Summary

For patients with prostate cancer, the Gleason score and grade group are crucial parts of their diagnosis and significantly influence treatment planning. Pathologists determine the score and group through microscopical assessment of tissue removed from the prostate after, e.g., a biopsy procedure. While the Gleason grading system is the strongest tissue-based marker, it is inherently subjective. Consequently, it exhibits substantial variability between pathologists and even within the same pathologist at different points in time. Prostate cancer diagnostics could thus benefit from more robust, reproducible Gleason grading. Artificial intelligence holds promise for improving this process with its ability to learn from large amounts of data.

Chapter 1 gave a global overview of the fields at the basis of this thesis. We started by explaining the epidemiology of prostate cancer and the diagnostic pathway. Second, we discussed the Gleason grading system and the difficulties that co-exist with this grading standard. Then we explained the digital revolution taking place in pathology and the subsequent introduction of computational pathology. Finally, we set out the overarching aim of this thesis: Improving prostate cancer diagnostics by developing artificial intelligence algorithms that can support pathologists.

Acquiring substantial amounts of highly detailed data for algorithm development is a challenge within computational pathology. Chapter 2 of this thesis aimed to overcome the annotation burden by employing immunohistochemistry (IHC) to annotate data for model training semi-automatically. Using CK8/18 and P63 to highlight the epithelial tissue and basal cell layer, we trained an IHC-based segmentation algorithm. We could then transfer this model’s output using image registration to H&E stained slides and train a highly accurate epithelium segmentation system in H&E. This proved to be an effective method to annotate large slides and drastically reduced the annotation burden for higher-level annotation tasks, e.g., cancerous glands.

Our next challenge was to build a system that could diagnose and grade prostate biopsies at the level of a pathologist (Chapter 3). For this, we required a large dataset of annotated biopsies, which was not readily available. Utilizing the algorithm from Chapter 2, we could semi-automatically annotate a large dataset of almost 6000 biopsies and used this for model training. We compared the AI system with the grading performance of a panel of pathologists and residents. We found that, on average, the AI system showed a higher agreement with experts than the panel, operating at the level of experienced pathologists, and giving concrete proof of the power of AI for Gleason grading.

Additional research was required to further assess the potential of AI for Gleason grading and see if results would translate to more extensive and diverse datasets. In Chapter 4, we introduced the PANDA challenge: an international competition to crowd-source AI innovations for automated Gleason grading. In this challenge, over 1200 algorithm developers worked towards building the best-performing algorithms. Based on performance during the challenge, we selected fifteen algorithms to undergo rigorous independent validation. Unique to this challenge was that we performed the validation entirely blinded to the algorithm developers, minimizing potential bias. We showed that these AI algorithms are capable of pathologist-level performance across cross-continental datasets with various reference standards.

This thesis started with the aim of empowering pathologists with AI to improve the diagnostic process. To this extent, we evaluated the potential benefits of AI when such algorithms are placed in the hands of pathologists. In Chapter 5 we invited pathologists to use our AI algorithm and compared their performance to an unassisted setting. We found that when pathologists used the algorithm, they significantly improved in determining the Gleason grade group. In Chapter 3, the AI system outperformed most pathologists. However, with AI support, pathologists now outperformed both the standalone AI as the pathologists without AI. These results show the existence of a pathologist-AI synergy: algorithms and pathologists both have their strengths and weaknesses that can complement one another.

In Chapter 6 we summarized the main findings of this thesis and outlined several key characteristics of computational pathology that shape current and future projects. First, we highlighted the importance of the availability of large high-quality datasets. Data is at the core of many AI innovations, and the projects conducted as part of this thesis are no exception. Second, we discussed steps that need to be taken to bring these AI innovations to the clinic. Developing robust models that generalize to external labs, introducing fail-safes, and human-algorithm-interaction design are crucial prerequisites for implementing AI in clinical practice. Third and finally, we discussed an outlook for the future, focusing on applications beyond human-based reference standards.

Concluding, we investigated how artificial intelligence can improve prostate cancer diagnostics. Specifically, we showed that AI algorithms are able to achieve pathologist-level performance at a highly complex task such as Gleason grading. AI can assist pathologists, and together they can improve the diagnostic process and ultimately deliver a more accurate diagnosis for patients.