A method and system for progressive deformation recovery medical image registration
By employing a progressive deformation recovery strategy and generating a smooth deformation field through the superposition of multi-scale random matrices, the problems of deformation folding and topological integrity in medical image registration under large deformation scenarios are solved, achieving high-precision image registration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-05-11
- Publication Date
- 2026-06-09
AI Technical Summary
Existing medical image registration methods are prone to deformation folding and image feature distortion in scenarios with large deformations, and it is difficult to guarantee the integrity of the topological structure.
A progressive deformation recovery strategy is adopted, which gradually reduces the amplitude of the random deformation field and generates a smooth deformation field by combining multi-scale random matrix superposition. The target deformation field is generated by iterative optimization using a medical image registration network model.
It significantly improves registration accuracy, avoids local optima, ensures the smoothness of the deformation field and the integrity of the image topology, and is suitable for large-deformation medical image registration scenarios.
Smart Images
Figure CN122176023A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and specifically to a medical image registration method and system for progressive deformation recovery. Background Technology
[0002] Medical image registration is a key technology that maps medical images (such as CT, MRI, ultrasound images, etc.) obtained at different times, in different modalities, from different perspectives, or from different individuals to the same spatial coordinate system. It is widely used in medical scenarios such as lesion localization, disease diagnosis, surgical planning, and efficacy evaluation.
[0003] Existing medical image registration methods are mainly divided into traditional registration methods and deep learning-based registration methods. Traditional registration methods rely on manually designed feature extraction and similarity measurement, resulting in low registration efficiency, poor robustness, and difficulty in adapting to complex deformation scenarios. Deep learning-based registration methods automatically learn image features and deformation patterns through neural networks, significantly improving registration efficiency, but still have many shortcomings: On the one hand, most methods directly perform one-time deformation registration on floating images, which is prone to getting trapped in local optima, leading to insufficient registration accuracy, especially in large deformation scenarios, where problems such as deformation folding and image feature distortion are likely to occur; on the other hand, the deformation field generated by existing methods lacks smoothness, making it difficult to guarantee the topological integrity of medical images.
[0004] Therefore, it is necessary to provide new medical image registration methods for progressive deformation recovery. Summary of the Invention
[0005] Based on the aforementioned problems in the existing technology, the purpose of this invention is to provide a medical image registration method and system for progressive deformation recovery. By using a progressive deformation recovery strategy and a multi-scale random deformation field optimization generation method, the accuracy of medical image registration is improved while effectively ensuring the smoothness of the deformation field and the integrity of the image topology.
[0006] To achieve the above objectives, the technical solution adopted by this invention is: a medical image registration method for progressive deformation recovery, comprising:
[0007] S1, acquire the original floating image and the original fixed image, generate the first random deformation field and perform spatial transformation on the original floating image to obtain the first floating image; S2, calculate the similarity between the first floating image and the original fixed image, and input the first floating image, the original fixed image and the similarity into the medical image registration network model to obtain the first registration deformation field; S3, Generate a second random deformation field, merge the first registration deformation field and the second random deformation field, and then perform a spatial transformation on the first floating image to obtain an updated first floating image. The amplitude of the second random deformation field is smaller than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly to 0 with the increase of the number of iterations. S4, preset the maximum number of iterations and determine whether the maximum number of iterations has been reached; if not, return to step S2 with the updated first floating image as the first floating image; if the maximum number of iterations has been reached, output all random deformation fields and registration deformation fields. S5, merge all random deformation fields and registered deformation fields in sequence to obtain the target deformation field; S6. The target deformation field is used to perform spatial transformation on the original floating image to obtain a registration image aligned with the original fixed image.
[0008] Furthermore, the generation process of the first random deformation field includes: Generate n random matrices of different sizes, each of which follows a normal distribution; Upsample n random matrices of different sizes to the maximum size, where the maximum size is the size of the original floating image; The random velocity field is obtained by summing the n upsampled random matrices of different sizes. The random velocity field is integrated using the scaling method to obtain a smooth deformation field, which is then multiplied by the amplitude parameter to obtain the first random deformation field.
[0009] Furthermore, the calculation of the similarity between the first floating image and the original fixed image includes: if the first floating image and the original fixed image are images of the same modality, then normalized cross-correlation is used to calculate the similarity; if the first floating image and the original fixed image are images of different modalities, mutual information is preferred to calculate the similarity; or the depth features of the first floating image and the original fixed image are extracted, and the mean squared error between the features is used to calculate the similarity.
[0010] Furthermore, the medical image registration network model adds a similarity encoding module and a feature fusion module to the existing VoxelMorph registration network. It uses fully connected layers to encode the similarity and adds it to the image features of each layer to achieve the fusion of similarity features and image features. The input of the medical image registration network model is a first floating image, the original fixed image and the similarity between the images, and the output is a deformation field or a velocity field. The velocity field is integrated to obtain the deformation field.
[0011] Furthermore, the training process of the medical image registration network model includes: Construct a medical image registration dataset, with each training data set including at least one floating sample image and one fixed sample image; The training data also includes segmentation labels for the target regions in both floating and fixed sample images; Following the method of generating the random deformation field in step S1, the amplitude parameter is set to a random number to generate a random deformation field with random amplitude. The sample floating image is then spatially transformed to obtain a distorted floating image. Calculate the similarity between the distorted floating image and the sample fixed image. Input the distorted floating image, the sample fixed image and the similarity into the medical image registration network model. If the output is a velocity field, integrate the velocity field to obtain the registration deformation field. Use the registration deformation field to perform spatial transformation on the distorted floating image to obtain the second floating image. The loss function is calculated, and the parameters of the medical image registration network model are updated through backpropagation and optimization methods. The training is iterated until convergence is achieved, thus completing the training of the medical image registration network model.
[0012] Furthermore, the loss function includes the similarity loss between the second floating image and the sample fixed image and the regularization term loss of the deformation field; the training data also includes segmentation labels, and the loss function also includes the Dice loss of the registered segmentation labels.
[0013] Furthermore, the step of generating a second random deformation field, merging the first registered deformation field and the second random deformation field, and then performing a spatial transformation on the first floating image to obtain an updated first floating image includes: A second random deformation field is generated. The amplitude parameter of the second random deformation field decreases linearly according to the current iteration number. The initial amplitude of the second random deformation field is smaller than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly with the increase of the iteration number until it becomes 0. The first registered deformation field is merged with the second random deformation field to obtain the merged deformation field; The formula for calculating the merging method of deformation fields is as follows:
[0014] in, For the first registration deformation field, For the second random deformation field, These are the pixel coordinates in the first floating image. The deformation field after merging; The merged deformation field is used to perform a spatial transformation on the first floating image to obtain an updated first floating image.
[0015] A progressive deformation recovery medical image registration system, applied to the aforementioned progressive deformation recovery medical image registration method, the system comprising: The image random deformation module is used to acquire the original floating image and the original fixed image, generate the first random deformation field, and perform spatial transformation on the original floating image to obtain the first floating image; The network model registration module is used to calculate the similarity between the first floating image and the original fixed image, and input the first floating image, the original fixed image and the similarity into the medical image registration network model to obtain the first registration deformation field; The deformation field merging module is used to generate a second random deformation field. After merging the first registered deformation field and the second random deformation field, the first floating image is spatially transformed to obtain an updated first floating image. The amplitude of the second random deformation field is smaller than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly to 0 with the increase of the number of iterations. The iteration judgment module is used to preset the maximum number of iterations and determine whether the maximum number of iterations has been reached; if not, the updated first floating image is used as the first floating image and the process returns to step S2; if the maximum number of iterations has been reached, all random deformation fields and registration deformation fields are output. The target deformation field generation module is used to sequentially merge all random deformation fields and registered deformation fields to obtain the target deformation field. The registration image generation module is used to perform spatial transformation on the original floating image using the target deformation field to obtain a registration image aligned with the original fixed image.
[0016] Embodiments of the present invention also provide a network-side server, comprising: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the aforementioned progressive deformation recovery medical image registration method.
[0017] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described progressive deformation recovery medical image registration method.
[0018] The beneficial effects of this invention are as follows: The progressive deformation recovery medical image registration method of this invention adopts a progressive deformation recovery strategy, which realizes progressive registration of floating images by gradually reducing the amplitude of the random deformation field. This effectively avoids the problem of getting trapped in local optima in one-time registration, and significantly improves the registration accuracy, especially suitable for medical image registration scenarios with large deformation. At the same time, the generation method of the random deformation field is optimized. A random velocity field is generated by superimposing multi-scale random matrices and then integrated to obtain a smooth deformation field, which can ensure the smoothness and continuity of the deformation field, avoid deformation folding and image feature distortion, and ensure the integrity of the medical image topology. Attached Figure Description
[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0020] In the picture: Figure 1 A flowchart of a medical image registration method for progressive deformation recovery provided in the first embodiment of the present invention; Figure 2 A schematic diagram illustrating the change of the random deformation field during the gradual deformation recovery process, provided for the first embodiment of the present invention; Figure 3 A schematic diagram illustrating the changes in the floating image during the progressive deformation recovery process provided in the first embodiment of the present invention; Figure 4 A schematic diagram of the modules of a medical image registration system for progressive deformation recovery provided in the second embodiment of the present invention; Figure 5 This is a schematic diagram of the network-side server provided according to the third embodiment of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] First implementation method: The first embodiment of the present invention provides a medical image registration method for progressive deformation recovery, comprising: acquiring an original floating image and an original fixed image; generating a first random deformation field and performing a spatial transformation on the original floating image to obtain a first floating image; calculating the similarity between the first floating image and the original fixed image, and inputting the first floating image, the original fixed image, and the similarity into a medical image registration network model to obtain a first registration deformation field; generating a second random deformation field, merging the first registration deformation field and the second random deformation field, and then performing a spatial transformation on the first floating image to obtain an updated first floating image; preset a maximum number of iterations and determining whether the maximum number of iterations has been reached; if not, returning the updated first floating image as the first floating image to calculate the similarity for iteration; if the maximum number of iterations has been reached, outputting all random deformation fields and registration deformation fields; sequentially merging all random deformation fields and registration deformation fields to obtain a target deformation field; and using the target deformation field to perform a spatial transformation on the original floating image to obtain a registration image aligned with the original fixed image. The progressive deformation recovery medical image registration method of this invention adopts a progressive deformation recovery strategy. By gradually reducing the amplitude of the random deformation field, progressive registration of floating images is achieved, which effectively avoids the problem of getting trapped in local optima in one-time registration and significantly improves the registration accuracy. It is especially suitable for medical image registration scenarios with large deformation. At the same time, the generation method of the random deformation field is optimized. A random velocity field is generated by superimposing multi-scale random matrices and then integrated to obtain a smooth deformation field. This can ensure the smoothness and continuity of the deformation field, avoid deformation folding and image feature distortion, and ensure the integrity of the medical image topology.
[0023] The following provides a detailed explanation of the implementation details of the progressive deformation recovery medical image registration method in this embodiment. The following content is for ease of understanding and is not essential for implementing this solution. The specific process of this embodiment is as follows: Figure 1 As shown.
[0024] Step S1: Obtain the original floating image and the original fixed image, generate the first random deformation field, and perform spatial transformation on the original floating image to obtain the first floating image.
[0025] Specifically, the original floating image is the medical image to be registered, and the original fixed image is the reference medical image for registration. Both can come from medical scan data of the same patient at different times and in different modalities, or from medical scan data of different patients in the same modality.
[0026] The generation process of the first random deformation field includes: Step S11: Generate n random matrices of different sizes, each random matrix following a normal distribution.
[0027] Step S12: Upsample the n random matrices of different sizes to the maximum size, where the maximum size is the size of the original floating image.
[0028] Step S13: Add the n upsampled random matrices of different sizes to obtain the random velocity field.
[0029] Step S14: The random velocity field is integrated using the scaling method to obtain a smooth deformation field, which is then multiplied by the amplitude parameter to obtain the first random deformation field.
[0030] As an example, the original floating image has a size of 256×256×128, and the original fixed image has the same size as the original floating image. The generation process of the first random deformation field includes: generating three random matrices of different sizes, namely 64×64×32, 128×128×64 and 256×256×128, each random matrix satisfying a standard normal distribution; The first two small-sized random matrices (64×64×32, 128×128×64) are upsampled to 256×256×128 and multiplied by the corresponding upsampling rate. The three matrices are then added together to obtain the random velocity field. Finally, the random velocity field is integrated using the scaling method and multiplied by the amplitude parameter to obtain the smooth first random deformation field.
[0031] Step S2: Calculate the similarity between the first floating image and the original fixed image, and input the first floating image, the original fixed image and the similarity into the medical image registration network model to obtain the first registration deformation field.
[0032] Specifically, the steps of calculating the similarity between the first floating image and the original fixed image, and inputting the first floating image, the original fixed image, and the similarity into the medical image registration network model to obtain the first registration deformation field include the following steps: Step S21: Calculate the similarity between the first floating image and the original fixed image.
[0033] Specifically, calculating the similarity between the first floating image and the original fixed image includes: if the first floating image and the original fixed image are images of the same modality, then normalized cross-correlation (NCC) is used to calculate the similarity; if the first floating image and the original fixed image are images of different modalities, mutual information (MI) is preferred to calculate the similarity; or the depth features of the first floating image and the original fixed image are extracted, and the mean squared error (MSE) between the features is used to calculate the similarity.
[0034] Step S22: Input the first floating image, the original fixed image, and the similarity into the pre-trained medical image registration network model.
[0035] Specifically, the medical image registration network model is any trainable deep learning network. The input to the model consists of a first floating image, the original fixed image, and the similarity between the images. The output is a deformation field or a velocity field; the velocity field is integrated to obtain the deformation field. Based on existing registration networks (such as VoxelMorph, Nice-trans, SACB, etc.), the model adds a similarity encoding module and a feature fusion module. It uses fully connected layers to encode the similarity and adds it to the image features of each layer, achieving the fusion of similarity features and image features. Through pre-training, the model can automatically learn the deformation mapping relationship between the first floating image and the original fixed image, and guided by the similarity of the current image pair, outputs a registration deformation field that aligns the two images.
[0036] Step S23: If the medical image registration network model outputs a velocity field, then the velocity field is integrated to obtain the first registration deformation field; if the medical image registration network model directly outputs a deformation field, then the deformation field is the first registration deformation field.
[0037] Furthermore, the training process of the medical image registration network model includes: Step S221: Construct a medical image registration dataset, where each training dataset includes at least one floating sample image and one fixed sample image.
[0038] Specifically, training data can also include segmentation labels for target regions in both floating and fixed sample images. Segmentation labels can mark key regions such as lesions and organs in medical images. By introducing segmentation labels, the registration accuracy of the registration network for key regions can be improved, meeting the needs of clinical diagnosis and surgical planning.
[0039] As an example, the abdominal CT dataset is used. Each training dataset includes one floating image, one fixed image, and the corresponding multi-organ segmentation label.
[0040] Step S222: Following the method of generating the random deformation field in steps S11 to S14, the amplitude parameter is set to a random number to generate a random deformation field with random amplitude. The sample floating image is then spatially transformed to obtain a distorted floating image.
[0041] Step S223: Calculate the similarity between the distorted floating image and the sample fixed image. Input the distorted floating image, the sample fixed image, and the similarity into the medical image registration network model. If the output is a velocity field, integrate the velocity field to obtain the registration deformation field. Use the registration deformation field to perform spatial transformation on the distorted floating image to obtain the second floating image.
[0042] During training, data augmentation is performed on the training data using random deformation fields to improve the network's generalization ability, enabling it to adapt to different deformation scenarios.
[0043] Step S224: Calculate the loss function, update the parameters of the medical image registration network model through backpropagation and optimization methods, iterate the training until convergence, and complete the training of the medical image registration network model.
[0044] The loss function includes the similarity loss between the second floating image and the fixed sample image, and the regularization term loss of the deformation field. If the training data includes segmentation labels, the loss function also includes the Dice loss of the registered segmentation labels. The image similarity loss measures the pixel-level similarity between the registered image and the fixed image, and can be measured using methods such as mean squared error or normalized cross-correlation. The regularization term loss of the deformation field constrains the smoothness of the deformation field, avoiding deformation folds and abrupt changes, and can use methods such as the L2 norm of the deformation field gradient. The Dice loss measures the overlap between the registered segmentation labels and the fixed image segmentation labels. The Dice loss effectively improves the registration accuracy of key regions, especially suitable for small target registration scenarios such as lesion areas. Through the joint constraints of multi-dimensional loss functions, registration accuracy, deformation field smoothness, and key region matching can be guaranteed simultaneously.
[0045] As an example, the loss function is a weighted sum of image similarity loss, deformation field regularization term loss, and segmentation label Dice loss, with weights of 1, 10, and 1, respectively. The Adam optimization method is used to iteratively train the medical image registration network model for 500 epochs.
[0046] Step S3: Generate a second random deformation field, merge the first registration deformation field and the second random deformation field, and then perform a spatial transformation on the first floating image to obtain the updated first floating image; wherein, the amplitude of the second random deformation field is smaller than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly to 0 with the increase of the number of iterations.
[0047] Specifically, generating the second random deformation field, merging the first registered deformation field and the second random deformation field, and then performing a spatial transformation on the first floating image to obtain the updated first floating image includes the following steps: Step S31: Generate the second random deformation field.
[0048] Specifically, the second random deformation field is generated in the manner described in steps S11 to S14, namely: n random matrices of different sizes are generated, each random matrix following a normal distribution; the n random matrices of different sizes are upsampled to their maximum size, which is the size of the original floating image; the upsampled n random matrices of different sizes are added together to obtain a random velocity field; the random velocity field is integrated using the scaling method to obtain a smooth deformation field, which is then multiplied by the amplitude parameter to obtain the second random deformation field.
[0049] The amplitude parameter of the second random deformation field decreases linearly according to the current iteration number. The initial amplitude of the second random deformation field is smaller than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly with the increase of the iteration number until it becomes 0.
[0050] Step S32: Merge the first registered deformation field with the second random deformation field to obtain the merged deformation field.
[0051] Specifically, the formula for calculating the merging method of the deformation field is as follows:
[0052] in, For the first registration deformation field, For the second random deformation field, These are the pixel coordinates in the first floating image. This represents the deformation field after merging.
[0053] This merging method ensures the continuity and smoothness of the deformation field after merging.
[0054] Step S33: The merged deformation field is used to perform spatial transformation on the first floating image to obtain the updated first floating image.
[0055] Progressive deformation recovery of floating images is achieved by gradually reducing the amplitude of the random deformation field: In the initial iteration, the amplitude of the random deformation field is large to explore the global optimal registration space; as the number of iterations increases, the amplitude of the random deformation field decreases linearly, gradually focusing on the local optimum to avoid getting trapped in a local optimum from the beginning; in the last iteration, the amplitude of the second random deformation field is 0, and at this time only the registration deformation field is used to update the floating image, ensuring that the final registration result is completely determined by the registration deformation field, thus improving the registration accuracy.
[0056] As an example, the maximum number of iterations is set to 5. In each iteration, the second random deformation field is obtained according to the generation method of steps S11 to S14, but its amplitude parameter decreases linearly from the initial value according to the current iteration number. The amplitude of the second random deformation field is 0 in the 5th iteration. After merging the first registration deformation field with the second random deformation field of the current iteration, the first floating image is spatially transformed to obtain the first floating image updated in this iteration.
[0057] Step S4: Preset the maximum number of iterations and determine whether the maximum number of iterations has been reached; if not, return to step S2 with the updated first floating image as the first floating image; if the maximum number of iterations has been reached, output all random deformation fields and registration deformation fields.
[0058] Specifically, the preset maximum number of iterations is determined, and it is determined whether the maximum number of iterations has been reached. If the maximum number of iterations has not been reached, the updated first floating image is used as the first floating image and the process returns to step S2. If the maximum number of iterations has been reached, the output of all random deformation fields and registration deformation fields includes the following steps: Step S41: Preset the maximum number of iterations.
[0059] Specifically, the maximum number of iterations can be set according to the actual registration requirements and image complexity, usually 3-10 times.
[0060] Step S42: Determine whether the current iteration count has reached the maximum iteration count.
[0061] Step S43: If the current iteration number has not reached the maximum iteration number, the updated first floating image obtained in step S3 is returned to step S2 as the first floating image, and steps S2 and S3 are continued; wherein, before returning to step S2, the current iteration number is increased by 1, and the amplitude of the second random deformation field decreases linearly according to the current iteration number.
[0062] Step S44: If the current iteration number reaches the maximum iteration number, output all random deformation fields and registration deformation fields generated during the iteration process.
[0063] Furthermore, in the final iteration, the amplitude of the second random deformation field is 0. At this point, only the registration deformation field is used to update the floating image, ensuring that the final registration result is entirely determined by the registration deformation field, thus improving registration accuracy. Through multiple iterations, the registration deformation field is gradually optimized, continuously improving the alignment accuracy between the floating and fixed images while maintaining the smoothness of the deformation field.
[0064] As an example, the maximum number of iterations is set to 5. Steps S2 and S3 are executed iteratively. In the first iteration, the second random deformation field has an initial amplitude. In the second to fourth iterations, the amplitude of the second random deformation field decreases linearly. In the fifth iteration, the amplitude of the second random deformation field is 0. At this point, only the registration deformation field is used to update the first floating image. After reaching the maximum number of iterations, the five random deformation fields and five registration deformation fields generated during the five iterations are output. Figure 2 The diagram illustrates the changes in the random deformation field during the gradual deformation recovery process. The random deformation fields 201-205 are shown in the five iterations, with the amplitude gradually decreasing, reaching zero in the fifth iteration (205). Figure 3 The diagram shows the changes in the floating image during the progressive deformation recovery process. The updated floating images 302-306 and the fixed image 307 after 5 iterations demonstrate that this embodiment of the invention achieves registration from a floating image to a fixed image in a progressive manner.
[0065] Step S5: Merge all random deformation fields and registered deformation fields sequentially to obtain the target deformation field.
[0066] Specifically, merging all random deformation fields and registration deformation fields in sequence to obtain the target deformation field includes: merging all output random deformation fields and registration deformation fields in the iterative order to obtain the target deformation field.
[0067] Specifically, the sequential merging method includes: merging the first registered deformation field and the second random deformation field generated in each iteration according to the iteration order, and then merging the deformation fields after each iteration in sequence to obtain the final target deformation field.
[0068] Furthermore, the target deformation field has good smoothness and continuity, which can avoid deformation folding problems and form a complete deformation mapping that can map the original floating image to a fixed image space.
[0069] Step S6: Use the target deformation field to perform spatial transformation on the original floating image to obtain a registration image that is aligned with the original fixed image.
[0070] Specifically, the target deformation field is used to perform spatial transformation on the original floating image to obtain a registered image aligned with the original fixed image, thus completing the registration from the floating image to the fixed image.
[0071] In addition, step S6 can also achieve differential homeomorphic registration, including: Step S61: Take the negative value of the velocity field corresponding to the random deformation field and the registered deformation field, and then integrate to obtain the inverse deformation field.
[0072] Specifically, the velocity field corresponding to the random deformation field is the random velocity field generated in step S1; the velocity field corresponding to the registration deformation field is the velocity field output by the registration network; after taking the negative value of the above velocity fields, the corresponding inverse deformation fields are obtained by integration.
[0073] Step S62: Merge all the reverse deformation fields generated during the iteration process in reverse order to obtain the reverse target deformation field.
[0074] Specifically, the merging order of the inverse deformation field is the reverse of the merging order of the target deformation field, that is, they are merged sequentially in descending order of iteration number to ensure the continuity and smoothness of the inverse deformation field.
[0075] Step S63: The original fixed image is spatially transformed using the inverse target deformation field to achieve registration from the fixed image to the original floating image, completing the differential homeomorphic registration and ensuring the reversibility of deformation and the invariance of topological structure.
[0076] Furthermore, the inverse target deformation field is applied to a fixed image to achieve differential homeomorphic image registration. Differential homeomorphic registration can guarantee the reversibility of deformation and topological invariance, avoid organ folding in medical images, and is suitable for clinical scenarios requiring bidirectional registration.
[0077] As an example, a spatial transformation is performed on the original floating image with a size of 256×256×128 using the target deformation field to obtain a registered image that is aligned with the original fixed image at the pixel level. At the same time, the velocity field output by the random deformation field and the registration network is negatively calculated and integrated to obtain the inverse deformation field. The inverse deformation fields generated by the five iterations are merged in reverse order to obtain the inverse target deformation field. This inverse target deformation field is then applied to the original fixed image to achieve reverse registration from the fixed image to the original floating image, ensuring the integrity of the organ topology in the medical image.
[0078] The progressive deformation recovery medical image registration method of this invention adopts a progressive deformation recovery strategy. By gradually reducing the amplitude of the random deformation field, progressive registration of floating images is achieved, effectively avoiding the problem of getting trapped in local optima in one-time registration, and significantly improving registration accuracy. It is especially suitable for medical image registration scenarios with large deformations. The generation method of the random deformation field is optimized. A random velocity field is generated by superimposing multi-scale random matrices, and then a smooth deformation field is obtained by integration. This can simulate the complex deformations in real clinical scenarios, while ensuring the smoothness and continuity of the deformation field, avoiding deformation folding and image feature distortion, and ensuring the integrity of the medical image topology. A loss function is designed, and the results are... By combining image similarity loss, deformation field regularization loss, and Dice loss for segmentation labels, the registration process is constrained in multiple dimensions, ensuring registration accuracy while improving the smoothness of the deformation field. It also optimizes registration results in key regions to meet clinical needs. Differential homeomorphic registration can be achieved through the inverse deformation field, ensuring the reversibility of deformation and topological invariance, adapting to bidirectional registration scenarios and further expanding the method's application scope. The registration network model has strong compatibility, allowing the use of any trainable deep learning network. Data augmentation during training enhances the network's generalization ability, enabling it to adapt to different modalities and types of medical image registration needs, making it highly practical and easy to promote.
[0079] Second implementation method: like Figure 4 As shown, the second embodiment of the present invention provides a medical image registration system for progressive deformation recovery. The system includes: an image random deformation module 401, a network model registration module 402, a deformation field merging module 403, an iterative judgment module 404, a target deformation field generation module 405, and a registration image generation module 406.
[0080] Specifically, the image random deformation module 401 is used to acquire the original floating image and the original fixed image, generate a first random deformation field, and perform spatial transformation on the original floating image to obtain the first floating image; the network model registration module 402 is used to calculate the similarity between the first floating image and the original fixed image, and input the first floating image, the original fixed image, and the similarity into the medical image registration network model to obtain the first registration deformation field; the deformation field merging module 403 is used to generate a second random deformation field, merge the first registration deformation field and the second random deformation field, and then perform spatial transformation on the first floating image to obtain the updated first floating image, wherein the second random deformation field... The amplitude is less than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly to 0 with the increase of the number of iterations; the iteration judgment module 404 is used to preset the maximum number of iterations and determine whether the maximum number of iterations has been reached; if it has not been reached, the updated first floating image is returned to step S2 as the first floating image; if the maximum number of iterations has been reached, all random deformation fields and registration deformation fields are output; the target deformation field generation module 405 is used to merge all random deformation fields and registration deformation fields in sequence to obtain the target deformation field; the registration image generation module 406 is used to perform spatial transformation on the original floating image using the target deformation field to obtain a registration image aligned with the original fixed image.
[0081] It is not difficult to see that this embodiment is a system implementation corresponding to the first embodiment, and this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the first embodiment.
[0082] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by this invention; however, this does not mean that other units are absent from this embodiment.
[0083] The third embodiment of the present invention relates to a network-side server, such as... Figure 5 As shown, it includes at least one processor 502; and a memory 501 communicatively connected to at least one processor 502; wherein the memory 501 stores instructions executable by at least one processor 502, the instructions being executed by at least one processor 502 to enable at least one processor 502 to perform the above-described data processing method.
[0084] The memory 501 and processor 502 are connected via a bus, which may include any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors 502 and memory 501 together. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 502 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 502.
[0085] Processor 502 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 501 can be used to store data used by processor 502 during operation.
[0086] The fourth embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the medical image registration method for progressive deformation recovery described in the first embodiment.
[0087] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0088] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A medical image registration method for progressive deformation recovery, characterized in that, include: S1, acquire the original floating image and the original fixed image, generate the first random deformation field and perform spatial transformation on the original floating image to obtain the first floating image; S2, calculate the similarity between the first floating image and the original fixed image, and input the first floating image, the original fixed image and the similarity into the medical image registration network model to obtain the first registration deformation field; S3, Generate a second random deformation field, merge the first registration deformation field and the second random deformation field, and then perform a spatial transformation on the first floating image to obtain an updated first floating image. The amplitude of the second random deformation field is smaller than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly to 0 with the increase of the number of iterations. S4, preset the maximum number of iterations and determine whether the maximum number of iterations has been reached; if not, return to step S2 with the updated first floating image as the first floating image; if the maximum number of iterations has been reached, output all random deformation fields and registration deformation fields. S5, merge all random deformation fields and registered deformation fields in sequence to obtain the target deformation field; S6. The target deformation field is used to perform spatial transformation on the original floating image to obtain a registration image aligned with the original fixed image.
2. The medical image registration method for progressive deformation recovery according to claim 1, characterized in that, The generation process of the first random deformation field includes: Generate n random matrices of different sizes, each of which follows a normal distribution; Upsample n random matrices of different sizes to the maximum size, where the maximum size is the size of the original floating image; The random velocity field is obtained by summing the n upsampled random matrices of different sizes. The random velocity field is integrated using the scaling method to obtain a smooth deformation field, which is then multiplied by the amplitude parameter to obtain the first random deformation field.
3. The medical image registration method for progressive deformation recovery according to claim 1, characterized in that, The calculation of the similarity between the first floating image and the original fixed image includes: if the first floating image and the original fixed image are images of the same modality, then normalized cross-correlation is used to calculate the similarity; if the first floating image and the original fixed image are images of different modalities, mutual information is preferred to calculate the similarity; or the depth features of the first floating image and the original fixed image are extracted, and the mean squared error between the features is used to calculate the similarity.
4. The medical image registration method for progressive deformation recovery according to claim 2, characterized in that, The medical image registration network model adds a similarity encoding module and a feature fusion module to the existing VoxelMorph registration network. It uses fully connected layers to encode the similarity and adds it to the image features of each layer to achieve the fusion of similarity features and image features. The input of the medical image registration network model is a first floating image, the original fixed image and the similarity between the images, and the output is a deformation field or a velocity field. The velocity field is integrated to obtain the deformation field.
5. The medical image registration method for progressive deformation recovery according to claim 4, characterized in that, The training process of the medical image registration network model includes: Construct a medical image registration dataset, with each training data set including at least one floating sample image and one fixed sample image; The training data also includes segmentation labels for the target regions in both floating and fixed sample images; Following the method of generating the random deformation field in step S1, the amplitude parameter is set to a random number to generate a random deformation field with random amplitude. The sample floating image is then spatially transformed to obtain a distorted floating image. Calculate the similarity between the distorted floating image and the sample fixed image. Input the distorted floating image, the sample fixed image and the similarity into the medical image registration network model. If the output is a velocity field, integrate the velocity field to obtain the registration deformation field. Use the registration deformation field to perform spatial transformation on the distorted floating image to obtain the second floating image. The loss function is calculated, and the parameters of the medical image registration network model are updated through backpropagation and optimization methods. The training is iterated until convergence is achieved, thus completing the training of the medical image registration network model.
6. The medical image registration method for progressive deformation recovery according to claim 5, characterized in that, The loss function includes the similarity loss between the second floating image and the sample fixed image and the regularization term loss of the deformation field; the training data also includes segmentation labels, and the loss function also includes the Dice loss of the registered segmentation labels.
7. The medical image registration method for progressive deformation recovery according to claim 1, characterized in that, The step of generating a second random deformation field, merging the first registered deformation field and the second random deformation field, and then performing a spatial transformation on the first floating image to obtain an updated first floating image includes: A second random deformation field is generated. The amplitude parameter of the second random deformation field decreases linearly according to the current iteration number. The initial amplitude of the second random deformation field is smaller than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly with the increase of the iteration number until it becomes 0. The first registered deformation field is merged with the second random deformation field to obtain the merged deformation field; The formula for calculating the merging method of deformation fields is as follows: in, For the first registration deformation field, For the second random deformation field, These are the pixel coordinates in the first floating image. The deformation field after merging; The merged deformation field is used to perform a spatial transformation on the first floating image to obtain an updated first floating image.
8. A medical image registration system for progressive deformation recovery, characterized in that, The system applied to the medical image registration method for progressive deformation recovery as described in claims 1-7 includes: The image random deformation module is used to acquire the original floating image and the original fixed image, generate the first random deformation field, and perform spatial transformation on the original floating image to obtain the first floating image; The network model registration module is used to calculate the similarity between the first floating image and the original fixed image, and input the first floating image, the original fixed image and the similarity into the medical image registration network model to obtain the first registration deformation field; The deformation field merging module is used to generate a second random deformation field. After merging the first registered deformation field and the second random deformation field, the first floating image is spatially transformed to obtain an updated first floating image. The amplitude of the second random deformation field is smaller than that of the first random deformation field, and the amplitude of the second random deformation field decreases linearly to 0 with the increase of the number of iterations. The iteration judgment module is used to preset the maximum number of iterations and determine whether the maximum number of iterations has been reached; if not, the updated first floating image is used as the first floating image and the process returns to step S2; if the maximum number of iterations has been reached, all random deformation fields and registration deformation fields are output. The target deformation field generation module is used to sequentially merge all random deformation fields and registered deformation fields to obtain the target deformation field. The registration image generation module is used to perform spatial transformation on the original floating image using the target deformation field to obtain a registration image aligned with the original fixed image.
9. A network-side server, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the medical image registration method for progressive deformation recovery as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the medical image registration method for progressive deformation recovery as described in any one of claims 1 to 7.