Predicting Prostate Cancer Aggressiveness via Diffusion Autoencoders with Multi-Parametric MR Images

This study proposes a novel framework that combines a diffusion autoencoder with an independent encoding structure for multi-parametric MR images (mp-FE), as a method for extracting imaging features for prostate cancer aggressiveness prediction. The diffusion autoencoder replaces the decoder part of a conventional autoencoder with a diffusion model, thereby guiding the encoder to produce latent vectors that are more informative and contain features useful for aggressiveness prediction. Additionally, to obtain latent vectors that capture the unique characteristics of each parametric MR image, the proposed mp-FE method constructs independent encoders for each parameter image and integrates the separately extracted features. Experimental results showed that, when reconstructing DWI images using the diffusion autoencoder with the mp-FE method, the model achieved the best performance, recording an accuracy of 70.97%, specificity of 79.41%, and AUROC of 0.72.