Extended Data Fig. 9: Development of MRI-PTPCa.

a) Training phase of MRI-PTPCa. b) Workflow of MRI-PTPCa using MRI for prostate cancer pathology prediction. c) Contrastive learning for 2-D images of MRI. d) Contrastive learning for 3-D images of mp-MRI. e) Interpretability of improvements from modeling view. The significant improvement of the model was from the progress of contrastive representation learning in image feature encoding and transformer in attention fusion. The mp-MRI foundation model of PCa provides the pre-trained parameters of the network and high-quality MRI features. Transformer-based model built attention and fusion among images of single sequences and multiple sequences. MRI-PTPCa enabled mp-MRI to surpass the limits of human vision, information association, and memory association in the characterization of prostate cancer. We also experimentally proved the importance of ground truth for modeling based on a supervised learning strategy. It was meaningful and efficient for modeling to explore the correlation between mp-MRI and pathology symmetrically. There is an inconsistency between needle biopsy evaluation and whole-mount histopathological analysis results of RP. Weakly supervised learning led to performance degradation from the modeling strategy, multi-classification networks were easier to detect differences between patients than binary classification networks. In addition, the richness of samples participating in training was also important to enhance performance, including the number of effective training samples and data augmentation types. Mp-MRI, multiparametric MRI; PCa, prostate cancer; CSPCa, clinically significant prostate cancer; NB, needle biopsy; RP, radical prostatectomy; MLP, multilayer perceptron; MLE, maximum likelihood estimation.