Uncertainty-Guided Nonrigid Registration Network for MR Image

Medical image registration is an important step in medical image analysis, which involves spatially aligning medical images obtained at different times, with different devices or imaging modalities. This process helps improve the accuracy of clinical diagnosis, assists in surgical planning, and promotes the development of medical research. Deep learning-based nonrigid medical image registration has achieved remarkable progress. However, variations in brightness, contrast, and noise levels across different data domains pose significant challenges for traditional similarity-based deep learning registration algorithms. These algorithms often struggle to maintain desired accuracy and stability in complex cross-domain scenarios, ultimately limiting their clinical generalizability. In this paper, we propose an uncertainty-guided nonrigid registration algorithm that produces robust registration outcomes and reliable uncertainty estimates without increasing computational complexity. This approach effectively mitigates noise and artifact interference, thereby enhancing overall performance. The proposed algorithm explicitly models uncertainty using Subjective Logic by parameterizing extracted feature representations as a Dirichlet distribution. This allows backbone network predictions to be interpreted as probabilistic descriptions, accurately quantifying uncertainty at each deformation field location and providing reliable confidence estimates for image mapping. Experimental results demonstrate that our method exhibits significantly superior generalization performance in cross-institutional datasets. Furthermore, comparisons with state-of-the-art medical image registration methods comprehensively validate the effectiveness and potential applicability of the proposed method in multi-domain registration tasks.