A medical image bidirectional non-rigid registration method and system

Abstract

The present application relates to the field of medical image processing, in order to solve the existing medical image registration in one-way registration leads to the accumulation of cyclic error, training strategy single and generalization and robustness are insufficient, etc., a kind of medical image two-way non-rigid registration method and system are provided;Medical image registration model is constructed, including forward registration network and backward registration network;The spatial transformation of mobile image to fixed image is learned using the forward registration network;The spatial transformation of fixed image to mobile image is learned using the backward registration network;In the training of medical image registration model, first, the parameters of the forward registration network and the parameters of the backward registration network are alternately trained, and then the parameters of the forward registration network and the parameters of the backward registration network are jointly optimized by the cyclic consistency loss;The present application scheme considers precision, robustness and generalization.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and in particular to a method and system for bidirectional non-rigid registration of medical images. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Medical image registration, a core supporting technology in the field of medical image processing, is a crucial step in achieving the integration of multi-source medical image information and empowering precise clinical decision-making. Its applications span multiple fields, including oncology diagnosis and treatment, neurosurgery, orthopedic repair, and chronic disease monitoring. For example, in tumor radiotherapy, registering CT images that locate anatomical structures with MRI images that distinguish tumors from normal tissues allows for precise determination of the radiotherapy target area, reducing damage to normal tissues. In neurosurgery, aligning preoperative MRI with intraoperative ultrasound images helps surgeons locate brain functional areas, reducing surgical risks. In chronic disease follow-up, registering liver ultrasound or kidney CT images at different time points allows for a direct assessment of lesion volume changes and disease progression. By aligning medical images from different times, devices, or modalities within a unified spatial coordinate system, registration technology effectively solves the problem of limited information from single-source images, providing doctors with multi-dimensional anatomical and functional references and serving as a vital bridge connecting medical imaging data with clinical practice.

[0004] Before the widespread application of deep learning technology, traditional medical image registration methods mainly revolved around three major ideas: "feature matching", "intensity similarity optimization" and "physical deformation modeling". Feature-based registration methods rely on manually labeled or automatically extracted anatomical landmarks (such as bone edges and blood vessel bifurcation) and local features (such as SIFT and SURF features). They calculate the spatial transformation matrix by matching feature points. However, manual labeling is time-consuming and highly subjective, and automatic feature extraction is prone to mismatches in low-contrast (such as ultrasound images) and noisy (such as low-dose CT) scenarios. Intensity-based registration methods (such as mutual information and normalized cross-correlation) find the optimal transformation by minimizing the gray-level differences between images without feature extraction. However, they are sensitive to differences in image gray-level distribution and are prone to getting stuck in local optima in scenarios with large nonlinear deformations (such as the deformation of abdominal organs due to respiratory movements). Physical model-based registration methods (such as elastic deformation models and optical flow methods) describe the deformation process by simulating the physical properties of tissues. However, the model assumptions often simplify the real anatomical structure and are difficult to adapt to the nonlinear deformation of complex tissues. Moreover, the iterative calculation process is complex and time-consuming, making it difficult to meet the real-time requirements of clinical scenarios such as intraoperative navigation and emergency image analysis.

[0005] In recent years, deep learning-based medical image registration methods have become a research hotspot in the field due to their advantages such as end-to-end prediction, fast inference speed, and strong ability to handle nonlinear deformations. These methods can be divided into supervised learning that relies on a large number of labeled gold standard deformation fields, unsupervised learning that optimizes solely through image intensity similarity loss, and weakly supervised learning that combines a small number of anatomical labels. Typical models such as VoxelMorph directly predict dense displacement fields through convolutional neural networks, and TransMorph introduces the Transformer architecture to improve global feature capture capabilities, significantly improving the efficiency and deformation handling capabilities of traditional methods.

[0006] However, existing deep learning registration methods still have general limitations: First, supervised learning's strong dependence on labeled data presents a significant contradiction with the scarcity of labeled data in medical scenarios. High-quality gold-standard deformation fields require professional teams to construct them using anatomical knowledge and image analysis tools, which is extremely costly and difficult to obtain on a large scale. Second, unsupervised methods rely excessively on image intensity information, which can easily lead to registration bias in scenarios with similar anatomical structures but significant functional differences (such as gray and white matter regions in the brain). Third, most methods adopt a one-way registration strategy (only calculating the deformation field from the moving image to the fixed image), ignoring the consistency constraints of the reverse deformation field, resulting in the accumulation of cyclic errors and affecting registration robustness. Fourth, the model's generalization ability is insufficient. When a model trained for a specific site (such as the brain) is applied to other sites (such as the lungs and liver), its performance drops significantly due to differences in tissue elasticity and deformation patterns. In addition, existing training strategies mostly use a single optimization objective or one-time joint optimization, failing to design an adapted training process for the differentiated learning needs of the forward and reverse deformation fields, further limiting registration accuracy and stability.

[0007] To address the aforementioned general technical bottlenecks, there is an urgent need to develop a medical image registration method that not only meets the accuracy requirements of static image registration (such as preoperative and postoperative comparison), but also possesses a certain degree of real-time capability to support dynamic clinical scenarios such as intraoperative navigation and real-time verification of radiotherapy, while taking into account accuracy, robustness, and generalization. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a bidirectional non-rigid registration method and system for medical images, which takes into account accuracy, robustness and generalization of medical image registration, and can ultimately be mapped to clinical intraoperative scenarios to support precision diagnosis and treatment.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a method for bidirectional non-rigid registration of medical images, comprising:

[0011] The medical image to be registered is obtained, and the obtained medical image to be registered is input into the trained medical image registration model to obtain the registration result;

[0012] The medical image registration model includes a forward registration network and a backward registration network. The forward registration network is used to learn the spatial transformation from a moving image to a fixed image. The backward registration network is used to learn the spatial transformation from a fixed image to a moving image. In the training of the medical image registration model, a cyclic cross-training strategy is executed. The parameters of the forward registration network are iteratively optimized based on the forward registration loss, and the parameters of the backward registration network are iteratively optimized based on the backward registration loss. A cyclic consistency loss is introduced to simultaneously and jointly optimize the parameters of the forward registration network and the backward registration network.

[0013] Secondly, the present invention provides a bidirectional non-rigid registration system for medical images, comprising:

[0014] The acquisition module is configured to acquire medical images to be registered;

[0015] The registration module is configured to input the acquired medical image to be registered into the trained medical image registration model to obtain the registration result;

[0016] The medical image registration model includes a forward registration network and a backward registration network. The forward registration network is used to learn the spatial transformation from a moving image to a fixed image. The backward registration network is used to learn the spatial transformation from a fixed image to a moving image. In the training of the medical image registration model, a cyclic cross-training strategy is executed. The parameters of the forward registration network are iteratively optimized based on the forward registration loss, and the parameters of the backward registration network are iteratively optimized based on the backward registration loss. A cyclic consistency loss is introduced to simultaneously and jointly optimize the parameters of the forward registration network and the backward registration network.

[0017] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0018] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0019] The above one or more technical solutions have the following beneficial effects:

[0020] This invention proposes a training method for medical image registration models based on bidirectional optimization and alternating training strategies. Through the collaborative design of dual-network architecture, alternating training, and cyclic constraints, it overcomes the limitations of unidirectional registration, avoids the error accumulation caused by traditional unidirectional registration, ensures logical consistency of the deformation field, and improves registration robustness.

[0021] The alternating training strategy proposed in this invention can specifically optimize the parameters of the dual-registration network, adapt to the differentiated learning requirements of bidirectional transformation, and combine normalized cross-correlation similarity loss and gradient smoothing loss to take into account both image grayscale overlap and anatomical structure continuity, thereby improving training efficiency and accuracy.

[0022] The present invention does not require a large amount of gold standard deformation field annotation. It achieves high-precision registration through bidirectional constraints and image grayscale similarity, thereby reducing the cost of clinical data preprocessing.

[0023] This invention is compatible with multimodal medical images such as CT, MRI, and ultrasound, and can be extended to registration scenarios in multiple locations such as the brain, abdomen, and orthopedics, rather than being limited to a single location. It supports integration with surgical navigation systems to achieve real-time fusion of preoperative and intraoperative images, providing visualization support for precision surgery (such as tumor radiotherapy and puncture biopsy), and helping to improve the safety and effectiveness of diagnosis and treatment.

[0024] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0026] Figure 1 The main flowchart of the bidirectional non-rigid registration method for medical images provided in the embodiments of the present invention is shown below;

[0027] Figure 2 A flowchart of the deep learning network architecture during the training phase provided in an embodiment of the present invention;

[0028] Figure 3 This is a flowchart of network optimization during the training phase provided in an embodiment of the present invention;

[0029] Figure 4 This is a flowchart of the image transformation process during the testing phase, provided in an embodiment of the present invention. Detailed Implementation

[0030] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0031] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0032] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0033] Example 1

[0034] This embodiment discloses a bidirectional non-rigid registration method for medical images, including:

[0035] The medical image to be registered is obtained, and the obtained medical image to be registered is input into the trained medical image registration model to obtain the registration result;

[0036] The medical image registration model includes a forward registration network and a backward registration network. The forward registration network is used to learn the spatial transformation from a moving image to a fixed image, and the backward registration network is used to learn the spatial transformation from a fixed image to a moving image. In the training of the medical image registration model, a cyclic cross-training strategy is implemented. The parameters of the forward registration network are iteratively optimized based on the forward registration loss, and the parameters of the backward registration network are iteratively optimized based on the backward registration loss. A cyclic consistency loss is introduced to simultaneously and jointly optimize the parameters of the forward and backward registration networks.

[0037] This embodiment addresses the technical bottlenecks in existing deep learning medical image registration methods, such as cyclic error accumulation due to unidirectional registration, lack of deformation field consistency constraints, single training strategies that are difficult to adapt to bidirectional transformation requirements, and insufficient generalization and robustness. It proposes a medical image registration method based on bidirectional optimization and alternating training strategies. By constructing a forward-backward dual registration network and designing hierarchical optimization objectives and cyclic consistency constraints, it achieves high-precision medical image registration with low annotation dependence, which can ultimately be mapped to clinical intraoperative scenarios to support precision diagnosis and treatment.

[0038] In this embodiment, the medical image registration model employs a cyclic cross-training strategy during training: First, the forward registration network parameters are iteratively optimized based on the forward registration loss, which is a weighted average of the forward image normalized cross-correlation similarity loss and the forward deformation field gradient smoothness loss. Similarly, the backward registration network parameters are iteratively optimized based on the backward registration loss, which is a weighted average of the backward image normalized cross-correlation similarity loss and the backward deformation field gradient smoothness loss. Then, a cyclic consistency loss, which measures the grayscale restoration of the image after both forward-backward and backward-forward bidirectional transformations, is introduced to simultaneously and jointly optimize both the forward and backward registration network parameters. This "separate optimization - joint optimization" process is repeated until a preset number of training rounds is reached. This training strategy avoids the accumulation of bidirectional deformation field bias, ensures bidirectional logical consistency, and prevents unidirectional bias from becoming "fixed."

[0039] The following is combined with Figure 1The medical image registration model proposed in this embodiment for use in medical image registration is described in detail below:

[0040] S1: Acquire multimodal or multi-time point medical image data (such as preoperative MRI, intraoperative ultrasound, CT at different time points, etc.), construct the medical image dataset to be registered, and divide the medical image dataset into training set and test set in an 8:2 ratio. The training set is used for subsequent model parameter optimization, and the test set is used to verify the performance of the registration model.

[0041] In S1, those skilled in the art can obtain medical image data through clinical collection or publicly available medical image databases to construct a medical image dataset. Subsequently, the medical image dataset is randomly divided into a training set and a test set in a ratio of 8:2. The training set is used for parameter optimization of the forward and backward registration networks and provides a sample basis for bidirectional transformation learning. The test set is used to verify the generalization ability and registration accuracy of the registration model and evaluate the performance of the model on unseen data.

[0042] The image data preprocessing stage must ensure grayscale normalization (eliminating brightness differences between modalities) and uniform voxel resolution to provide a consistent data format for subsequent network input.

[0043] S2: Initialize the parameters to be optimized in the medical image registration model related to the bidirectional transformation, i.e., the forward registration network parameters. Back-registration network parameters This provides initial values ​​for subsequent bidirectional optimization operations.

[0044] The medical image registration model is constructed using a deep learning network architecture. The forward registration network learns the transformation relationship from the moving image to the target image, and its network parameters are denoted as follows: The backward registration network is used to learn the transformation relationship from the target image to the moving image, and its network parameters are denoted as follows. ; and All of them include trainable parameters such as network weights and biases.

[0045] right and Initialization can be performed in the following ways: First, random initialization, that is, according to a preset probability distribution (such as normal distribution or uniform distribution). and First, assign initial values; second, initialize based on prior knowledge, such as according to the parameter range corresponding to common small-range rigid body transformations (such as translation and rotation) between medical images. and Setting initial values ​​enables the model to have preliminary registration capabilities in its initial state, providing a reasonable starting point for subsequent bidirectional optimization operations.

[0046] For example, both the forward registration network and the backward registration network can adopt typical network model architectures in the field of existing deep learning registration, such as the VoxelMorph network for end-to-end dense displacement field prediction and the TransMorph network that introduces Transformer to improve the ability to capture global features. The two networks can adopt the same architecture to simplify parameter initialization and training logic and ensure the consistency of bidirectional transformation learning.

[0047] The following is combined with Figure 2 A detailed explanation of deep learning network architectures is provided:

[0048] S21: Input the moving image to be registered. and fixed image To the forward registration network and the backward registration network; the forward registration network to and Features are extracted, fused, and mapped to learn from moving images. To a fixed image Forward deformation field Back-registration network pairs and Features are extracted, fused, and mapped to learn from fixed images. To moving images Backward deformation field The deformation field is used to characterize the displacement distribution of image voxels in three-dimensional space, providing a basis for subsequent image spatial transformation; the forward deformation field... Characterization voxel-oriented Displacement distribution in the coordinate system; backward deformation field Characterization voxel-oriented Displacement distribution in the coordinate system.

[0049] S22: Move the image With forward deformation field The input spatial transformation module, based on the trilinear interpolation algorithm, performs... according to The defined voxel displacement rules are used to perform spatial transformations to obtain the forward-registered image. , This represents a transformation operation. The symbol is preceded by the subject being transformed and followed by the specific transformation action. The spatial transformation module establishes a mapping relationship between the original coordinates of the voxels and the target coordinates to achieve image deformation in continuous space, ensuring the grayscale continuity of the transformed image.

[0050] The spatial transformation module consists of two parts: voxel coordinate mapping establishment and trilinear interpolation calculation. The former first determines the correspondence between the original coordinates and the target coordinates of the moving image voxels based on the forward deformation field, clarifying where the voxels should be moved to. The latter then, based on the target coordinates, calculates the weighted gray values ​​of the eight surrounding integer coordinate voxels to ensure grayscale continuity, and finally completes the spatial deformation from the moving image to the forward-registered image.

[0051] Furthermore, to meet the constraints of bidirectional optimization, the forward-registered image... Backward deformation field output by the back registration network The spatial transformation module is input again, still using the trilinear interpolation algorithm, according to... voxel displacement rules for Perform a quadratic spatial transformation to establish The mapping relationship between voxel coordinates and the target coordinates of the quadratic transformation is used to calculate the restored moving image. This provides data support for subsequent verification of the consistency between the forward and backward deformation fields.

[0052] S23: Fix the image With backward deformation field The input space transformation module, based on the same interpolation algorithm as S202, performs... according to The defined voxel displacement rules are used to perform spatial transformations to obtain the back-registered image. .

[0053] Furthermore, the fixed image after back-registration Forward deformation field output by the forward registration network The image is input into the spatial transformation module again to calculate the restored fixed image. .

[0054] S3: Use the training set to initialize the parameters and Optimization is performed using a strategy of bidirectional optimization and alternating training, enabling the model to learn accurate image transformation rules.

[0055] A training process alternating between hierarchical optimization and cyclic constraints is adopted for... and Iterative optimization is performed to ensure the consistency and registration accuracy of the two-way deformation field, such as... Figure 3 As shown, it specifically includes:

[0056] S31: Forward Registration Network Parameters The optimization objective is to calculate the total loss function of the forward optimization, taking "similarity between the forward-registered image and the fixed image + smoothness of the deformation field" as the optimization criteria, and then optimize the parameters of the forward registration network. The specific steps are as follows:

[0057] S311: Calculate forward image similarity loss based on normalized cross-correlation (NCC) ,measure and The grayscale overlap is calculated using the following formula:

[0058]

[0059] in, For the first The grayscale value of the moving image after forward transformation of the voxel position. The pixel mean of the forward-registered image; To fix the gray value of the i-th voxel in the image, The average pixel value of a fixed image.

[0060] S312: Employs displacement gradient Norm penalty for non-smooth displacements prevents abrupt changes in the deformation field that violate anatomical continuity, resulting in a loss of smoothness in the forward deformation field. The formula is:

[0061]

[0062] in, The total number of image pairs in the training set. For the first The voxel space gradient of the corresponding forward deformation field for the group of images; It is the square of the L2 norm.

[0063] S313: Construct the forward optimization total loss function :

[0064]

[0065] in, The smoothness loss weight coefficient can be set to 1.0; the total forward optimization loss function is minimized through backpropagation. Iterative updates Complete this round Preliminary optimization.

[0066] S32: Back-registration network parameters The optimization logic is consistent with the forward optimization. The total loss function for backward optimization is calculated with the objective of "similarity between the back-registered image and the moving image + smoothness of the deformation field," thus optimizing the back-registration network parameters. The specific steps are as follows:

[0067] S321: Calculate backward image similarity loss based on normalized cross-correlation (NCC). measure and The grayscale overlap is calculated using the following formula:

[0068]

[0069] in, For the first The fixed image grayscale value after backward transformation of the voxel position The pixel mean of the back-registered image; For the moving image grayscale value of a voxel This represents the average pixel value of the moving image.

[0070] S322: Employing displacement gradient Norm penalty for non-smooth displacements prevents abrupt changes in the deformation field that violate anatomical continuity, resulting in a loss of smoothness in the backward deformation field. The formula is:

[0071]

[0072] in The total number of image pairs in the training set. For the first The voxel space gradient of the corresponding forward deformation field for the group of images.

[0073] S323: Construct the total loss function for backward optimization :

[0074]

[0075] in, The smoothness loss weight coefficient can be set to 1.0; the total backward optimization loss function is minimized through the backpropagation algorithm. Iterative updates Complete this round Preliminary optimization.

[0076] S33: Calculate the cycle consistency loss and jointly optimize the parameters of the forward registration network. Back-registration network parameters To measure the grayscale restoration of an image after a bidirectional "forward-backward" or "backward-forward" transformation, and to avoid logical inconsistencies in the bidirectional deformation field, the normalized cross-correlation (NCC) loss is used to calculate similarity. The formula is as follows:

[0077]

[0078] in, The normalized cross-correlation loss between the moving image after the "forward-backward" transformation and itself is calculated in the same way. ; To fix the normalized cross-correlation loss between the fixed image after the "backward-forward" transformation and itself, the backpropagation algorithm is used to minimize... Synchronous iterative updates and .

[0079] S34: Determine the current training epoch. If the number of training epochs is less than 1000, start executing again from S31; if the number of training epochs reaches 1000, stop training and save the optimized results. and .

[0080] It should be noted that the training of the medical image registration model in this embodiment follows a cyclical pattern of S31-S32-S33-S31-S32-S33... That is, each round of training is completed by optimizing the parameters of the forward registration network, optimizing the parameters of the backward registration network, and performing a joint optimization under the cyclic consistency constraint.

[0081] S4: Apply the optimized forward registration network parameters from step S3 to the test set. and back-registration network parameters The registration transformation is performed on the medical images in the test set to obtain the transformed medical images.

[0082] In S4, such as Figure 4 As shown, the forward registration network parameters obtained from step S3 are first optimized. and back-registration network parameters The data is loaded into the forward registration network and the backward registration network respectively to ensure that the network architecture is consistent with that of the training phase.

[0083] For the medical images to be registered in the test set, pair them with moving images. With target image Perform the following registration operation:

[0084] S41: Synchronize the moving image With target image Input loading A forward registration network, from which the output of the moving image is obtained. To target image Forward deformation field ;

[0085] S42: Synchronize the moving image With target image Input loading The back-registration network outputs data from the target image. To target image Backward deformation field ;

[0086] S43: Forward Deformation Field With backward deformation field To perform fusion, according to the formula The fusion deformation field was calculated. This process uses the reverse deformation field to constrain and correct the forward deformation field, thereby enhancing the accuracy and robustness of single-direction registration.

[0087] S44: Move image With fusion deformation field The input space transformation module, based on the trilinear interpolation algorithm, establishes... voxel original coordinates and The mapping relationship of the defined target coordinates is used to calculate the gray value at the target coordinates, ultimately obtaining the registered medical image. .

[0088] During spatial transformation, trilinear interpolation ensures the continuity of gray levels and the integrity of anatomical structures in the registered image by weighting the gray levels of surrounding voxels.

[0089] S5: Map the medical image obtained after transformation in step S4 to a clinical scene (such as an intraoperative scene) to provide a registered image reference for clinical diagnosis and treatment;

[0090] In step S5, the registered and transformed medical image obtained in step S4 is mapped to the clinical intraoperative scene through a medical image fusion and spatial positioning system. First, based on the spatial tracking capability of the surgical navigation device, a transformation relationship is established between the coordinate system of the registered image and the actual anatomical spatial coordinate system of the patient during surgery, ensuring that the spatial position of anatomical structures (such as tumors, blood vessels, organ outlines, etc.) in the image accurately matches the actual anatomical position of the patient during surgery. Second, the registered image (which integrates anatomical details from preoperative multimodal images) is overlaid or displayed in conjunction with images acquired in real time during surgery (such as intraoperative ultrasound images), enabling the surgeon to intuitively refer to the registered medical image during surgery, accurately locate lesions, plan surgical operation paths, or assess the impact of operations on anatomical structures, providing a visual image reference for the precise implementation of clinical surgery and helping to improve the safety and effectiveness of surgery.

[0091] This embodiment avoids error accumulation caused by traditional unidirectional registration by using a forward-backward dual network and cyclic consistency constraints, ensuring logical consistency of the deformed field and improving registration robustness. The alternating training strategy in this embodiment can specifically optimize the parameters of the dual network, adapting to the differentiated learning needs of bidirectional transformations. Combining NCC similarity loss and gradient smoothing loss, it balances image grayscale overlap and anatomical structural continuity. This method does not require extensive gold-standard deformed field annotations, achieving high-precision registration through bidirectional constraints and self-supervised learning (image grayscale similarity), reducing clinical data preprocessing costs. This method is adaptable to multimodal medical images such as CT, MRI, and ultrasound, and can be extended to registration scenarios in multiple locations such as the brain, abdomen, and orthopedics, rather than being limited to a single location. This method supports integration with surgical navigation systems, enabling real-time fusion of preoperative and intraoperative images, providing visualization support for precision surgery (such as tumor radiotherapy and biopsy), and helping to improve the safety and effectiveness of diagnosis and treatment.

[0092] This embodiment overcomes the limitations of unidirectional registration by strengthening the bidirectional consistency constraint of the deformation field through innovative optimization strategies. It also designs a flexible training mechanism to adapt to the registration needs of different anatomical sites and imaging modalities, reducing reliance on scarce labeled data. This embodiment not only meets the accuracy requirements of static image registration (such as pre- and post-operative comparisons) but also possesses a certain degree of real-time capability to support dynamic clinical scenarios such as intraoperative navigation and real-time verification during radiotherapy. Ultimately, it provides an efficient and reliable image registration solution for diverse medical clinical needs, promoting the deep application of medical imaging technology in precision medicine.

[0093] Example 2

[0094] The purpose of this embodiment is to provide a bidirectional non-rigid registration system for medical images, including:

[0095] The acquisition module is configured to acquire medical images to be registered;

[0096] The registration module is configured to input the acquired medical image to be registered into the trained medical image registration model to obtain the registration result;

[0097] The medical image registration model includes a forward registration network and a backward registration network. The forward registration network is used to learn the spatial transformation from a moving image to a fixed image. The backward registration network is used to learn the spatial transformation from a fixed image to a moving image. In the training of the medical image registration model, a cyclic cross-training strategy is executed. The parameters of the forward registration network are iteratively optimized based on the forward registration loss, and the parameters of the backward registration network are iteratively optimized based on the backward registration loss. A cyclic consistency loss is introduced to simultaneously and jointly optimize the parameters of the forward registration network and the backward registration network.

[0098] In further embodiments, the following is also provided:

[0099] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0100] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0101] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0102] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0103] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0104] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0105] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0106] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0107] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0108] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0109] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for bidirectional non-rigid registration of medical images, characterized in that, include: The medical image to be registered is obtained, and the obtained medical image to be registered is input into the trained medical image registration model to obtain the registration result; The medical image registration model includes a forward registration network and a backward registration network. The forward registration network is used to learn the spatial transformation from a moving image to a fixed image. The backward registration network is used to learn the spatial transformation from a fixed image to a moving image. During the training of the medical image registration model, a cyclic cross-training strategy is implemented. The parameters of the forward registration network are iteratively optimized based on the forward registration loss, and the parameters of the backward registration network are iteratively optimized based on the backward registration loss. A cyclic consistency loss is introduced to simultaneously and jointly optimize the parameters of both the forward and backward registration networks. The total forward optimization loss function for training the forward registration network is: ； ； ； in, Forward image similarity loss; This represents the smoothness loss of the forward deformation field; The total loss function is optimized forward. For the first The grayscale value of the moving image after forward transformation of the voxel position. The pixel mean of the forward-registered image; To fix the gray value of the i-th voxel in the image, The average pixel value of a fixed image; The total number of image pairs in the training set. For the first The voxel space gradient of the corresponding forward deformation field for the group of images; The smoothness loss weighting coefficient; It is the square of the L2 norm; The total backward optimization loss function for training the backward registration network is: ； ； ； in, For backward image similarity loss, This is due to the loss of smoothness in the backward deformation field; The total loss function is optimized backward. For the first The fixed image grayscale value after backward transformation of the voxel position The pixel mean of the back-registered image; For the moving image grayscale value of a voxel The average pixel value of the moving image; The total number of image pairs in the training set. For the first The voxel space gradient of the corresponding forward deformation field for the group of images; The smoothness loss weighting coefficient; It is the square of the L2 norm.

2. The bidirectional non-rigid registration method for medical images as described in claim 1, characterized in that, In the training of the forward registration network, the parameters of the forward registration network are updated by calculating the total forward optimization loss and minimizing the total forward optimization loss through backpropagation; the total forward optimization loss is obtained by weighted summation of the forward image similarity loss and the forward deformation field smoothness loss.

3. The bidirectional non-rigid registration method for medical images as described in claim 1, characterized in that, In the training of the backward registration network, the parameters of the backward registration network are updated by calculating the total backward optimization loss and minimizing the total backward optimization loss through backpropagation; the total backward optimization loss is obtained by weighted summation of backward image similarity loss and backward deformation field smoothness loss.

4. The bidirectional non-rigid registration method for medical images as described in claim 1, characterized in that, The parameters of the forward registration network and the backward registration network are jointly optimized by using a cycle consistency loss. Specifically, the cycle consistency loss is calculated, and the cycle consistency loss is minimized through backpropagation to synchronously update the parameters of the forward registration network and the backward registration network. The cycle consistency loss is used to measure the grayscale restoration of the image after forward-backward or backward-forward bidirectional transformation.

5. A method for bidirectional non-rigid registration of medical images as described in claim 1 or 4, characterized in that, The cycle consistency loss is specifically as follows: ； in, The restored moving image is calculated by spatially transforming the forward-registered image with the backward deformation field output by the backward-registering network; The restored fixed image is calculated by spatially transforming the backward-registered image with the forward deformation field output by the forward-registering network; The normalized cross-correlation loss between the moving image and itself after the forward-backward transformation. The normalized cross-correlation loss between the fixed image and itself after the backward-forward transformation is used; For moving images; For fixed images.

6. A bidirectional non-rigid registration system for medical images, characterized in that, include: The acquisition module is configured to acquire medical images to be registered; The registration module is configured to input the acquired medical image to be registered into the trained medical image registration model to obtain the registration result; The medical image registration model includes a forward registration network and a backward registration network. The forward registration network is used to learn the spatial transformation from a moving image to a fixed image. The backward registration network is used to learn the spatial transformation from a fixed image to a moving image. During the training of the medical image registration model, a cyclic cross-training strategy is implemented. The parameters of the forward registration network are iteratively optimized based on the forward registration loss, and the parameters of the backward registration network are iteratively optimized based on the backward registration loss. A cyclic consistency loss is introduced to simultaneously and jointly optimize the parameters of both the forward and backward registration networks. The total forward optimization loss function for training the forward registration network is: ； ； ； in, Forward image similarity loss; This represents the smoothness loss of the forward deformation field; The total loss function is optimized forward. For the first The grayscale value of the moving image after forward transformation of the voxel position. The pixel mean of the forward-registered image; To fix the gray value of the i-th voxel in the image, The average pixel value of a fixed image; The total number of image pairs in the training set. For the first The voxel space gradient of the corresponding forward deformation field for the group of images; The smoothness loss weighting coefficient; It is the square of the L2 norm; The total backward optimization loss function for training the backward registration network is: ； ； ； in, For backward image similarity loss, This is due to the loss of smoothness in the backward deformation field; The total loss function is optimized backward. For the first The fixed image grayscale value after backward transformation of the voxel position The pixel mean of the back-registered image; For the moving image grayscale value of a voxel The average pixel value of the moving image; The total number of image pairs in the training set. For the first The voxel space gradient of the corresponding forward deformation field for the group of images; The smoothness loss weighting coefficient; It is the square of the L2 norm.

7. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-5.

Images