CaP is a disease with a broad spectrum of aggressiveness. While modest progress has been made in the treatment of the aggressive forms of CaP and refined active surveillance protocols have been adapted to better manage more indolent forms, it remains challenging to provide guidance for those at intermediate risk of recurrence. The adaptation of whole-slide scanners for digitizing entire tissue slides has promoted the rapid development of machine learning and deep learning tools that facilitate the mining of subvisual cell morphology features on H&E stained histology images.21,22,38,39 Multiple applications of CNNs have been developed to characterize histological images to improve both cancer diagnosis and prognostic risk classification 19.23,24,27,28,40,41 In prostate cancer histology, the majority of deep learning applications have focused on generating categorical GGs. However, the challenge of pathologist GG assessment centers mostly on intermediate GG 2 and 3 tumors, where pathologists estimation of GP percentages is required.13 In our study, we have taken the next step to generate these quantitative measures of GPs. Our results significantly add to the literature of artificial intelligence augmented pathological assessment by showing that computer generated quantitative measure of GPs can be achieved. Such a tool could have near term translational impact with the potential to improve both the accuracy and reproducibility in the pathologist’s assignment of GG 2 and 3.
Further, a growing number of studies have demonstrated that %GP4 can provide additional information to patient recurrence risk estimation.15,17,18 However, these studies relied on the pathologist’s subjective quantification of %GP4, which is prone to both intra- and inter-observer variability. Our results show that an increase in CNN-assessed %GP4 was significantly associated with increased BCR risk within GG 2 tumors, supporting the literature that %GP4 can provide more granular risk stratification.
The findings of this study also highlight the heterogenous nature of CaP. While the ordinal GG system remains the single most effective pathological predictor of recurrence risk in patients with localized CaP, our results show that there is more information in the quantitative cellular differences within these categories. While the current study is focused on %GP4 within intermediate GG 2 and 3 tumors, the overall approach could also be expanded to include the assessment of percentage GP5 (%GP5) GG 5 tumors. Further, within GG 3–5 tumors, there is also heterogeneity within GP4 tissue morphology ranging from poorly formed, glomeruloid, cribriform to fused glands, each with different clinical significance42,43. For example, the cribriform pattern of prostate cells has increasingly been recognized as a clinically relevant and aggressive sub-pattern of GP4, and similar to %GP4, notation of its presence and prevalence in pathology reports has recently been recommended by ISUP44. Additionally, recognition of a particularly aggressive expansile subtype of the cribriform pattern45 is emerging. Expanding our approach for computational identification and quantification of %GP5 and these sub-patterns could add information to move the Gleason grade system to include a full range of quantitative assessment of these various histological subtypes, capturing a greater degree of heterogeneity in CaP and providing more information for recurrence risk assessment.
While the multi-focal nature of CaP is established, the current study and the grading system is focused on the pathologically assessed primary tumor nodule GG. While the primary tumor is considered the disease driver, the secondary and tertiary foci are capable of metastatic seeding and impacting tumor prognosis.46 Applying the trained CNN model to multiple tumor foci obtained from whole mount prostatectomy specimens could provide another unique area of expansion for this area of digital pathology research. Specifically, such a prostate tumor-wide application could provide a global assessment of %GP4 as well as a characterization its spatial tumor heterogeneity. Neither of these ideas have been extensively studied, and given that their impact on CaP prognosis is unknown, their future study, facilitated by our novel computational approach, appears justified.
While there is clinical potential for translation of the CNN-assessed %GP4, additional steps are needed to realize it. First, the generalizability of the model beyond our single institution study needs assessment. Such an assessment would require not only specimens obtained from additional health care institutions but also study of the effects of both H&E staining variation and different whole slide image systems on the accuracy and reproducibility of %GP4. Second, the current model is constructed and validated on a modest number of GG 1 and 4 CaP tumor foci. As the accuracy and generalizability of the CNN model is dependent upon the number and diversity of the foci used to create it, future studies should also expand the number of foci images used for training. Third, the current study has focused on the assessment of %GP4 within pathologic specimens obtained from radical prostatectomies. The clinical applicability of the model could be further broadened by studies of its utility in the biopsy space prior to primary treatment selection. Such information from biopsy-based GG assessment may increase the diagnostic accuracy of those with more indolent disease, which may minimize the burden of unnecessary primary treatment within these individuals.
In summary, our findings show that a robust assessment of %GP4 is achievable using an artificial intelligence CNN approach based on high resolution histopathologic images of CaP tumors. Further, the application of this information shows clinically significant differentiation of BCR risk within GG 2 tumors. Consistent with the existing literature, this information could be used to refine risk stratification for GG 2 tumors. Further, this digital uropathology tool might be particularly impactful in smaller health care centers that are less likely to have pathologists with specialized urologic training. Future large and racially diverse studies should also evaluate the incorporation of GP percentages within biopsy- and surgery-based CaP nomograms to establish the inclusion of this quantitative histopathologic information in CaP recurrence risk assessment.