Personalized prostate cancer management : AI-assisted prostate pathology and improved active surveillance

Abstract

Prostate cancer is a major global health concern and is the most common cancer-related cause of death in Sweden. Prostate cancer screening using PSA has been shown to reduce prostate cancer mortality but also leads to significant overdiagnosis and overtreatment of low-risk cancers. Improved risk stratification and effective active surveillance are crucial to balancing the benefits of screening with the risk of overdiagnosis and overtreatment.

In Study I, we studied the uptake and the follow-up of active surveillance using a retrospective cohort of patients who were diagnosed with low-risk prostate cancer between 2008 and 2017 in Stockholm County. Our results showed that only 50% of eligible active surveillance patients received active surveillance as their primary treatment choice at diagnosis. Most men that enrolled in active surveillance remained on surveillance during the first years after diagnosis (82% during a median 3.5 years), but did not receive a follow up according to guidelines with regard to repeat biopsies and PSA tests.

Current clinical practice has seen an increase in the use of magnetic resonance imaging (MRI) and the incorporation of risk prediction models to select men with the highest suspicion of clinically significant prostate cancer for prostate biopsy. However, the effectiveness and how MRI and risk prediction models should be incorporated into active surveillance follow-up have yet to be established. Study II evaluated the performance of MRI-targeted biopsies and a blood-based risk prediction model (the Stockholm3 test) for monitoring disease progression in patients on active surveillance and compared this to the conventional follow-up using PSA and systematic biopsies. When MRI-targeted and systematic biopsies were combined, the detection rate of clinically significant prostate cancer increased when compared to conventional systematic biopsies. Biopsies performed in MRI-positive men resulted in a 49% reduction in performed biopsies, at the expense of failing to diagnose 1.4% clinically significant prostate cancer in MRI-negative men. The incorporation of the Stockholm3 test showed a 27% reduction in required MRI investigations and a 57% reduction in performed biopsies compared to performing only systematic biopsies.

In Study III, we digitized biopsy cores from STHLM3 participants to develop an artificial intelligence (AI) for prostate cancer diagnostics. The AI system demonstrated clinically useful performance that was comparable to that of the study pathologist for cancer detection (AUC of 0.986) and for predictions of cancer length (correlation of 0.87) and grading performance that was on par with that of expert prostate pathologists.

In Study IV, we developed a conformal predictor to estimate the uncertainty of the predictions for the model in Study III. The uncertainty estimates were used to control the error rate so that only predictions with high confidence are accepted and unreliable predictions can be detected. The conformal predictor was able to identify unreliable predictions as a result of variations in digital pathology scanners, preparation of tissue in different pathology laboratories, and the existence of unusual prostate tissue that the AI model was not exposed to during training.

Little is known about the relationships between prostate cancer genetic risk factors and the morphology of prostate tissue. In Study V:, we investigated whether weakly supervised deep learning can learn to detect such possible associations. The findings in this paper imply relationships between prostatic tissue morphology and genetic risk factors for prostate cancer, particularly in young men. These results provide proof of principle for exploring the use of morphological information in multi-modal prostate cancer risk prediction algorithms.

In conclusion, the purpose of this thesis was to describe possible extensions to improve prostate cancer active surveillance management, as well as to develop prediction models for improved prostate cancer diagnostics.