Objective Medical image registration is a spatial transformation process that aligns and matches the specific spatial structures contained in two medical images. It has been applied in disease detection, surgical diagnosis and treatment, and other fields. Traditional medical image registration methods are slow and computationally expensive. In recent years, researchers have made significant breakthroughs in medical image registration research using deep learning methods. Deep learning methods have increased the registration speed by hundreds of times, with a registration accuracy comparable to those of traditional methods. However, most patients have complex pathological conditions and lesions grow quickly, resulting in significant differences in the images collected at different stages. Existing deep learning-based registration methods have low registration accuracy and poor generalization performance when used for medical images of large deformations. Therefore, a multi-scale constraint network (MC-Net) for large-deformation 3D medical image registration Based on multi-scale constraints is proposed. Methods We propose a multi-scale constraint network (MC-Net) for large-deformation 3D medical image registration based on multi-scale constraints. Three multi-kernel (MK) modules are designed as parallel multi-channel and multi-convolution kernels for the encoder to accelerate the training speed. A convolutional block attention module (CBAM) is added to skip connections and enhance the ability to extract complex semantic information and fine-grained feature information from large-deformation images. In order to improve the registration accuracy, MC-Net combines multi-scale constrained loss functions to implement a layer-by-layer optimization strategy from low resolution to high resolution. Results and Discussions In an experiment, three publicly available 3D datasets (OASIS, LPBA40, and Abdomen CT-CT, with two modalities) were used for registration research. The effectiveness of MC-Net was demonstrated through original experiments, traditional comparison methods, deep learning comparison methods , ablation experiments, and multi-core fusion experiments. Based on the registration results shown in Figs. 5 and 6, MC-Net performed well in the registration of the OASIS and LPBA40 brain datasets, as well as for the Abdomen CT-CT abdominal dataset. In the brain image comparison experiment, the LPBA40 brain dataset was compared with a traditional registration method (ANTs) and three deep learning registration methods (VoxelMorph, CycleMorph, and TransMorph) in the same experimental environment. It was found that MC-Net outperformed the other methods in terms of detail registration in brain regions and overall brain contour deformation. The abdominal image comparison experiment compared two traditional methods (ANTs and Elastix) and two deep learning methods (VoxelMorph and TransMorph). It was found that MC-Net had some shortcomings in organ generation and contour deformation, but had better registration performance than the other methods in terms of blank area size and individual organ deformation. The ablation experiment was conducted using the LPBA40 dataset. It demonstrated the different roles of the MK and CBAM modules in processing medical images in MC-Net, which helped to improve the registration accuracy. In addition, this article also discusses the computational complexity of MC-Net. For For large target images such as medical images, this article discusses how a multi-kernel (MK) fusion module can be designed to effectively reduce the computational complexity. Conclusions In response to the low accuracy and poor generalization performance of current large-deformation image registration methods, this paper proposes a medical image registration network (MC-Net) based on multi-scale constraints, with LPBA40, OASIS, and Abdomen CT-CT medical image datasets used as research objects. Information loss can be avoided by designing CBAM modules in skip connections to enhance the ability to extract differential information from large-deformation images. In addition, considering the slow registration speed caused by the large number of parameters when processing large-deformation images, the MK module was designed with a parallel path large kernel convolution structure to improve the registration speed without affecting registration accuracy. When combined with the multi-scale constraint loss function proposed in this article, it iteratively optimizes the deformation fields at three scales from low resolution to high resolution to improve the registration accuracy. The experimental results show that compared with other methods, this method has improved registration accuracy, speed, and computational complexity. The good registration performances in three datasets with MRI and CT modalities demonstrate the generalization ability of our method. Subsequent research will focus on designing an adaptive adjustment module for multi-scale constrained loss function hyperparameters, in order to solve the problem of the time-consuming hyperparameter tuning needed for loss functions in experiments and improve the experimental efficiency. In summary, MC-Net has practical value in the registration of large-deformation images.
