基于深度学习的膀胱癌多参数磁共振成像无监督配准方法研究

Research on unsupervised registration method of bladder cancer multi-parameter MRI based on deep learning

针对传统基于迭代优化的医学影像配准方法速度慢、泛化性差的问题,本研究提出了一种基于深度学习的膀胱癌磁共振成像(magnetic resonance image,MRI)跨模态无监督配准方法,并采用具有随机块采样的标准互信息(patch normalized mutual information,Patch-NMI)进行无监督训练。相比传统的迭代配准方法,本研究算法在进行膀胱癌动态增强成像(dynamic contrast-enhanced imaging,DCE)和T2加权像(T2-weighted imaging,T2WI)配准时,精度提升了1.3%,速度提高了20.42倍。实验结果表明,本算法在进行膀胱癌DCE和T2WI配准时,精度更高,速度更快。

To solve the slow speed and poor generalization of traditional medical image registration based on iterative optimization,we proposed a cross-modal unsupervised registration method of bladder cancer MRI based on deep learning.The patch normalized mutual information(Patch-NMI)loss with random block sampling was used for unsupervised training.Compared with the traditional iterative registration method,the accuracy of the proposed algorithm was improved by 1.3%,and the speed was improved by 20.42 times in the registration of dynamic contrast-enhanced imaging(DCE)and T2-weighted imaging(T2WI).The results show that this algorithm has higher accuracy and faster speed in the registration of DCE and T2WI.