Prostate MRI registration method based on dense convolution and gated feature extraction network
By adopting dense convolution and gating feature extraction networks in medical image registration, the problem of insufficient comprehensive feature extraction and information loss in multimodal prostate MRI image registration is solved, and high-precision image registration and efficiency improvement are achieved.
Patent Information
- Application Number
- CN202510041537.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-10
- Publication Date
- 2025-05-09
AI Technical Summary
When processing multimodal prostate MRI images, existing medical image registration methods have problems such as insufficient feature extraction, loss of information and blurred image edges, resulting in poor registration results.
Using a method based on dense convolution and gated feature extraction network, personalized feature extraction is performed on prostate MRI-T2WI and DWI images, combining dense convolution module and Swin Transformer module to enhance feature flow and multiplexing, and high-precision image registration is achieved through spatial transformation network.
It improves the accuracy and efficiency of multimodal medical image registration, effectively solves the problems of blurred image edge information and loss of position information, and significantly improves the effect of prostate MRI image registration.
Smart Images
Figure CN119963405A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical information technology, and in particular to a prostate MRI registration method based on a dense convolution and gated feature extraction network. Background Art
[0002] Medical image registration is a key step to improve the accuracy of diagnosis and treatment. However, due to the complexity of prostate structure and the uniqueness of imaging mode, there are huge differences in image resolution, signal intensity and tissue contrast, making multimodal registration a core challenge in prostate cancer image analysis. In order to meet these challenges and improve the accuracy and efficiency of prostate disease diagnosis, researchers have begun to explore multimodal image registration technology. By aligning medical images from different modalities and integrating the information of each modality, the complementarity and enhancement between images can be achieved. This method can more comprehensively and accurately reflect the patient's pathological status and help doctors obtain more detailed anatomical and pathological information. Although deep learning methods have achieved significant accuracy improvements, these methods have not taken into account the extraction of multi-scale features, the loss of information during upsampling and downsampling, and the blurring of image edge information and the loss of position information caused by jump connections.
[0003] At present, medical image registration methods can be mainly divided into three categories: feature-based registration methods, pixel-based registration methods and deep learning-based registration methods. Feature-based registration methods achieve registration by extracting local feature points and calculating the transformation matrix between images, which is suitable for images with simple structures. However, they often show great limitations when processing medical images with low contrast, large noise or complex structures. Pixel-based registration methods achieve image registration by directly optimizing image similarity metrics, which is especially suitable for image registration with large modality differences. However, in practical applications, they face problems such as high computational complexity and high computational overhead, especially when processing high-resolution or large-scale data, computational efficiency becomes its main bottleneck. Deep learning-based registration methods automatically learn image features from data and directly output the transformation parameters required for registration through an end-to-end training framework, which greatly improves the accuracy of registration. However, deep learning-based registration methods generally input fixed images and moving images into the network together, which does not take into account the need for personalized feature extraction of different modality images, the loss of information during upsampling and downsampling, and the blurring of image edge information and the loss of position information caused by jump connections. Summary of the invention
[0004] In view of this, the object of the present invention is to provide a prostate MRI registration method based on dense convolution and gated feature extraction network to improve the above problems.
[0005] The present invention provides a prostate MRI registration method based on dense convolution and gated feature extraction network, which comprises:
[0006] Acquire a prostate MRI image sequence of each target user, wherein the MRI image sequence includes an MRI-T2WI image as a fixed image and a DWI image as a moving image;
[0007] Preprocessing the fixed image and the moving image;
[0008] Inputting the preprocessed fixed image and moving image into a gated residual fusion module, and extracting coarse features of the fixed image and moving image respectively through different convolutional networks;
[0009] The extracted coarse features and their corresponding source images are sliced and cross-joined and then input into the registration network to output the deformation field; wherein the registration network includes a dense convolution module and a Swin Transformer module, the dense convolution module is used to enhance the flow and reuse of features, and the Swin Transformer module is used to perform global information interaction;
[0010] According to the deformation field, a registered image is obtained through a spatial transformation network.
[0011] Preferably, the preprocessing of the fixed image and the moving image comprises:
[0012] De-noising the fixed image and the moving image, and removing image noise by using a median filter; resampling the de-noised fixed image and the moving image to unify the dimension and size of the image;
[0013] Performing affine transformation on the resampled fixed image and the moving image to obtain an affine-aligned moving image, and using the affine-aligned moving image as a new moving image;
[0014] The fixed image and the new moving image are grayscale normalized to uniformly map the image grayscale range to the [0, 1] interval.
[0015] Preferably, during resampling, the image is resampled to a uniform dimension and size by using a resampling image filter of the ITK library and cubic spline interpolation.
[0016] Preferably, the gated residual fusion module includes two parallel convolutional networks, wherein the convolutional network corresponding to the fixed image includes three convolutional layers, and the convolutional kernel size is 3x3;
[0017] The convolutional network corresponding to the moving image has a dual-channel structure, one channel extracts image edge information, and the other channel extracts image texture and low-level features.
[0018] Preferably, the dense convolution module connects the output of a convolution layer to each subsequent convolution layer by adding a densely connected convolution layer in the jump connection process of the registration network, thereby realizing the reuse and fusion of features between different layers and enhancing the feature extraction and expression capabilities of the network.
[0019] Preferably, the Swin Transformer module uses a shifted window method to perform self-attention calculations, model image features at different scales, capture global contextual information, and improve the registration network's ability to handle complex deformations.
[0020] Preferably, it also includes:
[0021] Obtain the prostate contour label corresponding to the prostate MRI image;
[0022] The MRI image and the contour label are input into a segmentation model for training to obtain a prostate segmentation result; the segmentation result is used as auxiliary supervision information of the registration network to guide the training process of the registration network.
[0023] Preferably, the method further comprises:
[0024] Construct the loss function of the registration network, which includes similarity loss and regularization loss;
[0025] Among them, the similarity loss uses the normalized mutual information NMI to measure the similarity between the registered image and the fixed image, and the regularization loss uses the L2 norm to constrain the smoothness of the deformation field;
[0026] By jointly optimizing the two loss functions, the optimal registration parameters are obtained.
[0027] Preferably, the method further comprises: using a Dice similarity coefficient and a non-positive Jacobian determinant to evaluate the registration result.
[0028] The present invention adopts an innovative network structure to solve the challenge of multimodal medical image registration for the scenario of MRI-T2WI and DWI image sequence registration. Personalized feature extraction is performed on fixed images and moving images through a gated residual fusion module, and then the improved Swin Transformer network is used to perform feature transformation and fusion, and finally the registration deformation field is output. The present invention introduces dense convolution in the jump connection of the Transformer to maximize the flow and reuse of feature information. Through a spatial transformation network and optimization strategy, the present invention achieves high-precision prostate MRI image registration. This method effectively improves the accuracy and efficiency of multimodal medical image registration, and provides important technical support for the diagnosis and treatment of prostate diseases. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 A flowchart of a prostate MRI registration method based on dense convolution and gated feature extraction network is provided in an embodiment of the present invention.
[0030] Figure 2 An architectural diagram of a gated residual fusion module according to an embodiment of the present invention.
[0031] Figure 3 An architectural diagram of a dense convolution module according to an embodiment of the present invention.
[0032] Figure 4 Figure 1. Schematic diagram of the prostate MRI registration method based on dense convolution and gated feature extraction network.
[0033] Figure 5 Visual comparison of prostate MRI registration method based on dense convolution and gated feature extraction network and other methods on two datasets. DETAILED DESCRIPTION
[0034] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0035] See also Figure 1 The first embodiment of the present invention provides a prostate MRI registration method based on dense convolution and gated feature extraction network, which can be performed by a prostate MRI registration device based on dense convolution and gated feature extraction network, and in particular, by one or more processors in the configuration device to implement the following method:
[0036] S101 , acquiring a prostate MRI image sequence of each target user, wherein the MRI image sequence includes an MRI-T2WI image as a fixed image and a DWI image as a moving image.
[0037] In this embodiment, for example, a prostate MRI image sequence in the .nii.gz format can be obtained from the prostate MRI data set generated by the National Cancer Institute of the United States between 2008 and 2010 in the Haikou People's Hospital and the Cancer Imaging Archive, respectively. The original prostate MRI image is an image of an MR-T2WI and DWI sequence, in which MRI-T2WI is used as a fixed image and DWI is used as a moving image.
[0038] Of course, it should be noted that the prostate MRI image sequence can also be obtained from other databases, and the present invention does not make any specific limitation.
[0039] S102: pre-processing the fixed image and the moving image.
[0040] Specifically, in this embodiment, the preprocessing includes the following steps:
[0041] Step 1: First, use a median filter to denoise all prostate MRI images. The formulas for median filtering are:
[0042] g(x, y)=meed{f(i-1:i+1, j-1:j+1), (i, j∈w)} (1)
[0043] Where g(x, y) is the denoised image, f(x, y) is the original image, meed{} means calculating the median of a set of numbers, and w represents all pixel domains of an image. The processed image maintains the original format of .nii.gz image.
[0044] Step 2: Use the ITK library's resampling image filter and use cubic spline function interpolation to resample the image to a uniform dimension and size. The cubic spline function interpolation formula is:
[0045] y=a i +b i x+c i x 2 +d i x 3 (2)
[0046] Where y is the cubic spline function, x represents each pixel interval, a i , b i , c i , d i These are the four unknowns in each pixel cell. Finally, by establishing the equation and solving the unknowns, we can get the pixel value of the target pixel.
[0047] Step 3: Arrange the MRI-T2WI and DWI image sequences of the same patient into several registration pairs in one-to-one correspondence;
[0048] First, the registration pair was loaded into Slicer (version 5.2.1), with MRI-T2WI as the fixed image and DWI as the moving image for affine transformation;
[0049] Then, the affine-aligned moving image obtained by the affine transformation is taken as a new moving image and saved in one-to-one correspondence with the fixed image.
[0050] In this embodiment, an affine transformation algorithm based on mutual information is used to optimize affine transformation parameters by minimizing the mutual information value between a fixed image and a moving image, thereby obtaining a new moving image.
[0051] Step 4: Unify the grayscale range of the image. Use the linear mapping formula to unify the grayscale range of the fixed image and the moving image to the [0, 1] interval. The linear mapping formula is as follows:
[0052]
[0053] Among them, I is the gray value in the original image, min is the minimum gray value of the image, max is the maximum gray value of the image, and I norm It is the normalized grayscale value after mapping, ranging from [0, 1].
[0054] Step 5: Experienced urologists and radiologists annotate the prostate contours on the normalized prostate MRI-T2WI and DWI image sequences;
[0055] First, the fixed and moving images were loaded into Slicer (version 5.2.1), and the urologists and radiologists jointly annotated the prostate contours;
[0056] Then, the annotated prostate contour is processed and the pixel value is set to 1;
[0057] Finally, after ensuring that all annotations cover the entire prostate gland, save the image in the same format as the original image, nii.gz;
[0058] S103, inputting the preprocessed fixed image and moving image into a gated residual fusion module, and extracting coarse features of the fixed image and moving image respectively through different convolutional networks.
[0059] In this embodiment, according to the resolution difference between the fixed image and the moving image, a gated residual fusion module is used to perform personalized coarse feature extraction respectively.
[0060] Among them, Figure 2 As shown in the figure, for high-resolution fixed images, a small convolutional network is used, such as 3 convolutional layers, and each convolutional kernel is 3x3. For low-resolution moving images, a two-channel convolutional network is used for feature extraction. One channel is dedicated to extracting edge information of the image, and the other channel is dedicated to extracting texture and other low-level features.
[0061] Exemplarily, in the gated residual fusion module, for fixed images, a small convolutional network containing 3 convolutional layers is used, each with a convolution kernel size of 3x3, a stride of 1, and a padding of 1. For moving images, a dual-channel convolutional network is used, the first channel uses the Sobel operator to extract edge information, and the second channel uses the Gabor filter to extract texture features.
[0062] S104, slicing and cross-stitching the extracted coarse features with their corresponding source images, and then inputting them into the registration network to output the deformation field; wherein the registration network includes a dense convolution module and a SwinTransformer module, the dense convolution module is used to enhance the flow and reuse of features, and the SwinTransformer is used to perform global information interaction.
[0063] In this embodiment, the extracted fixed image, coarse features of the moving image and their corresponding source images are dynamically fused through a gating mechanism, that is, the coarse features of the fixed image are sliced and crossed with the fixed image, and the coarse features of the moving image are sliced and crossed with the moving image, and then they are spliced and input into the registration network.
[0064] In this embodiment, the weights of certain features are selectively strengthened and irrelevant features are suppressed according to the learning status of the network. The residual learning mechanism is used to fine-tune the existing features. The network can effectively learn subtle transformations between images while retaining key information in the original image. The fused features are input into the SWinTransformer deep learning network for prostate MRI medical image registration. The registration network automatically learns image features from the data and directly outputs the transformation parameters required for registration through an end-to-end training framework. Dense convolution is added in the jump connection process of the registration network, such as Figure 3 As shown in the figure, the flow and reuse of feature information is maximized. Each layer can access the feature maps of all previous layers and reuse features with more information. Dense convolution solves the gradient vanishing problem in deep networks and improves the expressiveness of the model. The network can better capture information at different levels and provide stronger robustness and generalization capabilities.
[0065] Exemplarily, the gating mechanism controls the feature weights of the two channels between 0 and 1 through the activation function, and dynamically adjusts the feature fusion ratio. Residual learning adds the input features to the convolution output through identity mapping to achieve feature fine-tuning. The fused features are input into a deep learning network containing 4 Swin Transformer blocks. Each block contains a multi-head self-attention mechanism and a feedforward neural network, and global alignment is achieved through feature interactions of different scales. Three dense convolution blocks are added to the jump connection, each containing 4 convolution layers, which are concatenated with the feature maps of all previous layers as input to maximize feature reuse.
[0066] S105, obtaining a registered image through a spatial transformation network according to the deformation field.
[0067] In this embodiment, a registered deformation field is obtained from the registration network, and then the deformation field is applied to the moving image, and the registered image is obtained through the spatial transformation network.
[0068] The embodiment of the present invention adopts an innovative network structure to solve the challenge of multimodal medical image registration for the scenario of MRI-T2WI and DWI image sequence registration. Personalized feature extraction is performed on fixed images and moving images through a gated residual fusion module, and then the improved Swin Transformer network is used to perform feature transformation and fusion, and finally the registration deformation field is output. The present invention introduces dense convolution in the jump connection of the Transformer to maximize the flow and reuse of feature information. Through the spatial transformation network and optimization strategy, the present invention achieves high-precision prostate MRI image registration. This method effectively improves the accuracy and efficiency of multimodal medical image registration, and provides important technical support for the diagnosis and treatment of prostate diseases.
[0069] Some preferred embodiments of the present invention are further described below:
[0070] Preferably, it also includes:
[0071] Construct the loss function of the registration network, which includes similarity loss and regularization loss;
[0072] Among them, the similarity loss uses the normalized mutual information NMI to measure the similarity between the registered image and the fixed image, and the regularization loss uses the L2 norm to constrain the smoothness of the deformation field;
[0073] By jointly optimizing the two loss functions, the optimal registration parameters are obtained.
[0074] In this embodiment, if Figure 4 As shown, the registration network needs to be constructed and trained first, specifically:
[0075] First, the construction of the deep learning model and the optimization of the network are established. The adaptability and generalization ability of the benchmark model SwinTransformer network are limited by single-modality images, and it is easy to lose key detail information, especially when there are obvious differences between different imaging modalities, thereby limiting the registration accuracy. In view of the above limitations, this embodiment improves the network to make it more in line with the characteristics of the prostate multimodal dataset.
[0076] Then, this embodiment uses the preprocessed prostate dataset and divides the fixed images and moving images one by one to generate training sets, validation sets, and test sets. The distribution ratio of the dataset follows the principle of 3:1:1. Specifically, for each experiment, the dataset is randomly allocated to ensure the representativeness and diversity of the data. This study performed a total of five independent experiments to reduce the impact of randomness on the experimental results. Finally, the results of all experiments are reported as average values to provide more accurate and reliable research conclusions;
[0077] Secondly, for the construction of the model framework, it is based on the Pytorch-1.9 framework, and the interpreter is python3.8. The network architecture is based on the Swin Transformer, with a gated residual fusion module added. Coarse features are extracted through two independent convolutional networks. For high-resolution fixed images, a small convolutional network is used, with only 3 convolutional layers, and each convolutional kernel is 3x3. For low-resolution moving images, a dual-channel convolutional network is used, one channel is dedicated to extracting edge information of the image, and the other channel is dedicated to extracting texture and other low-level features. This ensures the independent extraction of fixed image and moving image features, while providing rich multimodal information for subsequent registration steps.
[0078] Next, this embodiment adds dense convolution to the jump connection process of the network architecture. Compared with the simple layer-by-layer connection in the traditional network model, dense convolution effectively promotes the fusion and transmission of features at different levels. This structure not only improves the model's ability to capture detailed features, but also optimizes the gradient flow and alleviates the gradient vanishing problem in deep networks, thereby accelerating the training and convergence of the network. In addition, dense convolution enhances the diversity of feature maps, enabling the model to better handle blur, deformation, and cross-modal image registration tasks.
[0079] Finally, the training set is used to train the model, and the validation set is used to select the optimal model. In the forward operation of the neural network, the loss function design of the network training is derived from the energy function of the traditional image registration. Its structure is usually divided into two parts. The first part is used to evaluate the similarity between the deformed motion image and the fixed image to ensure the image alignment effect; the second part introduces a regularization term to constrain the deformation field, thereby improving the smoothness and rationality of the transformation. The loss function can be expressed as:
[0080]
[0081] in represents the similarity measure, represents the regularization term, and λ represents the regularization weight, which controls the importance of the regularization term in the overall loss.
[0082] In the selection of similarity measurement, the common and widely used mean square error (MSE) method is adopted. MSE is based on calculating the square average of pixel intensity differences, which can intuitively measure the accuracy of image registration. It can be expressed as:
[0083]
[0084] where p represents the spatial position of the voxel and Ω represents the domain of the image.
[0085] In order to ensure the physical rationality and smoothness of the deformation field, the deformation field generated only by similarity measurement may not be smooth or realistic. It is also necessary to introduce a regularizer To suppress excessive deformation, thereby improving the smoothness and interpretability of the registration results. The model selects the diffusion regularizer for experiment, which can be expressed as:
[0086]
[0087] in represents the gradient of the deformation field at position p.
[0088] For a further understanding of the present invention, a practical example is used below to illustrate the application of the present invention.
[0089] 1) Experimental Dataset
[0090] The original datasets are the private medical prostate MRI image dataset (LH-MRI) provided by Haikou People's Hospital and the public medical image dataset (TCIA-MRI) of prostate MRI generated by the National Cancer Institute of the United States from 2008 to 2010. These source images are preprocessed to keep the size and parameters consistent, all of which are 160×160×96.
[0091] 2) Evaluation indicators
[0092] Two objective evaluation indicators are used to quantitatively evaluate the performance of the proposed method, namely Dice Similarity Coefficient, Non-positive Jacobian Determinant (% of |J φ |≤0). A higher Dice similarity coefficient indicates a more accurate registration result. A lower non-positive Jacobian value indicates a higher smoothness and better performance.
[0093] 1. Dice similarity coefficient
[0094] The DSC similarity coefficient (DSC) is an important evaluation index in medical image registration, which is used to quantitatively measure the accuracy of the registration results. DSC evaluates the registration accuracy by comparing the overlap between the registration result and the fixed image, and its value range is between 0 (completely dissimilar) and 1 (completely identical). The higher the DSC value, the higher the degree of consistency between the registration result and the actual structure, indicating that the registration effect is better. Its calculation formula is as follows:
[0095]
[0096] Among them I deform represents the image after registration deformation, I fixed Represents a fixed image. DSC can effectively reflect the consistency of overlapping areas of images and is an important basis for measuring algorithm performance.
[0097] 2. Non-positive Jacobian
[0098] In medical image registration, the Jacobian matrix is often used to describe the local change characteristics of image transformation. By calculating the determinant of the Jacobian matrix, the geometric change characteristics of the local area during the image deformation process can be evaluated. When the determinant value of the Jacobian matrix is zero or negative, it is called a non-positive Jacobian determinant, which usually indicates that the image voxels may overlap, flip or distort in this area. This phenomenon may cause geometric anomalies in the registration results. Therefore, the non-positive Jacobian determinant is an important evaluation indicator for monitoring the rationality of image registration and deformation processing. The formula is:
[0099]
[0100] where det(J) is the determinant of the Jacobian matrix and x′ is the new position of the deformation field mapped to the voxel position x.
[0101] N non-positive is the number of non-positive Jacobian voxels. II in the formula is the indicator function, which takes the value of 1 when the condition is met, otherwise it takes the value of 0. The indicator function is used to determine whether the Jacobian determinant is non-positive and count the corresponding voxels to reflect the abnormal areas that may appear in the image during the deformation process. Percentage of NPJ is the percentage of non-positive Jacobian, N total It represents the total prime number, and calculates the ratio of non-normal Jacobian voxels to the total voxels, reflecting the ratio of abnormal areas such as folding and distortion that may appear in the deformation field. The smaller the Percentage of NPJ, the higher the deformation quality, indicating that the deformation field is better in maintaining the rationality of the image structure.
[0102] 3) Experimental design and result analysis
[0103] This experiment will use the proposed prostate MRI registration method based on dense convolution and gated feature extraction network to perform registration experiments on two data sets, and also select several different mainstream registration methods for comparative experiments.
[0104] Figure 5 The visual registration results and the corresponding deformation fields are shown in . These figures clearly demonstrate the ability of the proposed method to achieve high-precision image registration. The registrations shown in the figures highlight the accuracy of the proposed method in handling multimodal image registration tasks. In addition, the deformation fields shown in these figures show minimal distortion, demonstrating the smoothness and stability of the transformations generated by our method.
[0105] As shown in Table 1, the method proposed in this embodiment outperforms several other excellent methods in terms of evaluation indicators. Specifically, the Dice similarity coefficient on the LH-MRI and TCIA-MRI datasets increased by more than 1.9% and 3.2%, respectively, and the appearance of non-positive Jacobian determinants in the deformation field was significantly reduced. Taking these indicators into consideration, the method proposed in this article can further improve the registration accuracy in the multimodal image registration task by effectively processing the feature information of different modalities compared to other methods.
[0106] Table 1
[0107]
[0108] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A prostate MRI registration method based on dense convolution and gated feature extraction network, characterized in that: include: Acquire a prostate MRI image sequence of each target user, wherein the MRI image sequence includes an MRI-T2WI image as a fixed image and a DWI image as a moving image; Preprocessing the fixed image and the moving image; Inputting the preprocessed fixed image and moving image into a gated residual fusion module, and extracting coarse features of the fixed image and moving image respectively through different convolutional networks; The extracted coarse features and their corresponding source images are sliced and cross-joined and then input into the registration network to output the deformation field; wherein the registration network includes a dense convolution module and a Swin Transformer module, the dense convolution module is used to enhance the flow and reuse of features, and the Swin Transformer module is used to perform global information interaction; According to the deformation field, a registered image is obtained through a spatial transformation network.
2. The prostate MRI registration method based on dense convolution and gated feature extraction network as claimed in claim 1, characterized in that: The preprocessing of the fixed image and the moving image comprises: Performing denoising processing on the fixed image and the moving image, and using a median filter to remove image noise; Resample the denoised fixed and moving images to unify the dimensions and sizes of the images; Performing affine transformation on the resampled fixed image and the moving image to obtain an affine-aligned moving image, and using the affine-aligned moving image as a new moving image; The fixed image and the new moving image are grayscale normalized to uniformly map the image grayscale range to the [0, 1] interval.
3. The prostate MRI registration method based on dense convolution and gated feature extraction network as claimed in claim 2, characterized in that: During resampling, the image is resampled to a uniform dimension and size using the ITK library's resample image filter and cubic spline interpolation.
4. The prostate MRI registration method based on dense convolution and gated feature extraction network according to claim 1, characterized in that: The gated residual fusion module includes two parallel convolutional networks, wherein the convolutional network corresponding to the fixed image includes three convolutional layers, and the convolutional kernel size is 3x3; The convolutional network corresponding to the moving image has a dual-channel structure, one channel extracts image edge information, and the other channel extracts image texture and low-level features.
5. The prostate MRI registration method based on dense convolution and gated feature extraction network as claimed in claim 1, characterized in that: The dense convolution module adds a densely connected convolution layer in the skip connection process of the registration network, connects the output of a convolution layer to each subsequent convolution layer, realizes the reuse and fusion of features between different layers, and enhances the feature extraction and expression capabilities of the network.
6. The prostate MRI registration method based on dense convolution and gated feature extraction network according to claim 1, characterized in that: The Swin Transformer module uses a shifted window method to perform self-attention calculations, model image features at different scales, capture global contextual information, and improve the registration network's ability to handle complex deformations.
7. The prostate MRI registration method based on dense convolution and gated feature extraction network as claimed in claim 2, characterized in that: Preprocessing also includes: Obtain the prostate contour label corresponding to the prostate MRI image; The MRI image and the contour label are input into a segmentation model for training to obtain a prostate segmentation result; the segmentation result is used as auxiliary supervision information of the registration network to guide the training process of the registration network.
8. The prostate MRI registration method based on dense convolution and gated feature extraction network according to claim 1, characterized in that: The method further comprises: Construct the loss function of the registration network, which includes similarity loss and regularization loss; Among them, the similarity loss uses the normalized mutual information NMI to measure the similarity between the registered image and the fixed image, and the regularization loss uses the L2 norm to constrain the smoothness of the deformation field; By jointly optimizing the two loss functions, the optimal registration parameters are obtained.
9. The prostate MRI registration method based on dense convolution and gated feature extraction network according to claim 1, characterized in that: Also includes: The Dice similarity coefficient and non-positive Jacobian were used to evaluate the registration results.