Enhanced Performance of Prostate Cancer Aggressiveness Prediction through Fusion of Multi-parametric MR Images and Additional Training of a Self-supervised Learning Model

This study proposes a method to improve cancer aggressiveness prediction performance by fusing multi-parametric MR images in various ways and additionally training a self-supervised learning model with data used for prostate cancer aggressiveness prediction. Experimental results showed that the prediction performance of prostate cancer aggressiveness improved when using multi-parametric MR images compared to using single-parametric MR image. The prediction fusion method of majority voting method with multi-parametric MR images showed the highest performance. Furthermore, the model with additional self-supervised learning on the dataset used for prostate cancer aggressiveness prediction showed an average performance improvement of 0.8% in accuracy, 4.49% in sensitivity, 1.77% in negative predictive value, and 0.02 in AUROC score across all single and multi-parametric MR image experiments.