Self-adaptive refined medical image registration method based on geometric blur

Abstract

The invention discloses an adaptive refined medical image registration method based on geometric blur, which belongs to the technical field of medical image processing and medical artificial intelligence, and comprises the following steps: 1, inputting a fixed image and a moving image, and extracting multi-scale geometric features of the medical image; 2, performing feature transformation and fusion on the multi-scale geometric features to obtain joint features; 3, carrying out weighted adjustment on the joint features through a fuzzy weighting module; 4, inputting the combined features into a deformation estimation module, refining the scale of the combined features through a scale-by-scale adaptive refining module, and then returning the deformation field at the moment to the step 2 for iterative operation until the deformation estimation of the finest scale is completed, so as to obtain a high-resolution deformation field; 5, processing the originally input moving image by using a high-resolution deformation field to obtain a deformation image, and performing accurate registration on the deformation image and the original fixed image; according to the method, the multiple modules are connected in sequence, and therefore accurate medical image registration from coarse to fine is achieved.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and medical artificial intelligence, and in particular to an adaptive thinning medical image registration method based on geometric blur. Background Technology

[0002] In recent years, deep learning has made significant progress in the field of medical image registration. End-to-end registration models based on convolutional neural networks (CNNs) have significantly improved inference speed and have become the mainstream research direction in medical image registration in recent years.

[0003] However, existing medical image registration techniques still suffer from several problems. First, traditional convolutional neural networks struggle to effectively represent the inherent spatial orientation, rotational relationships, and high-order geometric structures in medical images, leading to discontinuities, topological disruptions, or local instability in the predicted deformation field in complex anatomical regions. Second, medical images often contain noise, artifacts, and low-contrast structures. Existing methods lack sufficient feature representation capabilities in blurred regions, easily causing registration error accumulation and exhibiting poor robustness. Furthermore, existing registration networks often rely on single inference or a fixed number of iterations, failing to adaptively adjust the optimization process based on local structural differences, and thus struggling to balance local detail alignment with overall deformation stability. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive thinning medical image registration method based on geometric fuzziness, which is completed collaboratively by a geometric algebraic encoding module, a feature fusion module, a fuzzy weighting module, a deformation estimation module, and a scale-wise adaptive thinning module. These modules are connected sequentially, with the output of one module serving as the input of the next, and the deformation field is progressively passed and updated across a multi-scale structure, thereby achieving accurate medical image registration from coarse to fine.

[0005] To achieve the above objectives, this invention provides an adaptive thinning medical image registration method based on geometric blur, comprising the following steps: S1. Input a fixed image and a moving image, and extract multi-scale geometric features of the medical image through a geometric algebraic convolution module; S2. Perform feature transformation and fusion on the multi-scale geometric features extracted in S1 to obtain joint features; S3. The joint features are weighted and adjusted using the fuzzy weighting module; S4. Input the weighted joint features into the deformation estimation module, and refine the scale of the joint features through the scale-by-scale adaptive refinement module. Then return the deformation field at this time to S2 for iterative operation until the deformation estimation at the finest scale is completed, and obtain the high-resolution deformation field. S5. Process the original input moving image using a high-resolution deformation field to obtain a deformed image, and then accurately register it with the original fixed image.

[0006] Preferably, the process of S1 is as follows: S11. Obtain the fixed image and the moving image to be registered, and input both the fixed image and the moving image into the geometric algebraic convolution module; S12, the geometric algebraic convolution module inputs a fixed image into the geometric algebraic encoding unit; S13. The geometric algebra coding unit performs feature grouping on the feature map of the fixed image based on the multi-vector representation rules of geometric algebra, and maps each group of sub-features into the multi-vector space to form an initial multi-vector feature containing multi-order geometric components. S14. Based on the definition of geometric product, the input features and the learnable geometric algebraic convolution kernel are convolved in the multi-vector domain to extract the initial multi-vector features of the fixed image, including spatial features containing orientation, rotation relationship and high-order geometric structure information. S15. Then, the encoder uses a pyramid structure with five scale levels to extract multi-scale geometric features of spatial features from coarse to fine, and then recombines them into Clifford structured features. S16. Normalize the structured features and apply a non-linear activation function to obtain the GAConv output features, which in turn yields the fixed features of the fixed image. S17. Apply the same features as S12-S16 to the moving image to obtain the moving features of the moving image.

[0007] Preferably, the process of S2 is as follows: S21. Determine the current scale level, input the initial deformation field and decoded features; S22. Use the deformation field to perform a spatial transformation on the moving image features, map the moving image features to a fixed image coordinate system, and obtain the aligned moving image features. S23. The moving image features, fixed features of the fixed image, and decoded features obtained in S22 are concatenated in the channel dimension and input into the feature fusion module based on Swin Transformer to model the long-range correlation across images and output the fused features at this scale.

[0008] Preferably, the process of S3 is as follows: S31. Input the fusion features obtained in S2 into the fuzzy weighting module; S32. The fuzzy weighting module constructs a fuzzy membership degree based on the relevant indicators of the features to measure the degree to which each feature point belongs to the "reliable structure" and "fuzzy region". S33. The fuzzy weighting module adjusts the weights of feature channels or spatial locations based on membership degrees.

[0009] Preferably, the fuzzy membership degree in S32 is adaptively generated based on the relative relationship between features through learnable fuzzy mapping rules.

[0010] Preferably, the weighted adjustment in S33 is as follows: S331. Increase the feature weights of regions with high membership and clear structure to enhance the sensitivity of deformation estimation to important structures; S332. Reduce the weight of features with low membership or located in fuzzy or noisy regions to reduce the interference of these regions on deformation prediction.

[0011] Preferably, the process of S4 is as follows: S41. Input the fuzzy weighted features into the deformation estimation module to predict the deformation increment at the current scale. Then, superimpose the deformation increment with the initial deformation field to obtain the updated deformation field. S42. Introduce a scale-wise adaptive thinning mechanism to perform spatial transformation on the moving image based on the updated deformation field at this time to obtain the deformation image; S43. Measure the similarity between the deformed image obtained in S42 and the fixed image to obtain the similarity at this time; S44. Set a similarity enhancement threshold, take the updated deformation field obtained in S41 as the new initial deformation field and return to S2 using the same decoding features for iteration until the similarity enhancement is lower than the preset threshold, and obtain the optimal deformation field of the current layer of the multi-layer decoder. S45. Upsample the optimal deformation field and the fused features at this time to obtain a new initial deformation field and decoding features. Enter the next layer of the decoder and return to S2 using the new initial deformation field and decoding features. After 5 iterations of the multi-layer decoder, the finest scale deformation field, i.e., the high-resolution deformation field, is obtained.

[0012] Preferably, the process of S5 is as follows: S51. Use a high-resolution deformation field to perform a spatial transformation on the initial moving image to obtain the deformed image at this time; S52. Use the deformed image and the fixed image at this time to perform a similarity test to obtain the registration result. Combine the loss function to verify the high-resolution deformation field, taking into account both the registration accuracy and the geometric rationality of the deformation field, and complete the accurate registration of the moving image to the fixed image.

[0013] Preferably, the loss function in S52 is as follows: ; ; ; in, Indicates the total loss. This indicates the similarity between the deformed image obtained in S42 and the fixed image. This represents the smoothness and reversibility constraints imposed on the predicted deformation field. Indicates the trade-off coefficient. and These represent moving and stationary images, respectively. This represents the predicted deformation field. This represents the deformed image after the moving image has undergone deformation field transformation. Represents the deformation field In position Spatial gradient, Represents the balance coefficient. This represents a penalty term based on the Jacobian determinant.

[0014] Preferably, processes S2, S3, and S4 are all performed in a multi-layer decoder.

[0015] Therefore, this invention employs the above-described adaptive thinning medical image registration method based on geometric fuzziness, which is collaboratively completed by a geometric algebraic encoding module, a feature fusion module, a fuzzy weighting module, a deformation estimation module, and a scale-by-scale adaptive thinning module. These modules are connected sequentially, with the output of one module serving as the input of the next, and the deformation field is progressively transmitted and updated across the multi-scale structure, thereby achieving precise medical image registration from coarse to fine.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the adaptive thinning medical image registration method based on geometric fuzzing of the present invention. Detailed Implementation

[0018] Example To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0022] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] like Figure 1 As shown, this invention provides an adaptive thinning medical image registration method based on geometric fuzzing. This method enhances spatial geometric modeling capabilities, improves robustness in low-quality regions, and supports dynamic thinning during the inference stage, thereby improving the stability and accuracy of deformation field prediction in complex scenes. It includes a multi-scale geometric algebraic coding module, a feature fusion and fuzzy weighting module, a deformation estimation module, and a scale-wise adaptive thinning module. A high-resolution deformation field is generated step-by-step through a multi-scale decoding structure. The method comprises the following steps: S1. Input a fixed image and a moving image, and extract multi-scale geometric features of the medical image through a geometric algebraic convolution module; S11. Obtain the fixed image and the moving image to be registered, and input both the fixed image and the moving image into the geometric algebraic convolution module; S12, the geometric algebraic convolution module inputs a fixed image into the geometric algebraic encoding unit; S13. The geometric algebra coding unit performs feature grouping on the feature map of the fixed image based on the multi-vector representation rules of geometric algebra, and maps each group of sub-features into the multi-vector space to form an initial multi-vector feature containing multi-order geometric components. S14. Based on the definition of geometric product, the input features and the learnable geometric algebraic convolution kernel are convolved in the multi-vector domain to extract the initial multi-vector features of the fixed image, including spatial features containing orientation, rotation relationship and high-order geometric structure information. Geometric products consist of inner and outer product terms. The inner product reflects the metric relationship between two vectors, while the outer product reflects the directional information of the plane spanned by the vector pair. Through this calculation method, the convolution result can simultaneously contain local intensity dependence information and spatial geometric structure information.

[0025] S15. Then, the encoder uses a pyramid structure with five scale levels to extract multi-scale geometric features of spatial features from coarse to fine, and then recombines them into Clifford structured features. S16. Normalize the structured features and apply a non-linear activation function to obtain the GAConv output features, which in turn yields the fixed features of the fixed image. S17. Apply the same features as S12-S16 to the moving image to obtain the moving features of the moving image.

[0026] The geometric algebraic convolution module, based on the multi-vector representation capabilities of geometric algebra (Clifford Algebra), extends the feature representation method of traditional convolutional neural networks, which only performs weighted summation in the real number domain, to a multi-vector space. Geometric algebra can represent scalars, vectors, bivectors, and higher-order multivectors in a unified algebraic system, thereby achieving simultaneous encoding of geometric information of different orders and enabling features to fully reflect the spatial relationships contained in the image.

[0027] S2. Perform feature transformation and fusion on the multi-scale geometric features extracted in S1 to obtain joint features, which are used for deformation prediction at the current scale. S21. Determine the current scale level, input the initial deformation field and decoded features; S22. Use the deformation field to perform a spatial transformation on the moving image features, map the moving image features to a fixed image coordinate system, and obtain the aligned moving image features. S23. The moving image features, fixed features of the fixed image, and decoded features obtained in S22 are concatenated in the channel dimension and input into the feature fusion module based on Swin Transformer to model the long-range correlation across images and output the fused features at this scale.

[0028] S3. The joint features are weighted and adjusted using the fuzzy weighting module; The fuzzy weighting module introduces a membership modeling mechanism based on fuzzy set theory. By adaptively weighting and adjusting features, it effectively enhances the network's robustness to uncertain regions. At each scale, the fused features from fixed and moving images are input into the fuzzy weighting module.

[0029] S31. Input the fusion features obtained in S2 into the fuzzy weighting module; S32. The fuzzy weighting module constructs a fuzzy membership degree based on the relevant indicators of the features to measure the degree to which each feature point belongs to the "reliable structure" and "fuzzy region". The construction process of the fuzzy membership degree does not rely on manual annotation or additional supervision signals, but is adaptively generated based on the relative relationship between features through learnable fuzzy mapping rules. S33. The fuzzy weighting module adjusts the weights of feature channels or spatial locations based on membership degrees.

[0030] S331. Increase the feature weights of regions with high membership and clear structure to enhance the sensitivity of deformation estimation to important structures; S332. Reduce the weight of features with low membership or located in fuzzy or noisy regions to reduce the interference of these regions on deformation prediction.

[0031] S4. Input the weighted joint features into the deformation estimation module, and refine the scale of the joint features through the scale-by-scale adaptive refinement module. Then return the deformation field at this time to S2 for iterative operation until the deformation estimation at the finest scale is completed, and obtain the high-resolution deformation field. The adaptive refinement module is used to dynamically optimize the deformation field at each scale in multiple rounds during the inference phase. This allows the network to automatically adjust the number of iterations based on the difficulty of the local structure, avoiding insufficient or excessive updates caused by fixed iterations and improving the stability of deformation estimation. At each scale, the network first receives the deformation field upsampled from the previous scale as the initial deformation.

[0032] S41. Input the fuzzy weighted features into the deformation estimation module to predict the deformation increment at the current scale. Then, superimpose the deformation increment with the initial deformation field to obtain the updated deformation field. S42. Introduce a scale-wise adaptive thinning mechanism to perform spatial transformation on the moving image based on the updated deformation field at this time to obtain the deformation image; S43. Measure the similarity between the deformed image obtained in S42 and the fixed image to obtain the similarity at this time; S44. Set a similarity enhancement threshold, take the updated deformation field obtained in S41 as the new initial deformation field and return to S2 using the same decoding features for iteration until the similarity enhancement is lower than the preset threshold, and obtain the optimal deformation field of the current layer of the multi-layer decoder. S45. Upsample the optimal deformation field and the fused features at this time to obtain a new initial deformation field and decoding features. Enter the next layer of the decoder and return to S2 using the new initial deformation field and decoding features. After 5 iterations of the multi-layer decoder, the finest scale deformation field, i.e., the high-resolution deformation field, is obtained.

[0033] S5. Process the original input moving image using a high-resolution deformation field to obtain a deformed image, and then accurately register it with the original fixed image.

[0034] S51. Use a high-resolution deformation field to perform a spatial transformation on the initial moving image to obtain the deformed image at this time; S52. Using the deformed image and the fixed image at this point, a similarity test is performed to obtain the registration result. The high-resolution deformation field is then verified using a loss function, balancing registration accuracy with the geometric rationality of the deformation field, thus completing the accurate registration of the moving image to the fixed image. The loss function is as follows: ; ; ; in, Indicates the total loss. This indicates the similarity between the deformed image obtained in S42 and the fixed image. This represents the smoothness and reversibility constraints imposed on the predicted deformation field. Indicates the trade-off coefficient. and These represent moving and stationary images, respectively. This represents the predicted deformation field. This represents the deformed image after the moving image has undergone deformation field transformation. Represents the deformation field In position Spatial gradient, This represents the balance coefficient.

[0035] Considering the potential variations in brightness and contrast in medical images under different scanning conditions, this paper selects Local Normalized Cross-Correlation (NCC) as a similarity measure. NCC can measure structural similarity within a local range and exhibits strong robustness to linear changes in grayscale intensity. The negative sign in the calculation process is used to transform the similarity maximization problem into a minimization problem. In order to obtain physically reasonable registration results, the deformation field needs to satisfy the constraints of spatial smoothness and local invertibility.

[0036] The first term in the calculation process For diffusion-type regularization, it represents global average smoothing regularization. This is used to measure whether the deformation near a point is drastic. By penalizing the square of the deformation field gradient, it suppresses abrupt displacements in space and prevents discontinuous deformation. (Second term) Based on the Jacobian Determinant (JD), it is used to detect and penalize locations where local folds / folds occur in the deformation field.

[0037] The processes S2, S3, and S4 are all completed in the multi-layer decoder.

[0038] The usage process of this method is as follows: Figure 1 As shown: Fixed and moving images are input into a geometric algebraic encoder, where spatial structural features are extracted through multi-vector construction and geometric multiplication. The fixed features and the moving features (aligned via a warping module) are deeply fused in a feature fusion module to obtain joint features suitable for downstream deformation prediction. The fused features are then weighted by a fuzzy attention module to suppress low-contrast regions and noise interference, and to enhance the response to boundaries and salient structures.

[0039] Subsequently, the features are fed into the deformation estimation module to predict the deformation field at the current scale. The deformation field is used to perform spatial transformation on the moving image and calculate a similarity index with the fixed image, serving as the basis for adaptive thinning at the current scale. When the local registration improvement is lower than a threshold, early termination is triggered. If the termination condition is not met, a thinning iteration is performed at the current scale to further optimize the deformation estimation. The final output deformation field is upsampled and passed to the next scale, and the above steps are repeated until a high-resolution deformation field is obtained, achieving accurate registration of the moving image to the fixed image. Through a five-layer structure that progresses step by step from coarse to fine, this invention can achieve progressive optimization from global alignment to fine-grained local structure, ultimately outputting a high-resolution deformation field and achieving accurate registration of the moving image to the fixed image.

[0040] The technical effects of the method of the present invention are as follows: (1) By using the geometric algebraic convolution module to uniformly express scalar, orientation and multi-order geometric information, the network can still maintain its sensitivity to changes in spatial orientation and local morphological differences in complex anatomical regions, which is beneficial to obtaining a structurally continuous deformation field.

[0041] (2) The fuzzy weighting module can adaptively reduce the interference of noise and fuzzy boundaries, highlight the reliable structural area, and enable the network to maintain stable registration performance under conditions such as low contrast and bias field artifacts.

[0042] (3) By dynamically adjusting the number of iterations based on the similarity changes through the scale-by-scale adaptive refinement module, the excessive or insufficient deformation caused by fixed iterations can be avoided, and more refined local structure alignment can be obtained in the fine-scale region.

[0043] (4) The smoothing term in the comprehensive loss function works together with the Jacobian constraint to effectively suppress unreasonable deformations such as local folding and abnormal gradients, so that the final deformation field meets the requirements of spatial smoothness and reversibility.

[0044] Therefore, this invention employs the above-described adaptive thinning medical image registration method based on geometric fuzziness, which is collaboratively completed by a geometric algebraic encoding module, a feature fusion module, a fuzzy weighting module, a deformation estimation module, and a scale-by-scale adaptive thinning module. These modules are connected sequentially, with the output of one module serving as the input of the next, and the deformation field is progressively transmitted and updated across the multi-scale structure, thereby achieving precise medical image registration from coarse to fine.

[0045] 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 method of adaptive refinement of medical image registration based on geometric blur, characterized in that, The method comprises the following steps: S1, input fixed image and moving image, and extract multi-scale geometric features of medical image through geometric algebra convolution module; S2, perform feature transformation and fusion on the multi-scale geometric features extracted in S1 to obtain joint features; S3, perform weighted adjustment on the joint features through a fuzzy weighting module; S4, input the joint features after weighted adjustment into a deformation estimation module, and refine the scale of the joint features through a scale-by-scale adaptive refinement module, then return the deformation field at this time to S2 for iterative operation until the deformation estimation of the finest scale is completed, to obtain a high-resolution deformation field; S5, use the high-resolution deformation field to process the original input moving image to obtain a deformation image, and accurately register the original fixed image.

2. The geometric blur based adaptive refinement medical image registration method of claim 1, wherein, The process of S1 is as follows: S11, obtain the fixed image and the moving image to be registered, and input the fixed image and the moving image into the geometric algebra convolution module; S12, the geometric algebra convolution module inputs the fixed image into a geometric algebra encoding unit; S13, the geometric algebra encoding unit groups the feature maps of the fixed image based on the multi-vector representation rules of geometric algebra, and maps each group of sub-features to a multi-vector space to form initial multi-vector features containing multi-order geometric components; S14, according to the definition of geometric product, the input features and the learnable geometric algebra convolution kernel perform convolution operation in the multi-vector domain to extract the spatial features of the initial multi-vector features of the fixed image containing directionality, rotation relationship and high-order geometric structure information; S15, then the encoder adopts a pyramid structure with five scale levels to extract multi-scale geometric features of the spatial features from coarse to fine, and recombines the multi-scale geometric features into Clifford structured features; S16, perform normalization layer and nonlinear activation function processing on the structured features to obtain GAConv output features, and obtain fixed features of the fixed image at this time; S17, obtain moving features of the moving image by using the same features as S12-S16.

3. The geometric blur based adaptive refinement medical image registration method of claim 2, wherein, The process of S2 is as follows: S21, determine the current scale level, input the initial deformation field and the decoded features; S22, perform spatial transformation on the moving image features by using the deformation field to map the moving image features to the fixed image coordinate system to obtain the aligned moving image features; S23, splice the moving image features obtained in S22, the fixed features of the fixed image and the decoded features in the channel dimension, and input them into a feature fusion module based on Swin Transformer to model the long-range correlation across images and output the fusion features of this scale.

4. The geometric blur based adaptive refinement medical image registration method of claim 3, wherein, The process of S3 is as follows: S31, input the fusion features obtained in S2 into a fuzzy weighting module; S32, the fuzzy weighting module constructs a fuzzy membership degree according to the correlation index of the features to measure the degree of each feature point belonging to "reliable structure" and "fuzzy area"; S33, the fuzzy weighting module performs weighted adjustment on the feature channels or spatial positions according to the membership degree.

5. The geometric blur based adaptive refinement medical image registration method of claim 4, wherein: The fuzzy membership degree in S32 is adaptively generated according to the relative relationship between the features through a learnable fuzzy mapping rule.

6. The geometric blur based adaptive refinement medical image registration method of claim 5, wherein, The weighted adjustment in S33 is as follows: S331, for the higher membership, clear structure of the region to enhance its feature weight, To enhance the sensitivity of deformation estimation to important structures; S332, for the lower membership, in the fuzzy or noise interference area of the feature to reduce the weight, to reduce the interference of these areas to deformation prediction.

7. The geometric blur based adaptive refinement medical image registration method of claim 6, wherein, The process of S4 is as follows: S41, input the fuzzy weighted feature into the deformation estimation module, predict the deformation increment of the current scale, superimpose the deformation increment and the initial deformation field to obtain the updated deformation field; S42, introduce the scale-by-scale adaptive refinement mechanism to perform spatial transformation on the moving image according to the updated deformation field at this time to obtain the deformation image; S43, measure the similarity between the deformation image obtained in S42 and the fixed image to obtain the similarity at this time; S44, set a similarity improvement threshold, take the updated deformation field obtained in S41 as the new initial deformation field and use the same decoding feature to return to S2 for iteration, until the improvement of the similarity is lower than the preset threshold, the optimal deformation field of the current layer of the multi-layer decoder is obtained; S45, upsample the optimal deformation field and the fusion feature at this time to obtain the new initial deformation field and the decoding feature, enter the next layer of the decoder, and use the new initial deformation field and the decoding feature to return to S2, after the iteration of 5 layers of the multi-layer decoder, the finest scale deformation field, i.e. the high-resolution deformation field, is obtained.

8. The geometric blur based adaptive refinement medical image registration method of claim 7, wherein, The process of S5 is as follows: S51, use the high-resolution deformation field to perform spatial transformation on the initial moving image to obtain the deformation image at this time; S52, use the deformation image at this time and the fixed image to perform similarity test to obtain the registration result, and verify the high-resolution deformation field in combination with the loss function, taking into account the registration accuracy and the geometric rationality of the deformation field, to complete the accurate registration of the moving image to the fixed image.

9. The geometric blur based adaptive refinement medical image registration method of claim 8, wherein, The loss function in S52 is as follows: ； ； ； wherein, denotes the overall loss, denotes the similarity of the warped image and the fixed image in S42, denotes the smoothness and invertibility constraints imposed on the predicted deformation field, denotes the trade-off coefficient, and denote the moving image and the fixed image, respectively, denotes the predicted deformation field, denotes the warped image of the moving image transformed by the deformation field, denotes the deformation field the spatial gradient at position , denotes the balance coefficient, denotes the penalty term based on the Jacobian determinant.

10. The geometric blur based adaptive refinement medical image registration method of claim 9, wherein: The processes of S2, S3 and S4 are all completed in the multi-layer decoder.

Images