Unsupervised 3D deformable image registration method

Abstract

The application discloses a kind of unsupervised three-dimensional deformable image registration methods, first with U-Net backbone network as foundation, design and build big kernel ghost module deformable image registration network model, and through training set and verification set, the model is trained and verified;Then the image to be registered is regarded as fixed image, and the template image corresponding to the image to be registered is regarded as moving image;The moving image and fixed image are spliced;Then the spliced image is input into the deformable medical image registration network model, and the deformation field is output;Finally, the moving image is resampled using the deformation field, and the deformed moving image is obtained, which is the registered image.The application can reduce the model parameter quantity and GPU memory occupation while improving the registration accuracy.

Description

Technical Field

[0001] This invention relates to improvements in image registration techniques, and particularly to an unsupervised 3D deformable image registration method based on a large kernel ghost module, belonging to the field of deep learning computer vision. Background Technology

[0002] Deformable image registration (DIR) is a technique designed to find non-rigid voxel-space correspondences, or deformable fields, to visually align moving images with stationary images. In the medical field, DIR primarily improves the accuracy, efficiency, and safety of medical image analysis, as well as optimizing the effects of radiotherapy. Generally, DIR technology automatically or semi-automatically aligns multiple medical images in the same coordinate system, providing more comprehensive anatomical information and helping doctors or researchers identify regions of interest. DIR also has significant applications in non-medical fields. For example, in remote sensing, DIR can be used to register remote sensing images acquired at different times or from different sensors for change detection or feature extraction; in bioinformatics, DIR can be applied to protein structure analysis, comparing the three-dimensional structures of different proteins to discover their similarities and differences; and in virtual reality (VR) and augmented reality (AR), DIR technology can help achieve seamless integration between the real world and virtual environments, providing a more natural and immersive user experience.

[0003] Deformable registration (DIR) is a resource-intensive task. For example, deformable registration of two MRI images of the same modality typically requires tens of millions of independent voxel-space correspondences to accurately and smoothly deform the moving image to a stationary one. This demands substantial computational resources to ensure fast registration. Traditional methods typically iteratively compute dense transforms, quantifying alignment by measuring the similarity between the deformed and stationary images during the process, and employing specific optimization strategies to achieve the highest similarity. While these targeted pairwise optimization strategies do provide accurate and smooth transforms, they are time-consuming and computationally expensive. Furthermore, when faced with unseen image pairs, they require iterative optimization from scratch, making them a suboptimal choice for real-time registration tasks. In contrast, deep learning (DL)-based DIR achieves registration speeds several orders of magnitude faster than traditional methods and produces competitive registration results, becoming the mainstream approach in recent years, as iterative optimization is performed only during the training phase and the deformable field can be directly output during testing.

[0004] Currently, deformable image registration (DIR) frameworks based on deep learning (DL), including convolutional neural networks (CNNs) and attention-dependent Transformer architectures, have been widely applied. CNNs excel at extracting local details. Balakrishnan et al. (2019) proposed the well-known VoxelMorph based on U-Net, and its results are quite competitive with the well-known traditional registration methods ANTs and NiftyReg. However, using only 3×3×3 convolutional kernels results in a relatively small receptive field of the model, thus weakening its ability to extract global information. Jia et al. (2022) proposed LKU-Net, which uses large kernel parallel convolution operations to replace 3×3×3 convolutional kernels, effectively expanding the receptive field. However, using large kernels increases the number of model parameters. In addition, Chen et al. (2022) introduced TransMorph, integrating the Transformer module into the U-net architecture, which significantly enhances the ability to extract global information. However, this approach also leads to a significant increase in the number of model parameters. Furthermore, current deep learning models used for medical direct image re-identification (DIR) typically require significant amounts of GPU memory, posing a major obstacle in resource-constrained clinical environments. Han et al. (2020) proposed a lightweight model called GhostNet, discovering that certain image features can be generated from basic features through computationally inexpensive deep-level convolutional operations. They termed these features generated by deep-level convolutions "ghost features." In GhostNet, the ghost module generates a sufficient number of ghost features with minimal parameter overhead through inexpensive deep-level convolutions, ensuring excellent model performance. It demonstrates promising results in 2D image classification and detection tasks, but its performance in 3D deformable image registration tasks remains unverified.

[0005] In summary, existing deep learning models either have low registration accuracy or improve registration accuracy at the cost of increasing the number of parameters and GPU memory usage, making it difficult to meet the needs of practical applications and have low competitiveness. Therefore, there is an urgent need for a lightweight registration model with high registration accuracy. Summary of the Invention

[0006] To address the issues of low registration accuracy, large model parameter count, and high GPU memory consumption in existing deep learning-based image registration methods, this invention proposes an unsupervised 3D deformable image registration method. This invention can improve registration accuracy while reducing the number of model parameters and GPU memory consumption.

[0007] The technical solution of this invention is implemented as follows:

[0008] An unsupervised 3D deformable image registration method, comprising the following steps:

[0009] 1) Construct a deformable image registration network model;

[0010] 1.1) Obtain the dataset; obtain a public dataset including several subject images and their corresponding template images, where the subject images and the images to be registered are of the same type; preprocess the subject images and treat the preprocessed subject images as fixed images. The dataset is divided into training, validation, and test sets; the template image is treated as a moving image. ;

[0011] 1.2) Model Building; Based on the U-Net backbone network, a deformable image registration network model with a large kernel ghost (LKG) module was designed and built;

[0012] 1.3) Model training and validation; [This involves] processing motion images. Compare with each fixed image in the training set. The images are stitched together and then input into the deformable image registration network model built in step 1.2) for training. During training, the model performance is verified using a validation set, thus obtaining the deformable image registration network model.

[0013] 2) Treat the image to be registered as a fixed image and the template image corresponding to the image to be registered as a moving image; stitch the moving image and the fixed image together to obtain a stitched image;

[0014] 3) Input the stitched image obtained in step 2) into the deformable image registration network model constructed in step 1); the deformable image registration network model outputs the deformation field;

[0015] 4) Use the deformation field to resample the motion image from step 2) to obtain the deformed motion image, which is the registered image.

[0016] Furthermore, during model training in step 1.3), for each stitched image input, the deformable image registration network model outputs a deformation field. Using deformation fields For the motion image corresponding to this training Resampling is performed to obtain the deformed motion image. That is, the image after registration during the training phase. Then, iterative backpropagation is used to minimize the loss function in order to optimize the similarity between images, as shown in the following equation;

[0017]

[0018] in Representing the optimal deformation field, the loss function is measured by similarity. and smoothing regularization term composition, As a parameter for regularization trade-offs.

[0019] Further, in step 1.2), during model building, in the encoding stage, the network initially uses two consecutive convolutional layers, followed by alternating applications of multiple convolutional layers and LKG modules to extract encoder feature maps. During this process, the height, width, and depth of the feature maps gradually decrease, while the number of channels gradually increases. In the decoding stage, the network alternately uses multiple convolutional layers and transposed convolutional layers to extract decoder feature maps and gradually restores them to the original resolution, with the transposed convolutional layer responsible for the upsampling process. After each upsampling, the decoder feature map at that resolution is concatenated with the encoder feature map at the same resolution through skip connections to prevent the loss of information in the encoder feature map. Finally, the network outputs a three-channel deformation field through several 3×3×3 convolutional layers.

[0020] Furthermore, the LKG module extracts feature maps according to the following steps:

[0021] a) Input features are processed through a 3×3×3 convolution operation, followed by a BN layer and a PR layer to extract intermediate feature maps, the number of which is half the number of output channels;

[0022] b) Large kernel parallel deep-level convolution operation followed by BN and PR; firstly, 1×1×1, 3×3×3 and k×k×k deep-level convolution operations are used to extract ghost feature maps respectively, where the large kernel k>3; then, all ghost feature maps and intermediate feature maps are fused together by element-wise addition, and then passed through BN layer and PR layer to obtain fused feature map.

[0023] c) The intermediate feature map obtained in step a) and the fused feature map obtained in step b) are concatenated to obtain the final output feature map.

[0024] Furthermore, in step 1.3), the similarity loss function used for model training is the normalized cross-correlation (NCC), as shown in the formula below.

[0025]

[0026] in It is a deformation field. It is a position within the image domain Ω; It is the original intensity image of the fixed image. The image after subtracting the average intensity of the local region, the deformed motion image It is the original An image minus the average intensity of local regions;

[0027] The smoothing regularization term used in training is a diffusion regularization term, as shown in the formula below.

[0028]

[0029] in It is a deformation field. It is a displacement field, where p is a position in the image domain Ω. It is a gradient operator used to calculate the gradient of the displacement field at a certain position p; the relationship between the deformation field and the displacement field is as follows:

[0030]

[0031] in Identity mapping, i.e. .

[0032] Furthermore, in step 1.3), the validation metric used for model validation is the dice similarity coefficient (DSC); the calculation method is as follows:

[0033]

[0034] Where A represents the segmentation set of the deformed moving image, B represents the segmentation set of the fixed image, |A| and |B| represent the number of elements in sets A and B respectively, and |A ∩ B| represents the number of elements in the intersection of A and B.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] Compared to existing deep learning registration methods, the deformable image registration network model proposed in this invention is a model for image (such as 3D MRI) registration based on a large kernel ghost module (LKG module). This module achieves a larger receptive field through a large kernel parallel strategy while retaining the ability to capture local information, while depthwise convolution reduces the number of model parameters and inference time without sacrificing performance. Results on public datasets show that compared to LKU-Net and TransMorph, the proposed model has a significantly reduced number of parameters and achieves the best results in terms of dice similarity coefficient (DSC) and GPU memory usage. In summary, the proposed model achieves a balance between lightweight design and high accuracy, showing considerable potential, especially in resource-constrained clinical scenarios. Attached Figure Description

[0037] Figure 1 This is a diagram illustrating the overall framework of the unsupervised deformable image registration based on a large kernel ghost module, as described in this invention.

[0038] Figure 2 This is a flowchart illustrating the implementation of the deformable image registration network model of the present invention.

[0039] Figure 3 This is a schematic diagram of the deformable image registration network model structure of the present invention.

[0040] Figure 4 This is a schematic diagram of the LKG module structure of the present invention.

[0041] Figure 5 A visualization of the comparative test results provided by this invention. Detailed Implementation

[0042] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0043] See Figure 1 This invention discloses an unsupervised three-dimensional deformable image registration method, the steps of which are as follows:

[0044] 1) Construct a deformable image registration network model;

[0045] 2) Treat the image to be registered as a fixed image and the template image corresponding to the image to be registered as a moving image; stitch the moving image and the fixed image together to obtain a stitched image;

[0046] 3) Input the stitched image obtained in step 2) into the deformable image registration network model obtained in step 1); the deformable image registration network model outputs the deformation field;

[0047] 4) Use the deformation field to resample the motion image from step 2) to obtain the deformed motion image, which is the registered image.

[0048] The deformable image registration network model is constructed according to the following steps.

[0049] 1.1) Obtain the dataset; obtain a public dataset including several subject images and their corresponding template images, where the subject images and the images to be registered are of the same type; preprocess the subject images and treat the preprocessed subject images as fixed images. The dataset is divided into training, validation, and test sets; the template image is treated as a moving image. ;

[0050] 1.2) Model building; Based on the U-Net backbone network, a deformable image registration network model with a large kernel ghost module was designed and built;

[0051] 1.3) Model training and validation; [This involves] processing motion images. Compare with each fixed image in the training set. The images are stitched together and then input into the deformable image registration network model built in step 1.2) for training. During training, the model performance is verified using a validation set, thus obtaining the deformable image registration network model.

[0052] The core of the registration method in this invention is a deformable isomodal image registration network model (referred to as LKGDIR) based on a large kernel ghost module. This model can complete registration tasks for medical images, including 3D brain MRI. The implementation process of the network model in this invention is as follows: Figure 2 As shown, the embodiment includes four steps: dataset preprocessing, designing and building the LKGDIR registration network model, model training, and model performance testing. The following description, in conjunction with the embodiment, explains each step.

[0053] Step S1: Dataset Preprocessing. This invention focuses on three-dimensional magnetic resonance imaging (MRI) images of the brain. All subjects' three-dimensional MRI images were used as stationary images, and template atlas images obtained from Kim et al. (2021) were used as moving images to complete template-to-subject registration. Standard preprocessing steps, including skull dissection, resampling to a voxel size of 1mm × 1mm × 1mm, affine pre-registration, and brain region segmentation, were applied to the dataset. All images were then cropped to a size of 160 × 192 × 224. The dataset was subsequently divided into training, validation, and test sets, containing 403, 58, and 115 images respectively (7:1:2). The purpose of brain region segmentation was to use segmentation maps of 30 brain regions to validate and test the accuracy of the registration.

[0054] Step S2: Design and build the LKGDIR registration network model. Since unsupervised registration network models have been a popular research direction in recent years, this invention adopts an unsupervised strategy to design the overall registration framework based on the LKGDIR network model. Then, the specific design of the LKGDIR network model structure is based on the currently stable U-Net network, which is an encoder-decoder structure. In addition, this invention provides a detailed structural design for the key component of the model, the large kernel ghost module (LKG module). Finally, this invention uses the Python programming language in the PyTorch environment to write the LKGDIR model, completing the model construction preparation work before model training.

[0055] Step S21: Design of the overall registration framework based on LKGDIR. Figure 1 This illustrates the overall framework of the model design of the present invention. For ease of understanding, rectangles and rounded rectangles here represent medical images. Fixed Image and motion images spliced ​​( The data is then input into the LKGDIR network to infer the deformation field. Subsequently, a deformation field was used. For moving images Perform resampling ( ), to obtain the registered image Then, iterative backpropagation is used to minimize the loss function in order to optimize the similarity between images, as shown in the following equation.

[0056]

[0057] in Representing the optimal deformation field, the loss function is measured by similarity. and smoothing regularization term composition, As a parameter for regularization trade-offs.

[0058] Step S22: Structural design of the LKGDIR network model. Figure 3 The structure of the LKGDIR network model is shown. In the encoding stage, the network initially uses two consecutive convolutional layers, followed by alternating applications of multiple convolutional layers and LKG modules to extract encoder feature maps. During this process, the height (H), width (W), and depth (L) of the feature maps gradually decrease, while the number of channels (C) gradually increases. In the decoding stage, the network alternately uses multiple convolutional layers and transposed convolutional layers to extract decoder feature maps and gradually restores them to the original resolution, with the transposed convolutional layers handling the upsampling process. After each upsampling, skip connections are used to concatenate the decoder feature map at that resolution with the encoder feature map of the same resolution to prevent information loss in the encoder feature map. Finally, the network outputs a three-channel deformable field through several 3×3×3 convolutional layers.

[0059] Step S23: Structural design of the LKG module. Figure 4 The architecture of the proposed LKG module is demonstrated. The LKG module consists of three steps: a) regular convolutional operations followed by batch normalization (BN) and PReLU activation (PR), b) deep-level convolutional operations with large kernel parallelism followed by BN and PR, and c) concatenation ( The first step extracts intermediate feature maps using 3×3×3 convolution operations, BN layers, and PR layers, with the number of channels being half the number of output channels. In the second step, ghost feature maps are first extracted using depthwise convolution operations of 1×1×1, 3×3×3, and k×k×k, where the large kernel k>3. Then, these ghost feature maps are added to the intermediate feature maps element-wise. The feature maps from the first and second steps are then fused, followed by Batch Normalization (BN) and Backpropagation (PR) layers to obtain a fused feature map. In this step, the large kernel significantly expands the receptive field without a substantial increase in the number of parameters, while the parallel use of regular and small kernels ensures the capture of local information. The shortcut preserves intermediate feature information. Finally, the feature maps from the first and second steps are concatenated to obtain the final output feature map.

[0060] Step S24: LKGDIR network model implementation. This invention uses the PyCharm programming environment and the Python programming language to build the model within the PyTorch deep learning framework. PyTorch is a deep learning tensor library based on Python and optimized for Torch, primarily used for tasks requiring GPU and CPU computation. Compared to other deep learning frameworks such as TensorFlow and Keras, PyTorch is favored for its use of dynamic computation graphs and full compatibility with Python. It allows scientists, developers, and neural network debuggers to run and test parts of the code in real time. Therefore, users do not need to wait for the entire code implementation to be completed to check if a part of the code is effective. Two main features of PyTorch are:

[0061] A-Tensor computation (similar to the NumPy library). Unlike NumPy, which uses gradient-free array-type sequence data for computation, PyToch uses gradient-type tensor-type sequence data for related operations and has powerful GPU acceleration support.

[0062] B-Automatic Differentiation: PyTorch can automatically calculate the gradients of network model parameters for backpropagation, making it convenient for users to create and train deep neural networks.

[0063] Step S3: Model training and validation.

[0064] Step S31: Model Training. Training experiments were conducted in a PyTorch environment using an RTX 4090 graphics card. The optimizer used was Adam; the learning rate was set to 1e-4; the number of iterations was 40,300; the batch size was 1; and the regularization parameter was... The value is set to 5.0. This invention sets the kernel size k to 5 and 7 to analyze the impact of different kernel sizes on the registration results. The similarity loss function used in this invention is Normalized Cross-Correlation (NCC), which is more sensitive to intensity changes in images than the usual Mean Squared Error (MSE) similarity. The formula for NCC is shown below:

[0065]

[0066] in It is a deformation field. It is a position in the image domain Ω. It is the original intensity image of the fixed image. The image after subtracting the average intensity of the local region, the deformed motion image It is the original The image after subtracting the average intensity of local regions from the original image. For example, the calculation method is as follows:

[0067]

[0068] in It revolves around the position The location of the surrounding n×n×n region, where n is set to 9.

[0069] Furthermore, the smoothing regularization term used in this invention is a diffusion regularization term, and its formula is shown below:

[0070]

[0071] in It is a deformation field. It is a displacement field, where p is a position in the image domain Ω. This is the gradient operator, used to calculate the gradient of the displacement field at a certain position p. The relationship between the deformation field and the displacement field is as follows:

[0072]

[0073] in Identity mapping, i.e. .

[0074] Step S32: Model Validation. After every 403 training iterations, the LKGDIR model's performance is validated using a validation set. The validation metric used in this invention is the Dice Similarity Coefficient (DSC). DSC is a commonly used evaluation metric for image segmentation and image registration. In registration, it measures the similarity between the segmentation of a deformed moving image and the segmentation of a stationary image. It is calculated based on the overlapping regions of the segments from the two images. The calculation method is as follows:

[0075]

[0076] Where A represents the segmentation set of the deformed moving image, B represents the segmentation set of the fixed image, |A| and |B| represent the number of elements in sets A and B, respectively, and |A ∩ B| represents the number of elements in the intersection of A and B. The value of DSC ranges from 0 to 1. The closer the value is to 1, the more similar the segmentation result of the deformed moving image is to the segmentation of the fixed image; the closer the value is to 0, the less similar the segmentation result of the deformed moving image is to the segmentation of the fixed image.

[0077] Step S4: Model Testing and Comparison Experiments. After training, the network model parameters were frozen, and then the performance of the model in registering unknown images was evaluated using a test set of 115 samples. The evaluation metrics used were: first, DSC (Distributed Synchronization Function), second, GPU memory usage, and third, the number of model parameters. To demonstrate the advancements of the proposed LKGDIR, this invention compared it with three excellent deep learning models: VoxelMorph, LKU-Net, and TransMorph. The training of these comparison models followed their official hyperparameter settings.

[0078] The publicly available dataset used in this invention is IXI. 115 three-dimensional brain MRI images from the test set were used as stationary images, and one template image was used as the moving image. The performance of different methods on three metrics—DSC, GPU memory usage, and model parameter count—was tested by simulating a real template-subject registration task. The quantitative results of the comparative experiment provided by this invention are shown in Table 1 below. The bold black text indicates the best and second-best results. The units for DSC (higher is better), Memory (lower is better), and Param. (lower is better) are %, GiB, and M, respectively. The value in parentheses after the DSC value is the standard deviation. When the large kernel k in the LKGDIR model is set to 5, GPU memory usage is reduced by 38.39%, 4.29%, and 57.64% compared to VoxelMorph, LKU-Net, and TransMorph, respectively. It is worth noting that although VoxelMorph has the fewest parameters, it requires a large amount of memory allocation during registration due to the network needing to extract a large number of high-resolution feature maps, thus resulting in higher computational costs. This invention employs an LKG module to compute a limited number of high-resolution feature maps, significantly reducing memory consumption. Furthermore, the number of parameters in this invention's model is only 37.34% and 1.66% of those in LKU-Net and TransMorph, respectively. In terms of DSC metrics, this invention's model improves upon VoxelMorph, LKU-Net, and TransMorph by 4.60%, 0.78%, and 0.77%, respectively, demonstrating that expanding the receptive field significantly improves registration accuracy. Moreover, when the large kernel k is set to 7, this invention's model further improves DSC by 4.91%, 1.08%, 1.07%, and 0.29% compared to VoxelMorph, LKU-Net, TransMorph, and LKGDIR with k set to 5, respectively. Moreover, compared to LKGDIR with k set to 5, it does not significantly increase GPU memory usage or the number of parameters. In Table 1, Affine represents affine pre-registration processing in data preprocessing.

[0079] Table 1

[0080]

[0081] Figure 5 Visualized registration results from (a) to (e) are presented for five examples in the test set using VoxelMorph, LKU-Net, TransMorph, and LKGDIR (k=7). The registration images generated by VoxelMorph, LKU-Net, and TransMorph show structural connectivity disruptions and misregistrations in multiple regions near the brainstem, as well as in the white matter and cerebral cortex, as indicated by the boxes in the figure. In contrast, the lightweight model of this invention produces higher DSC while preserving the connectivity of brain structures, thus validating that this invention, in addition to producing higher DSC, can also generate smoother deformation fields.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. An unsupervised three-dimensional deformable image registration method, characterized in that: The steps are as follows: 1) Construct a deformable image registration network model; 1.1) Obtain the dataset; obtain a public dataset including several subject images and their corresponding template images, where the subject images and the images to be registered are of the same type; preprocess the subject images and treat the preprocessed subject images as fixed images. The dataset is divided into training, validation, and test sets; the template image is treated as a moving image. ; 1.2) Model building; Based on the U-Net backbone network, a deformable image registration network model with a large kernel ghost module was designed and built; 1.3) Model training and validation; [This involves] processing motion images. Compare with each fixed image in the training set. The images are stitched together and then input into the deformable image registration network model built in step 1.2) for training. During training, the model performance is verified using a validation set, thus obtaining the deformable image registration network model. 2) Treat the image to be registered as a fixed image and the template image corresponding to the image to be registered as a moving image; stitch the moving image and the fixed image together to obtain a stitched image; 3) Input the stitched image obtained in step 2) into the deformable image registration network model constructed in step 1); the deformable image registration network model outputs the deformation field; 4) Resample the motion image from step 2) using the deformation field to obtain the deformed motion image, which is the registered image; Step 1.2) During model building, in the encoding stage, the network initially uses two consecutive convolutional layers, followed by alternating applications of multiple convolutional layers and large kernel ghost modules to extract encoder feature maps. During this process, the height, width, and depth of the feature maps gradually decrease, while the number of channels gradually increases. In the decoding stage, the network alternately uses multiple convolutional layers and transposed convolutional layers to extract decoder feature maps and gradually restores them to the original resolution, with the transposed convolutional layer responsible for the upsampling process. After each upsampling, skip connections are used to concatenate the decoder feature map at that resolution with the encoder feature map at the same resolution to prevent information loss in the encoder feature map. Finally, the network outputs a three-channel deformation field through several 3×3×3 convolutional layers. The large kernel ghost module extracts feature maps according to the following steps: a) Input features are processed through a 3×3×3 convolution operation, followed by a BN layer and a PR layer to extract intermediate feature maps, the number of which is half the number of output channels; b) Large kernel parallel deep-level convolution operation followed by BN and PR; firstly, use 1×1×1, 3×3×3 and k×k×k deep-level convolution operations to extract ghost feature maps respectively, where the large kernel k>3; Then, all ghost feature maps and intermediate feature maps are fused together by element-wise addition, and then passed through BN and PR layers to obtain fused feature maps; c) The intermediate feature map obtained in step a) and the fused feature map obtained in step b) are concatenated to obtain the final output feature map.

2. The unsupervised three-dimensional deformable image registration method according to claim 1, characterized in that: Step 1.3) During model training, for each stitched image input, the deformable image registration network model outputs a deformation field. Using deformation fields For the motion image corresponding to this training Resampling is performed to obtain the deformed motion image. That is, the image after registration during the training phase. Then, iterative backpropagation is used to minimize the loss function in order to optimize the similarity between images, as shown in the following equation; in Representing the optimal deformation field, the loss function is measured by similarity. and smoothing regularization term composition, As a parameter for regularization trade-offs.

3. The unsupervised three-dimensional deformable image registration method according to claim 1, characterized in that: Step 1.3) The similarity loss function used for model training is Normalized Cross-Correlation (NCC), as shown in the formula below. in It is a deformation field. It is a position within the image domain Ω; It is the original intensity image of the fixed image. The image after subtracting the average intensity of the local region, the deformed motion image It is the original An image minus the average intensity of local regions; The smoothing regularization term used in training is a diffusion regularization term, as shown in the formula below. in It is a deformation field. It is a displacement field, where p is a position in the image domain Ω. It is a gradient operator used to calculate the gradient of the displacement field at a certain position p; the relationship between the deformation field and the displacement field is as follows: in Identity mapping, i.e. .

4. The unsupervised three-dimensional deformable image registration method according to claim 3, characterized in that: Step 1.3) The validation metric used during model validation is the dice similarity coefficient (DSC); the calculation method is as follows: Where A represents the segmentation set of the deformed moving image, B represents the segmentation set of the fixed image, |A| and |B| represent the number of elements in sets A and B respectively, and |A ∩ B| represents the number of elements in the intersection of A and B.

Images