Non-rigid medical image registration based on residual-in-residual dense block and GAN [融合密集残差块和GAN变体的医学图像非刚性配准]

作者：Zhang G.; Hu Q.; Gong L.

来源：Journal of Image and Graphics, 2020, 25(10): 2182-2194.

DOI：10.11834/jig.200130

• Generative adversarial network(GAN)
• Least square loss
• Medical image
• Non-rigid registration
• Relative average GAN(RaGAN)
• Residual-in-residual dense block(RRDB)

Objective: Medical image registration is an important research task in the field of medical image analysis, which is the basis of medical image fusion and medical image reconstruction. Conventional registration methods that build loss function based on normalized mutual information by using iterative gradient descent to achieve registration tend to be time consuming. The existing deep learning-based medical image registration methods have limitation in registering medical images with large non-rigid deformation, which cannot achieve high registration accuracy and have poor generalization ability. Thus, this paper proposes a method to register multi-modal medical images by combining residual-in-residual dense block (RRDB) with generative adversarial network (GAN). Method: First, RRDB are introduced to the standard generator network to extract high-level feature information from unpaired image pairs; thus, registration accuracy is improved. Then, a least-squares loss is used to substitute cross-entropy loss constructed by the logistic regression objective. The convergence condition of least-squares loss is strict and promotes the model to reach convergence at the optimal parameters, which can alleviate gradient disappearance and overfitting; thus, the robustness of model training is enhanced. In addition, relative average GAN (RaGAN) is embedded into the standard discriminator network, namely, adding a gradient penalty in the discriminator network, which reduce the error of discriminator to estimate the relative true and false probability between the fixed image and moving image. The enhanced discriminator can help the generator to learn clearer and texture information; therefore, the registration error can be decreased, and the registration accuracy can be stabilized. Result: This registration model is trained and validated on DRIVE(digital retinal images for vessel extraction) dataset. Generalization performance tests are performed on Sunybrook Cardiac Dataset and Brain MRI Dataset. Compared with state-of-the-art registration methods, extensive experiments demonstrate that the proposed model achieves good registration results; both registration accuracy and generalization ability have been improved. Compared with the basic literature, the registration Dice values of retinal images, cardiac images, and brain images are improved by 3.3%, 3.0%, and 1.5%, respectively. According to the stability verification experiment of the registration model, as the number of iterations augments, the Dice value of this paper gradually increases, and the change is more stable than that of baseline literature. The number of iterations in this paper is 80 000, whereas that of baseline literature is 10 000. This verification experiment shows that this registration model has been effectively improved in the training phase; not only the convergence speed is accelerated but also stability is higher compared with that of baseline literature. Conclusion: The proposed registration method can obtain high-level feature information; thus, the registration accuracy is improved. Simultaneously, the loss function is built on the basis of the least-squares method, and the discriminator is also strengthened, which can enable the registration model to quickly converge during the training phase and then improve model stability and generalization ability. This method is suitable for medical image registration with large non-rigid deformation.