Abstract: 3D/2D registration is a key technique for intraoperative guidance, existing deep learning based registration methods extract image features through the network and regress the corresponding pose transformation parameters. The method relies on real samples and their corresponding 3D labels for training, however, this part of expert-annotated medical data is scarce. The alternative solution trained the network with DRR images, which struggled to keep the original accuracy on X-rays due to the differences of image features across domains. For the above problems, an unsupervised cross-domain transfer registration network based on self-attention was designed, without relying on X-rays and it's 3D spatial labels as the training samples, the image features captured in the source domain and the correspondence between spatial transformations were migrated to the target domain. The public features reduced the disparity of features between domains to minimise the negative impact of cross-domain, the mTRE of the model predicted result is 2.66mm. Compared to the model without cross-domain transfer training there was a 70.61% reduction in the mTRE metrics, the effectiveness of the design method in the cross-domain registration task was indicated.

Key words: image registration, surgical navigation, cross-domain transfer, deep learning, self-attention

摘要： 3D/2D配准是术中导航的关键技术，现有基于深度学习的配准方法通过网络提取图像特征，继而回归出相应的姿态变换参数。此类方法依赖于真实的样本以及对应的三维标签用于训练，然而这部分专家标注的医疗数据十分稀缺。替代的方案以数字重建放射影像（DRR）图像训练网络，由于图像特征跨域的差异，在X射线上难以保持原有的配准精度。针对上述问题，设计了基于自注意力的无监督跨域迁移配准网络，无需依赖X射线与其对应的三维空间标签作为训练样本，将源域所捕获的图像特征与空间变换间的对应关系迁移到目标域，借助公共特征减少域间特征的差距降低跨域所带来的负面影响，模型预测结果的平均配准误差（mTRE）为2.66mm。与未经跨域迁移训练的模型相比，mTRE指标降低了70.61%，说明所设计方法的在跨域配准任务的有效性。

关键词: 图像配准, 手术导航, 跨域迁移, 深度学习, 自注意力