Training Unsupervised Deep Learning Model for Multimodal Prostate Image Registration
Diana Tsvetkova · 2025
Prostate cancer is a serious but slow-progressing disease that often develops without symptoms for decades and its early detection is critical for better patient outcomes. A technique called image registration is often used for enhancing early prostate cancer assessment, where mainly MRI/CT and MRI/CT-PSMA images are aligned into a common coordinate system. The registration task can be solved as an iterative-optimization problem as well as a deep learning-based problem. As computational power and data availability grow, deep learning continues to reshape clinical workflows and decision-making processes in the medical healthcare domain. This paper presents an unsupervised deep learning approach for MRI-CT prostate image registration based on the VoxelMorph U-net architecture. Specific challenges related to real-world clinical datasets are addressed through preprocessing and augmentation strategies. The performance of the proposed method is evaluated using Normalized Cross-Correlation (NCC), Mutual Information (MI), and Joint Entropy, supported by visual comparisons of the registration results and analysis of the training process.