exhaustively and serially sectioning a whole specimen for histologic analysis may mitigate such variability but would be destructive of valuable clinical specimens and would require extensive manual effort by histotechnologists, which is impractical.

The recent advent of high-throughput open-top light-sheet (OTLS) microscopy of optically cleared tissues provides an elegant solution for obtaining serial digital 2D sections throughout a 3D biopsy in a nondestructive way. OTLS can rapidly collect images in a z-stack arrangement (ie, 3D pathology) from entire biopsies or surgical excisions without tissue sectioning.^13,14 Previous studies^10,14,15 have also shown that the tissue processing and imaging methods used in this study are gentle and reversible (ie, the specimens may be returned as formalin-fixed, paraffin-embedded blocks after 3D pathology is performed), having no discernable adverse effects on tissue morphology and molecular expression.

The main focus of this study was to assess the extent of intrabiopsy variability in Gleason grading. In collaboration with 4 human readers (X.F., T.M., P.L.F., L.D.T), this study focused on Gleason grading of individual 2D slices that were virtually sampled from 5 evenly spaced levels spanning the OTLS-generated 3D pathology data sets. Unlike several previous studies that have focused on interreader variability of GG, the focus here was to evaluate intrabiopsy variability for individual human readers in terms of GGs across 5 widely spaced levels (~100 μm apart) from each 1-mm-diameter biopsy. The human readers graded every 2D image (5 levels from each of 20 biopsies) in a randomized sequence (100 total images) using a custom-developed web platform (Fig. 1).

Figure 1.

Overview of the present study. (A) Archived formalin-fixed, paraffin-embedded prostatectomy specimens were obtained from a cohort of 20 patients, from which 20 simulated (ex vivo) biopsies were extracted for the analysis. (B) The biopsies, which all contained cancer, were labeled with a fluorescent analog of hematoxylin and eosin (H&E) staining, optically cleared to render the tissues transparent to light, and then comprehensively imaged in 3D with an OTLS microscope. (C) Five 2D images were extracted at 100-μm intervals from each biopsy, resulting in a total of 100 images from the 20 biopsies. (D) 2D images were stored on a university server to allow our collaborating human readers access to the images remotely. (E) The human readers graded all 2D images in a randomized order using an in-house–developed web platform. (F) After collecting the grades from 4 readers, intrabiopsy variability was quantified. OTLS, open-top light-sheet; 2D, 2 dimension; 3D, 3 dimension.

Materials and Methods

Patient Collection and Volumetric Image Acquisition With Open-Top Light-Sheet

The study was reviewed and approved by the Institutional Review Board of the University of Washington (study 00004980), where research specimens were previously obtained from patients with informed consent. Archived formalin-fixed, paraffin-embedded prostatectomy specimens were gathered from ~200 patients with low-risk to intermediate-risk PCa as part of a previous active surveillance study.¹⁶ The cases were initially graded during post-RP histopathology as having GGs of 1, 2, or 3, where approximately half of them were low risk (GG1). Of the ~200 cases that were imaged, 20 cases were randomly selected to be used in this study. One cancerous biopsy per case was selected for the present study. Biopsies with evidence of cancer were initially identified by 2 experienced pathologists.⁸

Each biopsy core was imaged comprehensively in 3D using a published OTLS microscope system¹⁴ after the biopsies were stained with a fluorescent analog of hematoxylin and eosin (H&E) and optically cleared with ethanol dehydration and immersion in ethyl cinnamate. As described previously,^8,14 the H&E analog consisted of a nuclear dye, TO-PRO3, and eosin, with sampling of ~0.44 μm/voxel, roughly equivalent to what is achieved with a 10× objective on a standard transillumination bright-field light microscope.¹⁰ The volumetric imaging took ~0.5 min/mm³ of tissue for each illumination wavelength, resulting in ~50 GB of raw data per biopsy. However, these data sets were downsampled by a factor of 2 in all dimensions (8× reduction in file size) for this human observer study.

Data Set Preparation

A total of 5 cross-sectional images were extracted at 100-μm intervals spanning nearly the entire diameter of each biopsy (~0.9 mm diameter), resulting in a total of 100 digital images {1 ≤ i ≤ 5, 1 ≤ b ≤ 20}, where $l_b^i$ is an image extracted from the i^th level of a biopsy identified as biopsy number b. Two-channel fluorescence images were false colored to mimic H&E staining using a previously published method.¹⁷ 2D images were stored in a secure university server, in which external human readers were provided secure access.

Human Readers Involved in This Study

Four human readers (r₁, r₂, r₃, and r₄) were involved in this study. r₁ is a genitourinary (GU) pathologist with 24 years of experience. r₂ is a GU pathologist with 33 years of experience. r₃ is a general pathologist with 8 years of experience, and r₄ is a GU pathologist with 17 years of experience. Each human reader was given a task of grading all images using a custom-developed interactive web tool.

Collecting Grades From Human Readers Using an In-House Interactive Web Tool

We developed an interactive web tool to facilitate reviewing the 2D images. The tool was developed to support large high-resolution images and provide several important functionalities necessary for analyzing high-resolution images, such as zooming and panning. An account was created for each reader to access the platform with their username and password. The tool shuffled all images in a way that no 2 images from the same biopsy would be displayed to the reader within 3 consecutive grading events. This was done to minimize the likelihood that the reader would be able to recall previous images originating from the same biopsy. The tool allowed the reader to stop/resume their analysis at any time. When they logged back in, the tool would allow for resumption of assessment from the last image reviewed.

For each image, the reader was asked to assess GG as defined by the International Society of Urological Pathology.¹⁸ This involved identifying the 2 predominant Gleason patterns (which could be identical if only 1 pattern was seen). Grades were collected from all human readers, resulting in a total of 400 predictions. All predictions and comments were saved into a structured database created with SQLAlchemy, a Python SQL toolkit for database object mapping.¹⁹ The web tool was developed in Python 3.8 and JavaScript with “Flask,”²⁰ “Open-Slide,”²¹ and “OpenSeadragon” libraries.²²

Exploring Intrabiopsy Variability in Grading Based on 2-Dimensional Histology Images

Intrabiopsy variability was analyzed quantitatively and qualitatively. It was calculated as the worst pairwise agreement in GG between individual levels. Each level was considered an independent observation. The pairwise agreement between any 2 levels in a biopsy was measured using Cohen κ.²³ We used 5 individual levels from a biopsy, and therefore, there would be 10 different pairwise comparisons: {$k_j^i$, 1 ≤ i ≤ 5, 1 ≤ j ≤ 5, i ≠j}. This resulted in 10 different Cohen κ scores for a reader. The worst pairwise κ score was then identified as the smallest one among the 10 scores.

Interpretation of Cohen κ score was based on the study by Cohen,²³ where values ≤0.20 were taken to indicate no agreement, 0.21 to 0.39 as “minimal,” 0.40 to 0.59 as “weak,” 0.60 to 0.79 as “moderate,” 0.80 to 0.90 as “strong,” and >0.90 as “almost perfect” agreement. Violin and histogram plots were used to visualize distributions of GGs. All analyses were performed in Python 3.8 with the “scikit-learn,”²⁴ “Matplotlib,”²⁵ “plotnine,”²⁶ and “Pingouin”²⁷ libraries.

Results

Distributions of Grade Groups Collected Across 4 Human Readers

The distributions of GGs assigned by all human readers are reported in Table 1. The most frequently assigned grade was GG1, which appears to align with the original study. r₁, r₂, r₃, and r₄ assigned 50%, 44%, 56%, and 59% of images as GG1, respectively. r₄ did not assign any GG for 1 image because of challenges with poor image quality. Overall, across all readers, 52% of the grades were assigned as GG1, and 18%, 14%, 9%, and 7% of the grades were assigned as GG2, GG3, GG4, and GG5, respectively. The average and SD of the GGs were 1.92 ± 1.08, 2.16 ± 1.33, 2.07 ± 1.41, and 1.85 ± 1.26 for r₁, r₂, r₃, and r₄, respectively.

Table 1.

Distributions of GGs assigned by all human readers

	r 1	r 2	r 3	r 4	Total (%)
GG1 (%)	50	44	56	59	52
GG2 (%)	20	22	12	16	18
GG3 (%)	18	19	9	11	14
GG4 (%)	12	4	15	6	9
GG5 (%)	0	11	8	7	7
Total (%)	100	100	100	99	100

GG, grade group.

Among the 20 samples, the GG determinations that fluctuated the most were between GG1 and GG2 for r₁ (n = 5, tied with the flipped grades between GG2 and GG3), r₂ (n = 5), and r₃ (n = 6), whereas the second most common flipped grades were between GG2 and GG3 for r₂ (n = 4) and r₃ (n = 3) (Table 2). For r₄, the most common flipped grades were between GG2 and GG3 (n = 5), and the second most common was between GG1 and GG2 (n = 4, tied with the flipped grades between GG4 and GG5).

Table 2.

Number of samples (of 20) where 2 GGs were coassigned to the same sample by (a) r₁, (b) r₂, (c) r₃, and (d) r₄

(a) r 1
	GG2	GG3	GG4	GG5
GG1	5	3	0	0
GG2	—	5	1	0
GG3	—	—	1	0
GG4	—	—	—	0
(b) r 3
	GG2	GG3	GG4	GG5
GG1	6	2	2	0
GG2	—	3	2	0
GG3	—	—	2	3
GG4	—	—	—	2
(c) r 2
	GG2	GG3	GG4	GG5
GG1	5	1	0	1
GG2	—	4	1	2
GG3	—	—	3	3
GG4	—	—	—	3
(d) r 4
	GG2	GG3	GG4	GG5
GG1	4	2	1	0
GG2	—	5	1	1
GG3	—	—	1	2
GG4	—	—	—	4

GG, grade group.

Experiment: Pairwise Agreement in Grade Groups When Grading Different Levels Within a Biopsy

In this experiment, the agreement in GG for all combinations of 2 levels within the same biopsy was measured for each reader. The worst pairwise agreement between levels was calculated using the Cohen κ method to quantify intrabiopsy agreement and found to be κ = 0.43 (“weak agreement”) for r₁, 0.38 (“minimal agreement”) for r₂, 0.34 (“minimal agreement”) for r₃, and 0.34 (“minimal agreement”) for r₄ (Fig. 2). All readers assigned at least 2 different grades for 50% of the biopsies. Overall, the pairwise agreement between levels within a biopsy tended to decrease (worsen) as the distance between the levels increased, implying that the grade is spatially heterogeneous within the tumor.

Figure 2.

Intrabiopsy agreement. (A) An example of 3-dimensional volume illustrating the 2-dimensional images extracted at 100-μm intervals through the volume of a biopsy. Pairwise agreements in grade groups (Cohen κ) at different biopsy levels for different readers (B) r₁, (C) r₂, (D) r₃, and (E) r₄. As the distance between levels increases, the agreement tends to decrease.

Figure 3 illustrates an example showing variability in tumor morphology across 2 different levels within the same biopsy and its impact on grading. The zoomed-in region in Figure 3B was extracted from the first level of biopsy 19, $l_{19}^1$, with small, closely packed tumor glands (pattern 3 or 4). However, in the same zoomed-in region at the third level of the biopsy, $l_{19}^3$, which was 200 μm away in depth (Fig. 3C), the Gleason pattern 3 glands disappear, and benign glands are observed instead. Several Gleason pattern 3 glands are still observed in $l_{19}^3$ (outside the zoomed-in region). Overall, according to all readers, $l_{19}^1$ was assigned GG3, whereas $l_{19}^3$ was assigned GG1.

Figure 3.

Distributions of GGs assigned by the human readers. (A) Scatter plots of the GGs assigned by each reader for 10 of 20 biopsies. To emphasize the range of variations, we specifically chose those 10, with different GGs for visualization purposes. Each subplot shows the grades assigned by 1 reader. Each data marker represents a GG assigned by looking at a 2-dimensional image at a particular spatial level within a biopsy. Each level is represented by a specific marker shape (see legend below the plots). (B, C) Two example images extracted from different levels within the same biopsy, l¹₁₉ and l³₁₉. (B) The zoomed region of l¹₁₉ contains mostly poorly formed glands, the presence of which indicates higher GGs, whereas (C) in the same axial region of l³₁₉, 200 μm far away from l¹₁₉ in depth, the glands appear well formed and have recognizable patterns, with clear boundaries, mostly associated with lower GGs. GG, grade group; ISUP, International Society of Urological Pathology.

Discussion

This study focused on evaluating the extent of variability in tumor grading when human readers were given the task of grading individual 2D sections obtained from a 3D tumor volume. Using a global collaboration involving 4 pathologists, we attempted to ascertain intrabiopsy variability for each individual human reader in terms of Gleason grading across multiple levels spanning a 3D biopsy data set. Unlike previous studies^28–32 that have focused on the quantitative assessment of the extent of interreader variability in GG, this study focused on intrabiopsy variability because of tumor spatial heterogeneities.

Our experiment revealed low intrabiopsy agreement on GG grading. We found that for the majority of biopsies, readers assigned different GGs to different levels within individual biopsies. Moreover, the degree of intrabiopsy variability found in this study is significant in relation to other sources of “uncertainty” in Gleason grading, such as intrareader variability.^{29–31,33,34} For instance, r₁ regraded a subset of the samples (n = 25) after a washout period of ~10 months. Intraobserver agreement for the 25 samples for r₁ was 0.57, which was much higher than the intrabiopsy agreement for the same reader (κ = 0.43). Similarly, Melia et al³¹ reported a κ score κ of 0.66 when an observer graded the same slides at different time points. In this study, significantly lower κ values for intrabiopsy variability were seen, suggesting that this variability is due at least in part to intrinsic morphologic differences within each biopsy rather than variations in interpretation by a single reader over time.

To better understand the implications of histology sectioning in clinical decision making, Reyes and Humphrey⁹ found that some of the diagnostic needle biopsies tended to be undergraded because of incomplete sectioning. They exhaustively serially sectioned specimens and defined the following 2 sets of slice groups: “diagnostic” slices, which represented slices used in clinical practice for diagnostic purposes, and “residual” slices, which represented the slices other than the diagnostic ones. It was observed that 4 needle biopsies, which were identified as focal glandular atypia in their diagnostic slices, were identified definitively as carcinoma in the residual slices. The clinical implications of such findings are profound as diagnosis of carcinoma likely prompts RP, whereas diagnosis of atypia typically results in only clinical follow-up. Similarly, for all pathologist readers in our study, the first level of biopsy 19 was evaluated GG3, falling into the intermediate-risk group (Fig. 3A) typically requiring RP, whereas the third level of the same biopsy (200 μm away) was graded GG1 by the same readers, a low-risk group (Fig. 3B) typically assigned to active surveillance. Such discrepancies observed in different 2D images of the same tumor support the value of 3D analysis. Looking at the entirety of a 3D tumor volume to derive a composite 3D-based GG, instead of looking at individual 2D slices, may yield better predictions of disease outcome.

As mentioned previously, the main finding of this study is that there are significant morphologic variations within prostate needle biopsies as a function of spatial location within the biopsy volume. This level of variability is significantly greater than what can be attributed to the uncertainty in grade based on the visual assessment by pathologists (intraobserver variability). Although the focus of the study was not on interobserver variability between pathologists, κ metrics revealed weak agreement between human readers (κ = 0.46). This finding appears to align with previous studies,^32,35 which have found that although agreement between pathologists improves as their experience increases, significant variability between readers remains.^31,36,37

A limitation of this study is that the intact prostate biopsies were imaged using a lower-resolution prototype of OTLS microscopy, equivalent to what is achieved with a ~10× objective on a standard transillumination bright-field light microscope.¹⁰ Thus, pseudo-OTLS images lack the higher spatial resolution to see the nucleoli and make the diagnoses of cancer difficult for GG5 cancerous cells. These cells can potentially be confused with inflammatory cells, small cell carcinoma, or other cell types. Additionally, differential diagnosis between cancerous and noncancerous lesions is sometimes dependent on the assessment of nuclear morphology, such as adenosis versus well-differentiated prostate adenocarcinoma, hyperplasia versus well-differentiated prostate adenocarcinoma, and atypical small acinar proliferation lesions. This situation may contribute to some degree of uncertainty by the pathologists. However, all biopsies were preselected to contain low-risk to intermediate-risk PCa (GG1-GG3) and Gleason grading is based on gland morphology rather than high-resolution cytologic features, and therefore, the lower resolution of our data sets should have had a minimal effect on the grading process. Additionally, although pseudo-H&E images have the advantage of being familiar to practicing pathologists, there may be slight deviations from conventional H&E images.¹⁷ These factors may also contribute to some degree of uncertainty by the pathologists. Higher-resolution data sets will be obtained in the future (40× equivalent), leveraging the more recent OTLS systems.^38,39

Since Gleason grading evolved many decades ago from the visual assessment of 2D histology slides, there is a potential need for a more accurate grading system based on 3D pathology that better correlates with clinical outcome. Additionally, the integration of computational pathology and machine learning within 3D pathology could pave the way for a new generation of spatial biomarkers to predict disease outcome or treatment response more accurately and efficiently.^8,11 Our previous study⁴⁰ provided evidence along these lines, suggesting that 3D histomorphometric analysis was superior to analogous 2D analyses for determining PCa aggressiveness.

In conclusion, this preliminary study presents evidence of a high degree of intrabiopsy variability for a 2D-based tumor grading system. Our findings reveal that Gleason grading based on 2D images varies dramatically as a function of spatial position within a 3D biopsy volume that is being analyzed. Consequently, clinical decision making and patient management may be affected. We believe that morphologic assessments of whole biopsy specimens in 3D can enable more reliable and consistent tumor grading than the standard of care that is based on the limited numbers of 2D histology sections.

Data Availability

The data sets used during the current study are available from the corresponding author on reasonable request. The original 3-dimensional pathology data sets used in this study are available publicly through The Cancer Imaging Archive—https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=145754446