关键词: 医学图像配准, 无监督学习, 不确定性估计, 证据深度学习, 可信机器学习

Abstract: Uncertainty quantification in medical image registration is crucial for surgeons to evaluate surgical risk based on the trustworthiness of registered image data in real-world clinical trials. While deep unsupervised learning-based registration models have shown promise, there is a lack of methods to estimate appearance uncertainty, which directly affects registration accuracy and trustworthiness. Further, real-time applications necessitate registration models that are both highly accurate and fast in inference. To address these issues, the EvidentialMorph model was proposed, which applies EDL (Evidential Deep Learning), a promising approach for uncertainty quantification without additional computational cost, to unsupervised deformable medical image registration. The EvidentialMorph model was designed to first learn the DVF (Deformation Vector Field) through a registration backbone network module with a U-Net architecture, and then to learn the NIG (Normal Inverse Gamma) distribution of the registered image through an evidential STN (Spatial Transformer Network) module, providing parameters that directly indicate the registered image and its appearance uncertainty. Applied to several MRI (Magnetic Resonance Imaging) datasets, results show that EvidentialMorph improves DSC (Dice Similarity Coefficient) and NCC (Normalized Cross-Correlation) scores by 20 percentage points and 25 percentage points, respectively, over the CLMorph model and reduces inference time by 85 ms. EvidentialMorph outperformed baseline models, demonstrating effective uncertainty quantification and improved registration accuracy, offering potential for real-time applications and improved registration outcomes.

Key words: medical image registration, unsupervised learning, uncertainty estimation, evidential deep learning, trustworthy machine learning