OMMIR: a dynamic-optimization-based multi-modal medical image registration framework for CT-MR
Xian Liang, Yinwei Zhan, Yang Ouyang, Caizhen Wei · 2025
Deformable medical image registration plays a significant role in medical image analysis. However, when registering multimodal medical images, such as CT and MR images, direct registration often leads to artifacts and unclear correspondence, due to the differences in imaging principles and anatomical structures, and the absence of suitable similarity metrics. For this, we introduce OMMIR, a Transformer-CNN encoder-decoder model for multimodal CT-MR image registration that utilizes a fine-grained feature extractor(FE) module to extract multi-scale multi-channel features, and incorporates a real-time guidance optimization module (Opt) during training to optimize output and reduce unnatural distortions. Experiments with mainstream models and ablation demonstrate that OMMIR achieves optimal registration performance in CT-MR registration.