基于生成对抗网络的多模态可变形医学图像配准

Multimodal Deformable Medical Images Registration Based on Generative Adversarial Networks

多模态医学图像的精确配准对于医生进行病情分析和诊断至关重要。针对传统配准方法迭代优化周期长、成本高,以及现有基于深度学习的配准模型对不同模态医学图像之间固有的形变配准效果差等问题,提出了一种基于可变形卷积和级联结构的生成对抗配准模型。级联串接可变形卷积的Unet网络构造了新型的生成模型,并引入基于非扩展熵的相似性测度。通过多模态配准实验,证明了该模型配准精度高、用时少、具有一定的临床应用价值。

Accurate registration of multi-modal medical images is very important for doctors’condition analysis and diagnosis.The traditional registration method features a long iterative optimization cycle and high cost,and the existing deep learning-based registration model has inherently poor registration effects between different modalities of medical images.A generative adversarial registration model based on deformable convolution and cascaded structure is proposed.A new generative model is constructed by cascading the Unet network with deformable convolution in series,and the similarity measure based on non-expanded entropy is introduced.Through the multi-modal registration experiment,it is proved that the model has high registration accuracy,less time,and certain clinical application value.