Deep Learning-Based Prediction of Prostate Cancer Relapse Risk using Electronic Health Records and Histopathology Images – Patrick Fuhlert

For the selection of optimal patient treatment, survival prediction methods that estimate the expected time to an event of interest can be utilized. These models aim to provide accurate disease predictions such as for cancer patients based on characteristics of an individual patient or group. I will present how deep learning-based survival prediction models can be utilized in the context of prostate cancer diagnostics. We have developed a survival prediction model called Discrete Calibrated Survival. The developed approach outperforms the commonly used Cox model regarding relapse-free survival prediction on a high quality dataset of radical prostatectomy patients.

To assess prostate cancer severity, pathologists traditionally use Gleason grading. However, this grading system suffers from large interobserver variability. Our newly developed deep learning-based cancer risk prediction model, called Prostate Cancer Aggressiveness Index, can surpass the Gleason grading system. Instead of relying on subjective annotations by human experts, this model utilizes the objective endpoint of relapse-free survival to assess the risk of individual tissue images.

About the speaker