Prostate Cancer Risk Stratification via Nondestructive 3D Pathology with Deep Learning–Assisted Gland Analysis

Prostate cancer is the most common cancer in men and the second leading cause of cancer-related death for men in the United States (1). Currently, prostate cancer management is largely dependent upon examination of prostate biopsies via two-dimensional (2D) histopathology (2), in which a set of core-needle biopsies is formalin-fixed and paraffin-embedded (FFPE) to allow thin sections to be cut, mounted on glass slides, and stained for microscopic analysis. To quantify the aggressiveness of the cancer, the Gleason grading system is used, which relies entirely upon visual interpretation of prostate gland morphology as seen on a few histology slides (thin 2D tissue sections) that only “sample” approximately 1% of the whole biopsy. Gleason grading of prostate cancer is associated with high levels of interobserver variability (3, 4) and is only moderately correlated with outcomes, especially for patients with intermediate-grade prostate cancer (5). This contributes to the undertreatment of patients with aggressive cancer (e.g., with active surveillance; ref. 6), leading to preventable metastasis and death (7), and the overtreatment of patients with indolent cancer (e.g., with surgery or radiotherapy; ref. 8), which can lead to serious side effects, such as incontinence and impotence (9).

Motivated by recent technological advances in optical clearing to render tissue specimens transparent to light [i.e., iDISCO (10), CUBIC (11) etc.] in conjunction with high-throughput three-dimensional (3D) light-sheet microscopy, a number of groups have been exploring the value of nondestructive 3D pathology of clinical specimens for diagnostic pathology (12–17). Compared with conventional slide–based histology, nondestructive 3D pathology can achieve vastly greater sampling of large specimens along with volumetric visualization and quantification of diagnostically significant microstructures, all while maintaining intact specimens for downstream molecular assays (18). Like others before us (12, 19, 20), we hypothesized that 3D versus 2D pathology datasets could allow for improved characterization of the convoluted glandular structures that pathologists currently rely on for prostate cancer risk stratification. However, since the associated information content of a 3D pathology dataset of a biopsy is >100× larger than a 2D whole-slide image representation (in terms of total number of pixels), computational tools are necessary to analyze these large datasets efficiently and reproducibly for diagnostic and prognostic determinations.

For computational analysis of 3D pathology datasets, a multistage pipeline was chosen for classifying patient outcomes based on interpretable “hand-crafted” features (i.e., glandular features; refs. 21–23) rather than an end-to-end deep learning (DL) strategy for risk classification based directly on the imaging data (24–26). This was motivated by: (i) the attractiveness of an intuitive feature-based approach as an initial strategy to facilitate hypothesis testing and clinical adoption of an emerging modality in which datasets are currently limited and poorly understood (18, 27), and (ii) the observation that when case numbers are limited, a hand-crafted feature-based approach can be more reliable than an end-to-end DL classifier (28, 29).

Multistage feature-based classification approaches rely on the accurate segmentation of morphologic structures such as nuclei (30–32), collagen fibers (33, 34), vessels (14, 35), or in our case, prostate glands (21, 36). This is typically achieved in one of two ways: (i) direct DL-based segmentation methods (37–40) that require manually annotated training datasets, which are especially tedious and difficult to obtain in 3D (Fig. 1A; ref. 41); or (ii) traditional computer vision (CV) approaches based on intensity and morphology, provided that tissue structures of interest can be stained/labeled with high specificity (Fig. 1B; refs. 19, 42, 43). While immunolabeling can confer a high degree of specificity for traditional CV-based segmentation, it is not an attractive strategy for clinical 3D pathology assays due to the high cost of antibodies to stain large tissue volumes, and the slow diffusion times of antibodies in thick tissues (up to several weeks; refs. 10, 44).

To address these challenges, we developed a generalizable annotation-free 3D segmentation method, hereafter referred to as “image-translation-assisted segmentation in 3D (ITAS3D).” In our specific implementation of ITAS3D (Fig. 1C), 3D hematoxylin and eosin (H&E) analogue images of prostate tissues are synthetically converted in appearance to mimic 3D immunofluorescence (IF) images of cytokeratin 8 (CK8) – a low molecular weight keratin expressed by the luminal epithelial cells of all prostate glands – thereby facilitating the objective (biomarker-based) segmentation of the glandular epithelium and lumen spaces using traditional CV tools. The DL image translation algorithm is trained with a generative adversarial network (GAN), which has been previously used for 2D virtual staining applications (45–47). However, unlike those prior 2D image translation efforts, we developed a “2.5D” virtual-staining approach based on a specialized GAN that was originally designed to achieve video translation with high spatial continuity between frames, but which has been adapted within our ITAS3D framework to ensure high spatial continuity as a function of depth (see Results; ref. 48).

As a preliminary study to investigate the value of a computational 3D pathology workflow versus a computational 2D pathology workflow, 300 ex vivo biopsies were extracted from archived radical prostatectomy (RP) specimens obtained from 50 patients who underwent surgery over a decade ago. We stained the biopsies with an inexpensive small-molecule (i.e., rapidly diffusing) fluorescent analogue of H&E, optically cleared the biopsies with a simple dehydration and solvent–immersion protocol to render them transparent to light, and then used an open-top light-sheet (OTLS) microscopy platform to obtain whole-biopsy 3D pathology datasets. The prostate glandular network was then segmented using ITAS3D, from which 3D glandular features (i.e., histomorphometric parameters) and corresponding 2D features were extracted from the 118 biopsies that contained prostate cancer. These 3D and 2D features were evaluated for their ability to stratify patients based on clinical biochemical recurrence (BCR) outcomes, which serve as a proxy endpoint for aggressive versus indolent prostate cancer.

Key methodologic information is provided below for our preliminary clinical demonstration. The Supplementary Methods contain technical details on the training and validation of ITAS3D, as well as computational details for the clinical study, such as a description of gland features.

The following study was approved by the institutional review board (IRB) of the University of Washington (Seattle, WA; Study 00004980), where research specimens were previously obtained from patients with informed consent. Archived FFPE prostatectomy specimens were collected from 50 patients with prostate cancer (see Supplementary Table S1 for clinical data), of which, 46 cases were initially graded during post-RP histopathology as having Gleason scores of 3+3, 3+4 or 4+3 (Grade Group 1–3). All patients were followed up for at least 5 years post-RP as part of a prior study (Canary TMA; ref. 49). FFPE tissue blocks were identified from each case corresponding to the six regions of the prostate targeted by urologists when performing standard sextant and 12-core (2 cores per sextant region) biopsy procedures. The identified FFPE blocks were first deparaffinized by heating them at 75°C for 1 hour until the outer paraffin wax was melted, and then placing them in 65°C xylene for 48 hours. Next, one simulated core-needle biopsy (∼1-mm in width) was cut from each of the six deparaffinized blocks (per patient case), resulting in a total of n = 300 biopsy cores. All simulated biopsies were then fluorescently labeled with the To-PRO-3 and eosin (T&E) version of our H&E analogue staining protocol (Supplementary Fig. S1).

Biopsies were first washed in 100% ethanol twice for 1 h each to remove any excess xylene, then treated in 70% ethanol for 1 hour to partially rehydrate the biopsies. Each biopsy was then placed in an individual 0.5 mL Eppendorf tube (catalog no. 14–282–300, Thermo Fisher Scientific), stained for 48 hours in 70% ethanol at pH 4 with a 1:200 dilution of Eosin-Y (catalog no. 3801615, Leica Biosystems) and a 1:500 dilution of To-PRO-3 Iodide (catalog no. T3605, Thermo Fisher Scientific) at room temperature with gentle agitation. The biopsies were then dehydrated twice in 100% ethanol for 2 hours. Finally, the biopsies were optically cleared (n = 1.56) by placing them in ethyl cinnamate (catalog no. 112372, Sigma-Aldrich) for 8 hours before imaging them with open-top light-sheet (OTLS) microscopy.

We utilized a previously developed OTLS microscope (15) to image tissues slices (for training data) and simulated biopsies (for the clinical study). For this study, ethyl cinnamate (n = 1.56) was used as the immersion medium, and a custom-machined HIVEX plate (n = 1.55) was used as a multibiopsy sample holder (12 biopsies per holder). Multichannel illumination was provided by a four-channel digitally controlled laser package (Skyra, Cobolt Lasers). Tissues were imaged at near-Nyquist sampling of approximately 0.44 μm/pixel. The volumetric imaging time was approximately 0.5 minute per mm³ of tissue for each wavelength channel. This allowed each biopsy (∼1 × 1 × 20 mm), stained with two fluorophores (T&E), to be imaged in approximately 20 minutes.

Patient-level glandular features were obtained by averaging the biopsy-level features from all cancer-containing biopsies from a single patient. Patients who experienced BCR within 5 years post-RP are denoted as the “BCR” group, and all other patients are denoted as “non-BCR.” BCR was defined here as a rise in serum levels of prostate-specific antigen (PSA) to 0.2 ng/mL after 8 weeks post-RP (49). The box plots indicate median values along with interquartile ranges (25%–75% of the distribution). The whiskers extend to the furthest data points excluding outliers defined as points beyond 1.5× the interquartile range. The P values for the BCR group versus non-BCR group are calculated using the two-sided Mann–Whitney U test (50). To assess the ability of different 3D and 2D glandular features to distinguish between BCR versus non-BCR groups, we applied ROC curve analysis, from which an area-under-the-curve (AUC) value could be extracted. The t-SNE (51) analyses were performed with 1000 iterations at a learning rate of 100.

To develop multiparameter classifiers to stratify patients based on 5-year BCR outcomes, a least absolute shrinkage and selection operator (LASSO) logistic regression model was developed (22) using the binary 5-year BCR category as the outcome endpoint. LASSO is a regression model that includes a L1 regularization term to avoid overfitting and to identify a subset of features that are most predictive. Here, the optimal LASSO tuning parameter, λ, was determined with 3-fold cross validation (CV), where the dataset was randomly partitioned into three equal-sized groups: two groups to train the model with a specific λ, and one group to test the performance of the model. Along the LASSO regularization path, the λ with the highest R² (coefficient of determination) was defined as the optimal λ. Because of the lack of an external validation set, a nested CV schema was used to evaluate the performance of the multivariable models without any bias and data leakage between parameter estimation and validation steps (52). The aforementioned CV used for hyperparameter tuning was performed in each iteration of the outer CV. LASSO regression was applied on the training set of the outer CV once an optimal λ was identified in the inner CV. AUC values were then calculated from the testing group of the outer CV (Supplementary Fig. S2). This nested CV was performed 200 times to determine an AUC (average and SD). The exact same pipeline was used to develop multiparameter classifiers based on 3D and 2D features.

Kaplan–Meier (KM) analysis was carried out to compare BCR-free survival rates for high-risk versus low-risk groups of patients. This analysis utilized a subset of 34 cases, for which time-to-recurrence data is available (see Supplementary Table S1). The performance of the models, either based on 2D or 3D features, was quantified with P values (by log-rank test), HRs, and concordance index (C-index) metrics. For the multiparameter classification model used for KM analysis, the outer CV (3-fold) in our nested CV schema was replaced by a leave-one-out approach, where one case was left out of each iteration (50 total iterations) to calculate the probability of 5-year BCR for that patient (53). The samples were categorized as low- or high-risk by setting a posterior class probability threshold of 0.5. MATLAB was used for the KM analysis and all other statistical analysis was performed in Python with the “Scipy” and “Scikit-learn” packages.

Relevant clinical data for this study are provided in Supplementary Table S1 and Supplementary Table S5. Example prostate images for testing ITAS3D codes and models are available in a GitHub repository at https://github.com/WeisiX/ITAS3D. Full 3D prostate imaging datasets, “simulated 2D whole-slide images” extracted from those datasets (three levels per biopsy), as well as other clinical data are available upon reasonable request and with the establishment of a material-transfer or data-transfer agreement.

The Python code for the deep-learning models, and for 3D glandular segmentations based on synthetic-CK8 datasets, are available on GitHub at https://github.com/WeisiX/ITAS3D.

To segment the 3D glandular network within prostate biopsies, we first trained a GAN-based image-sequence translation model (see Supplementary Methods) to convert 3D H&E analogue images into synthetic CK8 IF images, which can be false colored to resemble chromogenic IHC (Fig. 2A). As mentioned, the CK8 biomarker is expressed by the luminal epithelial cells of all prostate glands. The image translation model is trained in a supervised manner with images from prostate tissues that are fluorescently tri-labeled with our H&E analogue and a CK8-targeted mAb (Supplementary Table S2).

As shown in step 1 of Fig. 2B, for whole-biopsy H&E-to-CK8 conversion, we first sub-divide the 3D biopsy (∼ 1 mm × 0.7 mm × 20 mm) datasets into ∼1 mm × 0.7 mm × 1 mm (∼ 1024 × 700 × 1024 pixel) blocks. Each 3D image block is treated as a 2D image sequence as a function of depth. At each depth level, a synthetic CK8 image is inferred from the H&E analogue image at that level while simultaneously utilizing the images (H&E analogue and CK8) from two previous levels to enforce spatial continuity as a function of depth. This “2.5D” image translation method is based on a previously reported “vid2vid” method for video translation (time sequences rather than depth sequences; ref. 48; see Supplementary Fig. S3; Supplementary Methods). However, our modified model omits the “coarse-to-fine” training strategy implemented in the original vid2vid method because this enables training times to be minimized with negligible performance loss (see Supplementary Note S1; Supplementary Fig. S4; Supplementary Video S1). Once the synthetic CK8 image blocks are generated, they are mosaicked to generate a whole-biopsy CK8 IHC dataset for gland segmentation. In step 2 of Fig. 2B, the synthetic CK8 dataset is used to segment the luminal epithelial cell layer (“Epithelium” in Fig. 2B) via a thresholding algorithm. The gland-lumen space, which is enclosed by the epithelium layer, can then be segmented by utilizing both the epithelium segmentation mask and the cytoplasmic channel (eosin-analogue images). Algorithmic details are provided in the Supplementary Methods and Supplementary Fig. S5.

Example 3D prostate gland-segmentation results are shown for benign and cancerous regions in Fig. 3A. While the glands can be delineated on the H&E analogue images by a trained observer, automated computational segmentation of the glands remains challenging (54, 55). Here we demonstrate that 3D image translation based on H&E analogue inputs results in synthetic CK8 outputs in which the luminal epithelial cells are labeled with high contrast and spatial precision. We further show that these synthetic-CK8 datasets allow for relatively straightforward segmentation of the gland epithelium, lumen, and surrounding stromal tissue compartments (Fig. 3A). Glands from various prostate cancer subtypes are successfully segmented as shown in Fig. 3A, including two glandular patterns that are typically associated with low and intermediate risk, respectively: small discrete well-formed glands (Gleason 3) and cribriform glands consisting of epithelial cells interrupted by multiple punched-out lumina (Gleason 4). Supplementary Video S2 shows depth sequences of an H&E analogue dataset, a synthetic-CK8 dataset, and a segmentation mask of the two volumetric regions shown in Fig. 3A. A whole-biopsy 3D segmentation is also depicted in Supplementary Video S3.

To demonstrate improved depth-wise continuity with our 2.5D image-translation strategy versus a similar 2D image-translation method (based on the “pix2pix” GAN), vertical cross-sectional views of a synthetic-CK8 dataset are shown in Fig. 3B. While obvious distortions and discontinuities are seen as a function of depth with 2D image translation, the results of our 2.5D image-sequence translation exhibit optimal continuity with depth. To further illustrate this improved continuity with depth, Supplementary Video S4 shows a depth sequence of en face images (z stack). Abrupt morphologic discontinuities between levels are again obvious with 2D translation but absent with the 2.5D translation approach. To quantify the performance of our image-translation method, a 3D structural similarity (SSIM) metric was calculated in which real CK8 IF datasets were used as ground truth. For images generated with 2.5D versus 2D image translation, the 3D SSIM (averaged over 58 test volumes that were 0.2-mm³ each) was 0.41 vs. 0.37, reflecting a 12% improvement at a P value of 7.8 × 10^–6 (two-sided paired t test). This enhanced image-translation performance facilitates accurate 3D gland segmentations in subsequent steps of our computational pipeline.

To assess segmentation performance, ground-truth gland-segmentation datasets were first generated under the guidance of board-certified genitourinary pathologists (L.D. True and N.P. Reder). A total of 10 tissue volumes from different patients (512 × 512 × 100 pixels each, representing 0.2-mm³ of tissue) were manually annotated. We then compared the accuracy of ITAS3D with that of two common methods: 3D watershed (as a 3D non-DL benchmark; ref. 56) and 2D U-Net (as a 2D DL benchmark; ref. 57). ITAS3D outperforms the two benchmark methods in terms of Dice coefficient (58) and 3D Hausdorff distance (Fig. 3C; ref. 59). As a visual comparison between ITAS3D and the two benchmark methods, Supplementary Video S5 displays image-stack sequences from three orthogonal perspectives of a representative segmented dataset, where the higher segmentation accuracy of ITAS3D can be appreciated. Note that a 3D DL-based benchmark method is not provided since there are currently insufficient 3D-annotated prostate gland datasets to train an end-to-end 3D DL segmentation model (40, 60, 61); again, this is one of the main motivations for developing the annotation-free ITAS3D method.

Due to the slow rate of progression for most prostate cancer cases, an initial clinical study to assess the prognostic value of 3D versus 2D glandular features was performed with archived prostatectomy specimens. Our study consisted of N = 50 patients with prostate cancer who were followed up for a minimum of 5 years post-RP as part of the Canary TMA case-cohort study (primarily low- to intermediate-risk patients; ref. 49). The Canary TMA study was based on a well-curated cohort of patients with prostate cancer in which the primary study endpoints were 5-year BCR outcomes and time to recurrence, which are also used as endpoints for our validation study. In the original Canary TMA study, approximately half of the patients experienced BCR within 5 years of RP, making it an ideal cohort for our study. We randomly selected a subset of 25 cases that had BCR within 5 years of RP (“BCR” group), and 25 cases that did not have BCR within 5 years of RP (“non-BCR” group).

FFPE tissue blocks were identified from each case corresponding to the six regions of the prostate targeted by urologists when performing standard sextant and 12-core (2 cores per sextant region) biopsy procedures (Fig. 4A). Next, a simulated core-needle biopsy was extracted from each of the six FFPE tissue blocks for each patient (n = 300 total biopsy cores). The biopsies were deparaffinized, labeled with a fluorescent analogue of H&E, optically cleared, and imaged nondestructively with a recently developed OTLS microscope (see Materials and Methods; ref. 15). Note that we have previously shown that there is no difference in the quality/appearance of our H&E-analogue 3D pathology datasets whether the tissue is formalin-fixed only versus FFPE (where de-paraffinization is necessary; refs. 15–17). Review of the 3D pathology datasets by pathologists (L.D. True and N.P. Reder) revealed that 118 out of the 300 biopsy cores contained cancer (1–5 biopsies per case). The ITAS3D pipeline was applied to all cancer-containing biopsies (see Supplementary Methods and Supplementary Fig. S6 for cancer-region annotation). We then calculated histomorphometric features from the 3D gland segmentations, and from individual 2D levels from the center region of the biopsy cores, which were then analyzed in terms of their association with BCR outcomes. For 2D analysis, average values from a total of 3 levels were calculated, in which the three levels were separated by 20 μm (mimicking clinical practice at many institutions) as shown in Fig. 4A.

We compared multiple 3D and 2D glandular histomorphometric features (see Supplementary Table S3 for a detailed list). For example, the curvature of the boundary between the lumen and epithelium is a feature that increases as glands become smaller or more irregular, as is often seen with aggressive prostate cancer (21). This can be quantified in the form of the average surface curvature of the object in 3D, or the curvature of the object's cross-sectional circumference in 2D (Fig. 4B). As another example (Fig. 4C), the gland-to-convex-hull ratio (G/H) is defined as the volume ratio (in 3D) or the area ratio (in 2D) of the gland mask (epithelium + lumen) divided by the convex hull that circumscribes the gland. This G/H feature is inversely related to the irregularity or “waviness” of the periphery of the gland (at the scale of the gland itself rather than fine surface texture), which is generally expected to increase with aggressive prostate cancer (21). For various 3D and 2D features (Fig. 4D and E; Supplementary Table S3), ROC curves were generated to quantify the ability of the features to stratify patients based on 5-year BCR outcomes. When comparing analogous 3D and 2D glandular features, the 3D features largely exhibit an improved correlation with 5-year BCR outcomes in comparison with their 2D counterparts. This is exemplified by the significant P values for the 3D features showcased in Fig. 4B and C (between BCR and non-BCR groups) and higher area-under-the-ROC-curve (AUC) values (Fig. 4D and E).

We also extracted the 3D skeleton of the lumen network and quantified its branching parameters (skeleton-derived features). Here, a “gland skeleton” is defined as a collection of lines that approximate the center axes of various glands as they propagate in 3D space (similar to a line representation of a vessel network). Example skeleton networks for benign and cancerous glands are shown in Supplementary Video S6. Due to the complex 3D branching-tree architecture of the gland-lumen network, there are no straightforward 2D analogues for these skeleton-derived features. In Fig. 4F, we show two examples of skeleton-derived features: the average branch length and the variance of the branch lengths. Both features are correlated with BCR outcomes based on P values and AUC values (Fig. 4F and G). Our analysis reveals that aggressive cancers (BCR cases) have shorter branch lengths and a smaller variance in branch lengths, which agrees with prior observations from 2D histology that glandular structures in higher grade prostate cancer are smaller and more abundant (i.e., less differentiated and varied in size). A histogram of branch lengths (Supplementary Fig. S7) demonstrates that the vast majority of branches are < 200-μm long, which suggests that the diameter of standard prostate biopsies (∼1-mm) is sufficient for whole-biopsy 3D pathology to quantify prostate cancer branch lengths with reasonable accuracy.

To explore the prognostic value of combining multiple glandular features, we used logistic regression models for feature selection and classification based on 3D versus 2D features (see Materials and Methods). Brief descriptions and AUC values for the 3D and 2D glandular features involved in training the multiparameter models are shown in Supplementary Table S3. The ROC curve of a model that combined 12 non-skeleton 3D features (“3D nonskeleton model”) yielded an AUC value of 0.80 ± 0.05 (average ± SD; Fig. 4H), which is considerably higher than the AUC value (0.65 ± 0.06) of the model trained with 12 analogous 2D features (2D model). By adding five skeleton-derived features to the 12 non-skeleton 3D features, a retrained 3D multiparameter model (3D model) yielded a slightly higher AUC value of 0.81 ± 0.05. The distribution of the 50 cases, based on their glandular features, can be visualized using t-distributed stochastic neighbor embedding (t-SNE), where a clearer separation between BCR and non-BCR cases is evident based on 3D versus 2D glandular features (Supplementary Fig. S8). Multiparameter classification models based on 3D features alone (nonskeleton) or 2D features alone were used to divide patients into high- and low-risk groups based on 5-year BCR outcomes (Supplementary Table S1), from which KM curves of BCR-free survival were constructed for a subset of cases in which time-to-recurrence (BCR) data are available (Fig. 4I). Compared with the 2D model, the 3D model is associated with a higher HR and C-index, along with a significant P value (P < 0.05), suggesting superior prognostic stratification.

As high-resolution biomedical imaging technologies continue to evolve and generate increasingly larger datasets, computational techniques are needed to derive clinically actionable information, ideally through explainable approaches that generate new insights and hypotheses. Interpretable feature-based analysis strategies in digital pathology generally hinge upon obtaining high-quality segmentations of key structural primitives (e.g., nuclei, glands, cells, collagen; refs. 18, 29). However, a common bottleneck to achieving accurate segmentations is the need for large amounts of manually annotated datasets (41). In addition to being tedious and difficult to obtain (especially in 3D), such annotations are often performed by one or more individuals who are not representative of all pathologists, thereby introducing an early source of bias. The use of simulated data has been explored to alleviate the need for manual annotations, and has been reported to be effective for training DL-based segmentation models for highly conserved and predictable morphologies [e.g., ellipsoidal nuclei (62, 63), or tubular vessel networks (63, 64)]. However, the 3D glandular networks of prostate tissues are highly irregular and variable, making it challenging to computationally generate simulated datasets. This complex and varied 3D morphology is also in part why 2D Gleason patterns may not be ideal for characterizing prostate glands.

The ITAS3D pipeline is a general approach for the volumetric segmentation of tissue structures (e.g., vasculature/endothelial cells, neurons, collagen fibers, lymphocytes) that can be immunolabeled with high specificity and that are also discernable to a deep-learning model when labeled with small-molecule stains like our H&E analogue or similar covalent stains (65). ITAS3D obviates the need for tedious and subjective manual annotations and, once trained, eliminates the requirement for slow/expensive antibody labeling of thick tissues (Fig. 1). The 2.5D segmentation approach employed in ITAS3D (i.e., image-sequence translation) offers an attractive compromise between computational speed/simplicity and accuracy for 3D objects that are relatively continuous in space (e.g., prostate glands). Details regarding 2.5D versus 3D image translation are provided in Supplementary Note S2 and Supplementary Table S4. In addition, a video summary of our ITAS3D-enabled gland-segmentation approach for prostate cancer assessment is provided in Supplementary Video S7. Note that in this specific implementation of ITAS3D, intermediate images are synthetically generated to mimic an IHC stain (CK8) that is routinely used by genitourinary pathologists. Therefore, it has the added advantage of enabling intuitive troubleshooting and facilitating clinical acceptance of our computational 3D pathology approach. A limitation of our implementation of ITAS3D is that the tissues used to train our deep-learning image-sequence translation modules were predominantly from Gleason pattern 3 and 4 regions, as well as benign regions. The reason for this was that the Canary TMA case–cohort study from which we derived our tissue specimens mainly consisted of low- to intermediate-grade (i.e., pattern 3 and 4) patients. In the future, improved ITAS3D performance over a wider range of prostate cancer grades would be facilitated by a more-diverse set of training specimens.

In this initial clinical study, we have intentionally avoided comparing our method with extant risk classifiers or nomograms that incorporate parameters based on human interpretation of 2D histopathology images [e.g., Kattan (66), CAPRA (67), and Canary-PASS (68)]. Rather, our goal has been to demonstrate the basic feasibility and value of 3D pathology by providing a direct comparison of intuitive 3D versus 2D glandular features analyzed computationally. A direct comparison of computational 3D pathology versus computational 2D pathology allows us to avoid the challenging variability and subjectivity of human interpretation. A human-observer study would require much-larger patient cohorts and a large panel of pathologist observers to account for inevitable interobserver variabilities. Nonetheless, to encourage the clinical adoption of 3D pathology methods, such large-scale studies will be necessary in the future, including prospective randomized studies on active surveillance versus curative therapies for low- to intermediate-risk patients (e.g., the PROTECT study; ref. 69), as well as studies to demonstrate the ability of computational 3D pathology to predict the response of individual patients to specific treatments such as androgen deprivation and both neoadjuvant and adjuvant chemotherapy. For completeness, in Supplementary Table S5, we provide Gleason scores for the 118 cancerous biopsies based on single-pathologist review (N.P. Reder) of whole-biopsy 3D pathology datasets with level-by-level examination of all 2D planes.

When trained with large numbers of images/cases and an optimal set of histomorphometric features, computational 2D pathology (based on whole slide images) has already been shown to be highly prognostic (21, 70–72). The goal of our study is not to suggest otherwise, but to provide early evidence of the additional prognostic value that 3D pathology can provide. The metrics presented in this study (Fig. 4) are intended to be comparative in nature (between computational 3D vs. 2D pathology) rather than regarded as definitive figures from a large prospective study. Our results show clear improvements in risk stratification based on 3D glandular features, both individually and in combination (Fig. 4B–I). As mentioned, the added prognostic value of 3D pathology is due in part to the significantly increased microscopic sampling of specimens (e.g., whole biopsies vs. sparse tissue sections). In addition, there are a number of advantages of 3D pathology datasets for computational analyses: (i) more-reliable segmentation of tissue structures due to the ability to leverage out-of-plane information (e.g., through continuity constraints) (ii); the ability to quantify tissue structures more accurately in 3D while avoiding 2D artifacts (73, 74), and (iii) the ability to extract novel prognostic features that cannot be derived from 2D tissue sections (e.g., gland-skeleton features).

As future work, ITAS3D can be used for the extraction and analysis of other 3D features (e.g., nuclear features, vascular features, and stromal features) to develop powerful classification models based on multiple morphologic primitives for a variety of tissue types. Annotation-free ITAS3D segmentation results, once available in sufficient quantities, can also be used to train end-to-end DL-based segmentation methods that bypass the image-translation step within ITAS3D. In the context of prostate cancer, studies are underway to identify additional prognostic 3D features based on our unique 3D-pathology datasets. A tiered approach to analyzing prostate cancer glandular features could be useful, such as first identifying broad classes of glandular morphologies (e.g., cribriform glands) and then analyzing class-specific features, as has been suggested in recent studies based on 2D whole-slide images of prostate cancer (70). Future studies should also aim to combine computational 3D pathology with patient metadata, such as radiomics, genomics, and electronic health records, to develop holistic decision-support algorithms (18). Nonetheless, as an initial step towards these goals, the results of this study, as enabled by the ITAS3D computational approach, provide the strongest evidence to date in support of the value of computational 3D pathology for clinical decision support, specifically for patients with low- to intermediate-risk prostate cancer.

W. Xie reports a patent for U.W. Ref. No. 49141.01US1, “Annotation-free segmentation of 3D digital pathology datasets with image-sequence translation” pending. N.P. Reder reports grants from Department of Defense and grants from Prostate Cancer Foundation during the conduct of the study; other support from Lightspeed Microscopy and grants from NIH outside the submitted work; in addition, N.P. Reder has a patent for “Annotation-free segmentation of 3D digital pathology datasets with image-sequence translation” pending. C.F. Koyuncu reports a patent for US20210049759A1 pending. P. Leo reports a patent for automated analysis of prostate pathology tissue pending and issued, and is presently employed by Genentech. K.W. Bishop reports grants from National Science Foundation during the conduct of the study. P. Fu reports grants from Case Western Reserve University during the conduct of the study. J.L. Wright reports grants from Merck, Altor Biosciences, Nucleix, Janssen, Movember; and other support from UpToDate outside the submitted work. A. Janowczyk reports grants from NIH during the conduct of the study; personal fees from Roche, Merck, and personal fees from Lunaphore outside the submitted work. A.K. Glaser reports other support from Lightspeed Bio outside the submitted work; in addition, A.K. Glaser has a patent for US10409052B2 issued and licensed to Lightspeed Bio, a patent for US20210033838 pending and licensed to Lightspeed Bio, a patent for US20210325656 pending and licensed to Lightspeed Bio, a patent for WO2021097300A1 pending and licensed to Lightspeed Bio, and a patent for WO2020167959A1 pending and licensed to Lightspeed Bio. A. Madabhushi reports grants from Bristol Myers-Squibb, AstraZeneca, Boehringer-Ingelheim; other support from Elucid Bioimaging and other support from Inspirata Inc; personal fees from Aiforia Inc, Caris Inc, Roche, Cernostics, and personal fees from Biohme outside the submitted work; in addition, A. Madabhushi has a patent 10,528,848 issued, a patent 10,769,783 issued, a patent 10,783,627 issued, and a patent 10,902,256 issued. L.D. True reports other support from LightSpeed Microscopy, Inc., outside the submitted work; in addition, L.D. True has a patent for patent no. 10,409,052 issued to University of Washington. J.T. Liu reports grants from NIH, US Department of Defense, and grants from National Science Foundation during the conduct of the study; other support from Lightspeed Microscopy, Inc., outside the submitted work; in addition, J.T. Liu has a patent for U.S. Patent No. 10,409,052. “Inverted light-sheet microscope,” filed on September 28, 2017, issued and licensed to Lightspeed Microscopy, a patent for U.S. Patent Application 17/356,135 “Apparatuses, Systems and Methods For Solid Immersion Meniscus Lenses,” filed 6/23/2021 pending and licensed to Lightspeed Microscopy, a patent for US Patent Application 17/310,309 “Apparatuses, Systems and Methods For Microscope Sample Holders,” filed 7/27/2021 pending and licensed to Lightspeed Microscopy, a patent for U.S. Patent Application No. 16/901,629, “Analysis of Prostate Glands Using Three-Dimensional (3D) Morphology Features of Prostate From 3D Pathology Images,” filed June 15, 2020 pending, and a patent for patent application 63/136,605 (provisional), “Annotation-free segmentation of 3D digital pathology datasets with image-sequence translation” filed on 1/12/2021 pending. No disclosures were reported by the other authors.

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation, the NIH, the Department of Defense, the Department of Veterans Affairs, or the United States Government.

W. Xie: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. N.P. Reder: Conceptualization, data curation, formal analysis, supervision, validation, investigation, methodology, project administration, writing–review and editing. C. Koyuncu: Software, formal analysis, validation, investigation, methodology, writing–review and editing. P. Leo: Software, formal analysis, validation, investigation, methodology, writing–review and editing. S. Hawley: Resources, data curation, writing–review and editing. H. Huang: Methodology, writing–review and editing. C. Mao: Resources, methodology, writing–review and editing. N. Postupna: Resources, methodology, writing–review and editing. S. Kang: Supervision, methodology, writing–review and editing. R. Serafin: Software, investigation, methodology, writing–review and editing. G. Gao: Investigation, methodology, writing–review and editing. Q. Han: Methodology. K.W. Bishop: Methodology, writing–review and editing. L.A. Barner: Methodology, writing–review and editing. P. Fu: Formal analysis, validation, methodology, writing–review and editing. J.L. Wright: Formal analysis, validation, writing–review and editing. C.D. Keene: Resources, formal analysis, supervision, methodology, writing–review and editing. J.C. Vaughan: Resources, supervision, methodology, writing–review and editing. A. Janowczyk: Formal analysis, investigation, methodology, writing–review and editing. A.K. Glaser: Software, methodology, writing–review and editing. A. Madabhushi: Resources, formal analysis, supervision, funding acquisition, validation, investigation, methodology, project administration, writing–review and editing. L.D. True: Conceptualization, data curation, formal analysis, supervision, funding acquisition, validation, investigation, methodology, project administration, writing–review and editing. J.T.C. Liu: Conceptualization, resources, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing.

The authors acknowledge funding support from the Department of Defense (DoD) Prostate Cancer Research Program (PCRP) through W81XWH-18-10358 (to J.T. Liu, L.D. True, and J.C. Vaughan), W81XWH-19-1-0589 (to N.P. Reder), W81XWH-15-1-0558 (to A. Madabhushi), and W81XWH-20-1-0851 (to A. Madabhushi and J.T. Liu). Support was also provided by the NCI through K99 CA240681 (to A.K. Glaser), R01CA244170 (to J.T. Liu), U24CA199374 (to A. Madabhushi), R01CA249992 (to A. Madabhushi), R01CA202752 (to A. Madabhushi), R01CA208236 (to A. Madabhushi), R01CA216579 (to A. Madabhushi), R01CA220581 (to A. Madabhushi), R01CA257612 (to A. Madabhushi), U01CA239055 (to A. Madabhushi), U01CA248226 (to A. Madabhushi), and U54CA254566 (to A. Madabhushi). Additional support was provided by the National Heart, Lung and Blood Institute (NHLBI) through R01HL151277 (to A. Madabhushi), the National Institute of Biomedical Imaging and Bioengineering (NIBIB) through R01EB031002 (to J.T. Liu) and R43EB028736 (to A. Madabhushi), the National Institute of Mental Health through R01MH115767 (to J.C. Vaughan), the VA Merit Review Award IBX004121A from the United States Department of Veterans Affairs (to A. Madabhushi), the National Science Foundation (NSF) 1934292 HDR: I-DIRSE-FW (to J.T. Liu), the NSF Graduate Research Fellowships DGE-1762114 (to K.W. Bishop) and DGE-1762114 (to L. Barner), the Nancy and Buster Alvord Endowment (to C.D. Keene), and the Prostate Cancer Foundation Young Investigator Award (to N.P. Reder). The training and inference of the deep learning models were facilitated by the advanced computational, storage, and networking infrastructure provided by the Hyak supercomputer system, as funded in part by the student technology fee (STF) at the University of Washington.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.