Development and Prospective Validation of a Pathology-Based Artificial Intelligence Model for Predicting the Time to Castration Resistance of Prostate Cancer

The goal of this predictive test is to prospectively test the performance of pre-developed artificial intelligence (AI) predictive model for predicting the time to castration resistance of prostate cancer. Investigators had developed this AI model based on deep learning algorithms in preliminary research, and it performed well in retrospective tests.

Hormone therapy is an important treatment method for prostate cancer and can effectively extend the survival of patients. However, almost all patients will progress to castration-resistant prostate cancer at different times. Current Hormone therapy options include androgen deprivation therapy(ADT), anti-androgen receptor(AR), and chemotherapy, with combination therapy being more effective in the early stages but associated with greater side effects. Therefore, predicting the time to castration-resistant progression and using this information to apply personalized treatment plans can ensure efficacy while reducing drug side effects. Therefore, we have developed an artificial intelligence predictive model for predicting the time to castration resistance of prostate cancer, which is expected to accurately predict the progression time for different patients and assist doctors in making personalized and precise treatment plans based on individual progression risks. This study is a predictive test with no intervention measures, planning to collect pathological slides of prostate biopsy from the enrolled patients and digitise them into whole-slide images (WSIs). The AI model will analyse the WSIs and generate slide-level predictive results (within 12 months, between 12 to 24months or over 24 months). The routine therapy and examination will be performed as usual. These two processes will not interfere with each other. Then we will follow-up the patients for 24 months, to record the time to castration-resistant progression, then we will compare the results with predictive model.