A multi-modal medical image registration method, system, computer device and medium

Abstract

The application provides a kind of multimodal medical image registration method, system, computer device and medium, belong to medical image processing technical field.The method is first to the pair of multimodal medical image is preprocessed, then using the modal conversion network of structure keeps and converts source modal image into the synthetic image of target modal gray scale distribution, by constructing positive and negative sample pairs and combining contrast loss function and gray consistency loss, structure keeping loss constitutes compound loss function to optimize network parameters.Finally, the synthetic image and fixed image are input into the unsupervised registration network, and the deformation field is output to obtain the registered medical image.The application adopts the above-mentioned multimodal medical image registration method, system, computer device and medium, which can effectively improve the registration accuracy, enhance the robustness and discriminability of the model to the local area, ensure the accurate alignment of the key anatomical structure, and has good clinical application prospect.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a multimodal medical image registration method, system, computer equipment, and medium. Background Technology

[0002] Medical image registration is a key technology for clinical diagnosis and treatment planning, especially in multimodal scenarios (such as CT and MRI fusion), where the goal is to achieve complementary integration of information from different imaging modalities through spatial alignment. However, multimodal images exhibit significant inconsistencies in grayscale distribution and structural representation due to differences in imaging principles. The grayscale similarity index relied upon by traditional registration methods fails when there are significant grayscale differences between modalities, leading to the optimization process easily getting trapped in local optima and causing problems such as structural mismatch and blurred boundaries, severely limiting registration accuracy and clinical applicability.

[0003] Existing multimodal medical image registration methods mainly include traditional optimization, supervised deep learning, unsupervised learning, and modality transformation strategies, but all have significant shortcomings. Traditional optimization methods rely on manually designed features, resulting in poor registration performance and high computational complexity when there are large differences in modal grayscale. Supervised deep learning methods rely on labels from real or simulated deformation fields, but real deformation fields are difficult to obtain, and there are discrepancies between simulated samples and real-world conditions, limiting the model's generalization performance. The similarity measurement of unsupervised learning methods fails when there are significant differences in grayscale, easily leading to structural mismatches. Modality transformation-based strategies suffer from problems such as structural information distortion, blurred boundaries, and complex training, especially in fine-grained anatomical regions where it is difficult to maintain structural integrity.

[0004] While existing research attempts to mitigate modal discrepancies through modality transformation or feature alignment, two major shortcomings remain: First, current modality transformation methods fail to effectively preserve structural boundaries and fine-grained texture information, resulting in transformed images lacking structural expressiveness and unable to provide accurate supervision for registration. Second, most methods neglect the registration accuracy of local regions, especially small-scale anatomical structures, leading to unstable matching in key areas and limiting their clinical application value. Therefore, improving the structural consistency and alignment accuracy of multimodal image registration requires addressing two key issues: modal grayscale difference interference and robustness of local region registration. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal medical image registration method, system, computer device, and medium that can effectively solve the problem of decreased registration accuracy caused by differences in grayscale distribution in multimodal medical image registration, ensure dual consistency of the converted image at both the grayscale and structural levels, and effectively improve the accuracy of image positioning and registration robustness.

[0006] To achieve the above objectives, the present invention provides a multimodal medical image registration method, comprising the following steps:

[0007] Step S1: Acquire paired multimodal medical images and preprocess them to obtain preprocessed source modality images and target modality images;

[0008] Step S2: Input the preprocessed source modal image into the structure-preserving modal conversion network, and output a synthesized image with the same grayscale distribution as the preprocessed target modal image;

[0009] Based on the source modality image processed by the structure-preserving modality transfer network and the preprocessed target modality image, positive sample pairs are constructed by extracting spatially corresponding image patches, while multi-source negative sample pairs are generated through cross-location selection and noise perturbation.

[0010] The contrast loss is calculated based on normalized cosine similarity, and combined with grayscale consistency loss and structure preservation loss to form a composite loss function, thereby optimizing the parameters of the structure-preserving mode transformation network.

[0011] Step S3: Input the synthesized image and the fixed image into the unsupervised registration network, extract multi-scale features through the cascaded encoder-decoder structure, and output the deformation field after fusion;

[0012] Step S4: Based on the deformation field and the synthesized image, output the registered medical image.

[0013] Preferably, in step S1, the preprocessing includes: cropping, noise reduction, and grayscale normalization.

[0014] Preferably, in step S2, the structure-preserving modality transformation network is based on the DenseUNet architecture, introducing an early transition module in the encoding path and employing a three-dimensional linear upsampling module combined with a post-convolution module in the decoding path.

[0015] Preferably, in step S2, the process of constructing positive sample pairs is as follows:

[0016] From the preprocessed source modal image and preprocessed target modal images Extract image patches at the same location and ; Represents the source modality image block. Represents the target modal image patch;

[0017] The source modal image is converted using a structure-preserving modal transfer network. Mapped to pseudo-target modal image patches Construct positive sample pairs ,in This represents a mode transition network.

[0018] Preferably, in step S2, the cross-location selection of negative samples includes selecting from the target modality image that matches the target modality image. Other image patches with mismatched positions And noise perturbation including target modal image patches Add a mixed perturbation to generate a noisy negative sample. ,in, Represents the Gaussian noise term. Represents a symbolic function. Indicates comparative loss , For target modal image blocks.

[0019] Preferably, in step S2, the composite loss function is as follows:

[0020] ;

[0021] in, Indicates compound loss. This indicates the loss of grayscale consistency. Indicates structural retention loss. Indicates a composite image. Indicates weight, This represents the gradient operator (implemented using Sobel filtering). This indicates the number of blocks in each pair of images. This represents the number of negative samples corresponding to each image patch. This indicates the image patch number that is currently a positive sample. Indicates the pseudo-target modal image patch and the first Feature similarity between negative samples across locations Indicates temperature parameter, Indicates the relationship with the first Feature similarity of noisy negative samples The normalized cosine similarity is expressed by the following formula:

[0022] ;

[0023] in, This represents the inner product of the pseudo-target modal image patch and the target modal image patch. express Norm.

[0024] Preferably, in step S3, the loss function of the unsupervised registration network... It consists of similarity loss and smoothing loss, and the calculation formula is as follows:

[0025] ;

[0026] in, The hyperparameter representing the balance between similarity and regularization terms. Represents the image domain. This represents the gradient of the deformation field. Indicates normalized cross-correlation. Indicates the deformation field Applied to images , This represents the composition of functions. This indicates a fixed image.

[0027] The present invention also provides a multimodal medical image registration system, comprising:

[0028] The data acquisition and preprocessing module is used to acquire paired multimodal medical images and perform preprocessing to obtain preprocessed source modality images and target modality images;

[0029] The structure-preserving modality conversion module, built on the DenseUNet architecture, includes an early transition module in the encoding path and a 3D linear upsampling combined with post-convolution module in the decoding path. The structure-preserving modality conversion module receives the preprocessed source modality image and outputs a synthetic image with the same grayscale distribution as the target modality.

[0030] The comparative learning optimization module includes:

[0031] Positive sample construction unit: Positive sample pairs are constructed by extracting image patches with corresponding spatial locations from the source modality image processed by the structure-preserving modality transfer network and the preprocessed target modality image;

[0032] Multi-source negative sample generation unit: generates negative sample pairs through cross-location selection and noise perturbation;

[0033] Composite loss calculation unit: Based on normalized cosine similarity, the comparison loss is calculated, and the gray-level consistency loss and structure preservation loss are combined to form a composite loss function to optimize the parameters of the mode conversion module;

[0034] The unsupervised registration module adopts a cascaded encoder-decoder structure, receives synthetic images and fixed images, and outputs a deformation field after multi-scale feature fusion;

[0035] The image registration generation module outputs registered medical images based on the deformation field and the synthesized image.

[0036] The present invention also provides a computer device including a memory and a processor, the memory being used to store instructions and the processor being used to execute the instructions to implement the multimodal medical image registration method as described above.

[0037] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multimodal medical image registration method as described above.

[0038] Therefore, the present invention employs the above-described multimodal medical image registration method, system, computer equipment, and medium, and the beneficial technical effects are as follows:

[0039] (1) In terms of network design, based on the DenseUNet backbone architecture, an early transition module is introduced in the encoding path, and a three-dimensional linear upsampling module combined with a post-convolution module is used in the decoding path to achieve direct gray-level mapping from the source modality to the target modality, effectively avoiding error accumulation caused by intermediate modalities. At the same time, the representation of anatomical regions is enhanced by the structure preservation module, and a composite loss function is proposed by combining gray-level consistency loss, structure preservation loss and ratio loss function to further enhance the ability of key anatomical structures to be preserved during modality conversion. This design ensures the dual consistency of the converted image at the gray-level and structural levels, providing a high-quality strength foundation for subsequent single-modality image registration.

[0040] (2) An image patch-level contrastive learning mechanism is introduced to fully explore the spatial correspondence between the source modality and the target modality in local regions. By constructing image patch pairs with the same spatial location as positive samples and different spatial locations as negative samples, the network can focus on spatial consistency modeling of key structures, thereby significantly improving the semantic alignment ability of local regions and cross-modal structural recognition ability. In the registration task, this perceptual characteristic based on structural alignment can effectively improve the model's localization accuracy and registration robustness for key parts (such as lesions or cerebellar regions).

[0041] (3) A multi-source negative sample strategy is introduced into the contrastive learning mechanism. By combining positional perturbation and noise perturbation, the construction forms of negative samples are enriched, and the discriminative and generalization abilities of contrastive features are improved. Among them, positional perturbation changes the spatial semantic relationship of image patches, while noise perturbation simulates the intensity uncertainty commonly found in modal changes. The two work together to enhance the network's robust modeling ability for structural alignment, effectively suppressing the interference of modal artifacts and noise on structural preservation and alignment accuracy. Attached Figure Description

[0042] Figure 1 Flowchart for modality switching and contrastive learning;

[0043] Figure 2 For the registration flowchart;

[0044] Figure 3 Comparison of qualitative visualization results for models with the same modality registration;

[0045] Figure 4 Comparison of the effects of multimodal registration models. Detailed Implementation

[0046] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0048] Example 1

[0049] This invention provides a medical image registration method, comprising the following steps:

[0050] Step S1, Data Preparation and Preprocessing:

[0051] Acquire and preprocess paired multimodal medical images that have been aligned or coarsely registered to obtain preprocessed source modality images and target modality images.

[0052] The preprocessing steps include:

[0053] Image cropping: Removes irrelevant areas from an image and retains the region of interest.

[0054] Noise reduction: Use filtering or noise reduction methods to remove noise interference.

[0055] Gray-level normalization: Linear normalization is used to map image pixel values ​​to a uniform range. This is to reduce the difference in grayscale between modes.

[0056] Step S2: Joint optimization of modality transfer and contrastive learning:

[0057] Step S21: Input the preprocessed source modal image into the structure-preserving modal conversion network and output a synthesized image with the same grayscale distribution as the preprocessed target modal image.

[0058] like Figure 1 As shown, the structure-preserving modality transformation network is based on the DenseUNet architecture. Specifically, it first constructs an end-to-end feature learning framework (DenseUNet architecture) through convolution, pooling, and upsampling operations; then, an early transition module is introduced in the encoding path to expand the receptive field through early downsampling, thereby assisting the model in establishing global grayscale mapping between modalities; and in the decoding path, a three-dimensional linear upsampling module combined with a post-convolution module is used to avoid the checkerboard artifact problem introduced by traditional deconvolution operations, improving the smoothness of edge regions and the detail restoration effect.

[0059] The input to the structure-preserving mode transfer network is the preprocessed source mode image. The output is a synthesized image whose grayscale distribution is consistent with the preprocessed target modal image. ,Right now:

[0060] ;

[0061] in, This represents a mode transition network.

[0062] Step S22: Based on the preprocessed source modal image and target modal image, extract the corresponding image patches in space to construct positive sample pairs, and at the same time generate multi-source negative sample pairs through cross-location selection and noise perturbation.

[0063] The process of constructing positive sample pairs is as follows:

[0064] From the preprocessed source modal image and preprocessed target modal images Extract image patches at the same location and ; Represents the source modality image block. Represents the target modal image patch;

[0065] The source modal image is converted using a structure-preserving modal transfer network. Mapped to pseudo-target modal image patches Construct positive sample pairs .

[0066] The process of constructing multi-source negative sample pairs is as follows:

[0067] Cross-location negative samples: Cross-location selection includes selecting from the target modality image that matches the target modality image. Other image patches with mismatched positions This is used to simulate situations where anatomical regions do not match.

[0068] Noisy negative samples: Noisy perturbations include target modality image patches Add a mixed perturbation to generate a noisy negative sample. ,in, Represents the Gaussian noise term. Represents a symbolic function. This indicates a comparative loss.

[0069] Step S23: Calculate the contrast loss based on normalized cosine similarity, and combine it with grayscale consistency loss and structure preservation loss to form a composite loss function, thereby optimizing the parameters of the structure-preserving mode conversion network.

[0070] The composite loss function is as follows:

[0071] ;

[0072] in, Indicates compound loss. This indicates the loss of grayscale consistency. Indicates structural retention loss. Indicates a composite image. Indicates weight, Represents the gradient operator, This indicates the number of blocks in each pair of images. This represents the number of negative samples corresponding to each image patch. This indicates the image patch number that is currently a positive sample. Indicates the pseudo-target modal image patch and the first Feature similarity between negative samples across locations Indicates temperature parameter, Indicates the relationship with the first Feature similarity of noisy negative samples The normalized cosine similarity is expressed by the following formula:

[0073] ;

[0074] in, This represents the inner product of the output image patch and the target modal image patch. express Norm.

[0075] Step S3: Input the synthesized image and the fixed image into the unsupervised registration network. The registration process is as follows: Figure 2 As shown.

[0076] The registration network is designed based on UNet. It uses the modality-transformed image... and fixed image The data, after being concatenated along the channel dimension, is used as input. A cascaded encoder and decoder structure extracts multi-scale features and generates a high-quality deformation field. The encoder consists of five layers, each using 32 channel feature maps, and performs downsampling at 1 / 2, 1 / 4, 1 / 8, and 1 / 16 scales sequentially. The decoder uses a symmetrical structure to progressively upsample to the original resolution, finally outputting the final deformation field. , Indicates the height of the corresponding image. Indicates the width of the corresponding image. 3 represents the depth of the corresponding image, and 3 represents the displacement component.

[0077] Loss function of unsupervised registration network It consists of similarity loss and smoothing loss, and the calculation formula is as follows:

[0078] ;

[0079] in, The hyperparameter representing the balance between similarity and regularization terms. Indicates the image domain. This represents the gradient of the deformation field. Indicates normalized cross-correlation. Indicates the deformation field Applied to images , This represents the composition of functions. This indicates a fixed image.

[0080] Step S4: Based on the deformation field and the synthesized image, output the registered medical image.

[0081] The invention will be further illustrated below with specific examples.

[0082] This example introduces the network into several typical registration frameworks and conducts systematic comparative experiments. The experiments cover monomodal registration frameworks (Voxelmorph, LapIRN) and multimodal registration frameworks (MIND, multiGranICON, MambaMorph), and use the DICE coefficient (DSC) and 95% Hausdorff distance (HD95) as evaluation metrics to comprehensively evaluate the registration performance of each method.

[0083] In the same-modal registration experiment, the modality transfer network was integrated into the Voxelmorph and LapIRN frameworks (labeled as Voxelmorph-T and LapIRN-T, respectively) and evaluated on the standard CT-MR dataset.

[0084] Table 1 summarizes the quantitative evaluation results of each method. The results show that after adding the modality transformation module, the DICE of Voxelmorph-T increased from 68.88 to 74.04, and the HD95 decreased from 3.15 to 3.02; the DICE of LapIRN-T increased from 70.81 to 75.86, and the HD95 decreased from 2.76 to 2.54. These results indicate that the proposed modality transformation network can effectively mitigate the adverse effects of cross-modal grayscale differences on the same-modal registration model and significantly improve the alignment performance of anatomical structures.

[0085] Table 1. Quantitative evaluation results of the same-modal registration task

[0086] ;

[0087] To further evaluate the structural consistency of each method from a visual perspective, Figure 3The qualitative results of each model in the CT-to-MR registration task are presented. Each group of images in the figure consists of a stationary image (MR), the original moving image (CT), and the registered image, with three types of contours superimposed: red represents anatomical structures in the stationary image, pink represents structures in the original moving image, and orange represents structures in the registered image. By observing the degree of overlap between the red and orange contours, the consistency of the registered anatomical structures can be visually assessed. It can be seen that the method incorporating the modality conversion module exhibits a higher degree of contour overlap in key areas, further validating its ability to improve registration accuracy.

[0088] In multimodal registration experiments, the proposed method was compared with existing state-of-the-art multimodal registration models (MIND, multiGranICON, MambaMorph) (see Table 2), and a unified evaluation was performed on the same CT-MR dataset. Since these methods natively possess the ability to handle modal differences, this invention did not integrate additional modules to reflect the intrinsic performance of each method. The results show that the proposed method outperforms the comparison models in both DICE and HD95 metrics, indicating that it has stronger robustness and accuracy in maintaining anatomical consistency and handling modal grayscale differences, making it suitable for multimodal registration scenarios with significant differences in grayscale distribution.

[0089] Table 2 Quantitative evaluation results of the multimodal registration task

[0090] ;

[0091] Figure 4 The qualitative visualization results of each method in the multimodal registration task are further presented. Figure 3 Similarly, the degree of overlap between the red and orange outlines in the registered image can be used to reflect the effectiveness of structural alignment. It can be seen that the method proposed in this invention exhibits better structural alignment capability in the main anatomical regions, indicating that it also has superior performance in multimodal registration tasks.

[0092] Example 2

[0093] A multimodal medical image registration system, comprising:

[0094] The data acquisition and preprocessing module is used to acquire paired multimodal medical images and perform preprocessing to obtain preprocessed source modality images and target modality images;

[0095] The structure-preserving modality conversion module, built on the DenseUNet architecture, includes an early transition module in the encoding path and a 3D linear upsampling combined with post-convolution module in the decoding path. The structure-preserving modality conversion module receives the preprocessed source modality image and outputs a synthetic image with the same grayscale distribution as the target modality.

[0096] The comparative learning optimization module includes:

[0097] Positive sample construction unit: Positive sample pairs are constructed by extracting image patches with corresponding spatial locations from the source modality image processed by the structure-preserving modality transfer network and the preprocessed target modality image;

[0098] Multi-source negative sample generation unit: generates negative sample pairs through cross-location selection and noise perturbation;

[0099] Composite loss calculation unit: Based on normalized cosine similarity, the comparison loss is calculated, and the gray-level consistency loss and structure preservation loss are combined to form a composite loss function to optimize the parameters of the mode conversion module;

[0100] The unsupervised registration module adopts a cascaded encoder-decoder structure, receives synthetic images and fixed images, and outputs a deformation field after multi-scale feature fusion;

[0101] The image registration generation module outputs registered medical images based on the deformation field and the synthesized image.

[0102] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0104] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0105] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0106] Therefore, the present invention employs the above-mentioned multimodal medical image registration method, system, computer equipment and medium, which can effectively improve registration accuracy, enhance the robustness and discriminative ability of the model to local areas, ensure accurate alignment of key anatomical structures, and has good clinical application prospects.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multimodal medical image registration method, characterized in that, Includes the following steps: Step S1: Acquire paired multimodal medical images and preprocess them to obtain preprocessed source modality images and target modality images; Step S2: Input the preprocessed source modal image into the structure-preserving modal conversion network, and output a synthesized image with the same grayscale distribution as the preprocessed target modal image; Based on the source modality image processed by the structure-preserving modality transfer network and the preprocessed target modality image, positive sample pairs are constructed by extracting spatially corresponding image patches, while multi-source negative sample pairs are generated through cross-location selection and noise perturbation. The contrast loss is calculated based on normalized cosine similarity, and combined with grayscale consistency loss and structure preservation loss to form a composite loss function, thereby optimizing the parameters of the structure-preserving mode transformation network. The structure-preserving modality transformation network is based on the DenseUNet architecture, which introduces an early transition module in the encoding path and uses a three-dimensional linear upsampling module combined with a post-convolution module in the decoding path. Step S3: Input the synthesized image and the fixed image into the unsupervised registration network, extract multi-scale features through the cascaded encoder-decoder structure, and output the deformation field after fusion; Step S4: Based on the deformation field and the synthesized image, output the registered medical image.

2. The multimodal medical image registration method according to claim 1, characterized in that, In step S1, preprocessing includes cropping, denoising, and grayscale normalization.

3. The multimodal medical image registration method according to claim 1, characterized in that, In step S2, the construction process of positive sample pairs is as follows: From the preprocessed source modality image and preprocessed target modal images Extract image patches at the same location and ; Represents the source modality image block. Represents the target modal image patch; The source modal image is converted using a structure-preserving modal transfer network. Mapped to pseudo-target modal image patches Construct positive sample pairs ,in This represents a mode transition network.

4. The multimodal medical image registration method according to claim 3, characterized in that, In step S2, the cross-location selection of negative samples includes selecting samples from the target modality image that match the target modality image. Other image patches with mismatched positions The noise perturbation includes the target modal image patch. Add a mixed perturbation to generate a noisy negative sample. ,in, Represents the Gaussian noise term. Represents a symbolic function. This indicates a comparative loss.

5. The multimodal medical image registration method according to claim 4, characterized in that, In step S2, the composite loss function is as follows: ； in, Indicates compound loss. This indicates the loss of grayscale consistency. Indicates structural retention loss. Indicates a composite image. Indicates weight, Represents the gradient operator. This indicates the number of blocks in each pair of images. This represents the number of negative samples corresponding to each image patch. This indicates the image patch number that is currently a positive sample. Indicates the pseudo-target modal image patch and the first Feature similarity between negative samples across locations Indicates temperature parameter, The normalized cosine similarity is expressed by the following formula: ； in, This represents the inner product of the pseudo-target modal image patch and the target modal image patch. express Norm.

6. The multimodal medical image registration method according to claim 5, characterized in that, In step S3, the loss function of the unsupervised registration network It consists of similarity loss and smoothing loss, and the calculation formula is as follows: ； in, The hyperparameter representing the balance between similarity and regularization terms. Indicates the image domain. This represents the gradient of the deformation field. Indicates normalized cross-correlation. Indicates the deformation field Applied to images , This represents the composition of functions. This indicates a fixed image.

7. A multimodal medical image registration system, characterized in that, For performing the multimodal medical image registration method as described in any one of claims 1-6, comprising: The data acquisition and preprocessing module is used to acquire paired multimodal medical images and perform preprocessing to obtain preprocessed source modality images and target modality images; The structure-preserving modality conversion module, built on the DenseUNet architecture, includes an early transition module in the encoding path and a 3D linear upsampling combined with post-convolution module in the decoding path. The structure-preserving modality conversion module receives the preprocessed source modality image and outputs a synthetic image with the same grayscale distribution as the target modality. The comparative learning optimization module includes: Positive sample construction unit: Positive sample pairs are constructed by extracting image patches with corresponding spatial locations from the source modality image processed by the structure-preserving modality transfer network and the preprocessed target modality image; Multi-source negative sample generation unit: generates negative sample pairs through cross-location selection and noise perturbation; Composite loss calculation unit: Based on normalized cosine similarity, the comparison loss is calculated, and the gray-level consistency loss and structure preservation loss are combined to form a composite loss function to optimize the parameters of the mode conversion module; The unsupervised registration module adopts a cascaded encoder-decoder structure, receives synthetic images and fixed images, and outputs a deformation field after multi-scale feature fusion; The image registration generation module outputs registered medical images based on the deformation field and the synthesized image.

8. A computer device, characterized in that, It includes a memory and a processor, the memory being used to store instructions, and the processor being used to execute the instructions to implement the multimodal medical image registration method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multimodal medical image registration method as described in any one of claims 1-6.

Images