Registration method, intraoperative guided intervention system and computer equipment readable storage medium

Abstract

The invention relates to a registration method, an intraoperative guided intervention system and a computer readable storage medium. The method comprises the following steps: generating low-dimensional deformation data according to an initial network, and inputting the low-dimensional deformation data into a first target network model to obtain an initial deformation field; registering the first medical image and the second medical image based on the initial deformation field to obtain a registration result; and adjusting the parameters of the initial network, generating new low-dimensional deformation data according to the adjusted initial network, and returning to execute the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration condition. By using the registration method provided by the invention, the registration accuracy can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a registration method, an intraoperative guidance interventional system and a computer readable storage medium. BACKGROUND

[0002] Interventional surgery is a minimally invasive surgical method. A doctor determines the position of a lesion area in the human body according to the patient's condition and imaging examination results, and then formulates a detailed surgical plan. However, in the specific surgical process, the doctor usually needs to rely on experience and tactile guidance to operate, lacking accurate positioning and navigation.

[0003] In the traditional technology, the preoperative image and the intraoperative real-time image of the patient are obtained, and the preoperative image and the intraoperative real-time image are registered to guide the doctor to perform accurate puncture.

[0004] However, the preoperative image and the intraoperative real-time image registration in the traditional technology is inaccurate. SUMMARY

[0005] Therefore, it is necessary to provide a registration method, an intraoperative guidance interventional system and a computer readable storage medium capable of improving registration accuracy in view of the above technical problems.

[0006] In a first aspect, the present application provides a registration method, which comprises:

[0007] generating low-dimensional deformation data according to an initial network, inputting the low-dimensional deformation data into a first target network model to obtain an initial deformation field;

[0008] registering the first medical image and the second medical image based on the initial deformation field to obtain a registration result;

[0009] adjusting the parameters of the initial network, and generating new low-dimensional deformation data according to the adjusted initial network, returning to execute the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration condition.

[0010] In one of the embodiments, the generating of the low-dimensional deformation data according to the initial network comprises:

[0011] inputting the first medical image and the second medical image into the initial network to generate the low-dimensional deformation data;

[0012] or inputting the first random noise into the initial network to generate the low-dimensional deformation data.

[0013] In one of the embodiments, the registering of the first medical image and the second medical image based on the initial deformation field to obtain the registration result comprises:

[0014] According to the initial deformation field, the first medical image is deformed to obtain a deformed medical image;

[0015] A similarity between the deformed medical image and the second medical image is determined to obtain a registration result.

[0016] In one of the embodiments, the method for determining the first target network model comprises:

[0017] Obtaining training samples, the training samples comprising image samples of different phases and first standard deformation fields between the image samples of different phases;

[0018] Training the first initial network model based on the training samples, the first initial network model comprising an encoder and a decoder;

[0019] In a case where a first training stop condition is met, the decoder in the trained first initial network model is determined as the first target network model.

[0020] In one of the embodiments, the training of the initial network model based on the training samples comprises:

[0021] Inputting the training samples into the encoder to obtain first latent deformation data;

[0022] Sampling the first latent deformation data to generate new first latent deformation data;

[0023] Inputting the new first latent deformation data into the decoder to obtain first deformation field samples;

[0024] Updating parameters of the encoder and parameters of the decoder according to the first deformation field samples and the first standard deformation fields.

[0025] In one of the embodiments, the method for determining the first target network model comprises:

[0026] Obtaining training samples, the training samples comprising image samples of different phases and first standard deformation fields between the image samples of different phases;

[0027] Training the second initial network model based on the training samples, the second initial network model comprising a generator and a discriminator;

[0028] In a case where a second training stop condition is met, the generator in the trained second initial network model is determined as the first target network model.

[0029] In one of the embodiments, the training of the second initial network model based on the training samples comprises:

[0030] Inputting second random noise into the generator to generate second latent deformation data;

[0031] inputting the first standard deformation field and the second latent deformation data into the discriminator to obtain a discrimination probability;

[0032] updating parameters of the generator and parameters of the discriminator according to the discrimination probability.

[0033] In one of the embodiments, the method for determining the first target network model comprises:

[0034] obtaining training samples, the training samples comprising image samples of different phases and first standard deformation fields between the image samples of different phases;

[0035] training a third initial network model based on the training samples, the third initial network model comprising a transformation function;

[0036] in a case where a third training stop condition is met, determining an inverse function of the transformation function in the trained third initial network model as the first target network model.

[0037] In one of the embodiments, the training of the third initial network model based on the training samples comprises:

[0038] inputting the standard deformation field into the transformation function to obtain third latent deformation data;

[0039] sampling the third latent deformation data to generate new third latent deformation data;

[0040] inputting the new third latent deformation data into the inverse function of the transformation function to obtain a second deformation field sample;

[0041] updating parameters of the transformation function according to the second deformation field sample and the first standard deformation field.

[0042] In one of the embodiments, the first target network model comprises any one of a target phase model, a full phase model and a continuous phase model;

[0043] The target phase model is a phase model corresponding to a target phase selected from a plurality of independent phase models, the independent phase models being used to describe deformation fields between one phase and another phase; the full phase model being used to describe deformation fields between any phases; and the continuous phase model being used to describe continuous deformation fields between one phase and another phase.

[0044] In one of the embodiments, inputting the low-dimensional deformation data into the first target network model to obtain an initial deformation field comprises:

[0045] selecting a target phase model from a plurality of independent phase models according to the target phase of the second medical image, and determining the target phase model as the first target network model;

[0046] The low-dimensional deformation data is input into the first target network model to obtain the initial deformation field.

[0047] Secondly, one embodiment of this application provides a registration method, the method comprising:

[0048] Acquire first and second medical images;

[0049] The first medical image and the second medical image are input into the second target network model to determine the target deformation field. The second target network model includes a first encoding module and a second encoding module. The second target network model is obtained by inputting the first image sample and the second image sample into the first encoding module to obtain the first deformation sample, inputting the second standard deformation field into the second encoding module to obtain the second deformation sample, and training the second target network model based on the first deformation sample and the second deformation sample.

[0050] The second medical image and the first medical image are registered based on the target deformation field.

[0051] Thirdly, this application provides an intraoperative guidance intervention system, including a processor, which processes the following steps:

[0052] Acquire the first medical image and the second medical image. The first medical image is a preoperative image, and the second medical image is an intraoperative image.

[0053] The first and second medical images are registered according to the registration method provided in the first aspect above, so as to achieve real-time three-dimensional guidance during the operation.

[0054] In one embodiment, a respiratory gating system is used to collect respiratory status data from preoperative and intraoperative images. When the first target network model includes a target phase model or a continuous phase model, the preoperative and intraoperative images include respiratory phase information.

[0055] Fourthly, one embodiment of this application provides a registration apparatus, the apparatus comprising:

[0056] The first acquisition module is used to acquire the first medical image and the second medical image;

[0057] The input module is used to generate low-dimensional deformation data based on the initial network, and input the low-dimensional deformation data into the first target network model to obtain the initial deformation field.

[0058] The first registration module is used to register the first medical image and the second medical image based on the initial deformation field to obtain the registration result;

[0059] The generation module adjusts the parameters of the initial network and generates new low-dimensional deformation data based on the adjusted initial network. It then returns to execute the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration conditions.

[0060] Fifthly, one embodiment of this application provides a registration apparatus, the apparatus comprising:

[0061] The second acquisition module is used to acquire the first medical image and the second medical image.

[0062] A determination module is used to input a first medical image and a second medical image into a second target network model to determine the target deformation field. The second target network model includes a first encoding module and a second encoding module. The second target network model obtains a first deformation sample by inputting the first image sample and the second image sample into the first encoding module, and obtains a second deformation sample by inputting the second standard deformation field into the second encoding module. The second target network model is trained based on the first deformation sample and the second deformation sample.

[0063] The second registration module is used to register the second medical image and the first medical image according to the target deformation field.

[0064] In a sixth aspect, one embodiment of this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the methods provided in the first and second aspects described above.

[0065] In a seventh aspect, one embodiment of this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the methods provided in the first and second aspects described above.

[0066] Eighthly, one embodiment of this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the methods provided in the first and second aspects described above.

[0067] The aforementioned registration method, intraoperative guidance system, and computer-readable storage medium acquire a first medical image and a second medical image; generate low-dimensional deformation data based on an initial network, input the low-dimensional deformation data into a first target network model to obtain an initial deformation field; register the first and second medical images based on the initial deformation field to obtain a registration result; adjust the parameters of the initial network, and generate new low-dimensional deformation data based on the adjusted initial network; return to the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration conditions. In this embodiment, a pre-trained first target network model is used to transform the low-dimensional deformation data into an initial deformation field. The parameters of the initial network are adjusted based on the registration result, and new low-dimensional deformation data is generated based on the adjusted initial network. Returning to the step of inputting the low-dimensional deformation data into the first target network model, this iterative update of the initial deformation field improves the accuracy of the determined initial deformation field, thereby improving the accuracy of registration of the first and second medical images, reducing the risk of intraoperative puncture, and improving the precision and safety of the surgery. Furthermore, the initial deformation field is determined using a pre-trained first target network model, and the initial deformation field is updated iteratively. This improves the efficiency of determining the deformation field corresponding to the registration result that meets the registration conditions. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of preoperative and intraoperative images in one embodiment;

[0069] Figure 2 This is a schematic diagram of the structure of a computer device in one embodiment;

[0070] Figure 3 This is a flowchart illustrating the steps of a registration method in one embodiment;

[0071] Figure 4 This is a flowchart illustrating the registration method in another embodiment;

[0072] Figure 5 This is a flowchart illustrating the registration method in another embodiment;

[0073] Figure 6 This is a flowchart illustrating the registration method in another embodiment;

[0074] Figure 7 This is a flowchart illustrating the registration method in another embodiment;

[0075] Figure 8 This is a flowchart illustrating the registration method in another embodiment;

[0076] Figure 9This is a flowchart illustrating the registration method in another embodiment;

[0077] Figure 10 This is a flowchart illustrating the registration method in another embodiment;

[0078] Figure 11 This is a flowchart illustrating the registration method in another embodiment;

[0079] Figure 12 This is an image schematic diagram and deformation field schematic diagram of each phase within a respiratory cycle in one embodiment;

[0080] Figure 13 This is a schematic diagram of the independent phase model training process in one embodiment;

[0081] Figure 14 This is a schematic diagram of the full-cycle phase model training process in one embodiment;

[0082] Figure 15 This is a schematic diagram of the continuous phase model training process in one embodiment;

[0083] Figure 16 This is a structural block diagram of a registration method in one embodiment;

[0084] Figure 17 Here is a structural block diagram of the registration method in another embodiment;

[0085] Figure 18 This is a flowchart illustrating the registration method in another embodiment;

[0086] Figure 19 This is a flowchart illustrating the registration method in another embodiment;

[0087] Figure 20 This is a training block diagram of the second target network model in one embodiment;

[0088] Figure 21 This is a schematic diagram of the registration device in one embodiment;

[0089] Figure 22 This is a schematic diagram of the registration device in another embodiment. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0091] The serial numbers assigned to components in this article, such as "first" and "second", are used only to distinguish the objects being described and have no sequential or technical meaning.

[0092] First, before introducing the technical solutions of the embodiments disclosed in this application, the background technology or technological evolution on which the embodiments of this application are based will be introduced. In the medical field, interventional surgery is a minimally invasive surgical procedure. Based on the patient's condition and imaging results (medical images), doctors use enhanced Magnetic Resonance Imaging (MRI) or Computed Tomography (CT) images to determine the location of the lesion within the body, and then formulate a detailed surgical plan. This allows doctors to use medical images and surgical instruments to accurately reach the lesion site for treatment or removal without traditional open surgery, by inserting long needles or other tools into the body cavity. However, in the actual surgical process, doctors usually rely on experience and tactile guidance, lacking accurate positioning and navigation. Furthermore, patients are usually under local anesthesia during the procedure, so respiratory movements can cause movement of the target point, organs, blood vessels, etc., leading to the risk of deviation, accidental damage to normal tissues or organs, and affecting treatment effectiveness and safety.

[0093] To overcome the aforementioned problems, traditional techniques have proposed image-guided robotic puncture surgery. Specifically, based on real-time CT imaging (CT Fluoroscopy), an extremely narrow area (approximately 10 mm) of the target region can be observed at the level of hundreds of milliseconds. Through master-slave operation of the robotic arm, the robotic arm can be controlled outside the radiology room, achieving faster and more accurate punctures under the guidance of real-time images.

[0094] However, due to the extremely narrow imaging range of real-time imaging (typically close to two-dimensional imaging thickness), surgeons can only observe a very narrow area around the target region, resulting in a weak overall three-dimensional perception during surgery. For example... Figure 1 As shown, Figure 1 The image on the left is a preoperative image of the diaphragm, the white area in the middle image is an image acquired during the operation, and the image on the right is a combination of multiple images acquired during the operation. For lesions significantly affected by respiration, such as the lung base or liver apex, the registration between the intraoperative two-dimensional image and the preoperative three-dimensional image becomes inaccurate due to the fluctuations in respiratory movements during the puncture process, making accurate puncture difficult and thus limiting the precision and safety of the procedure.

[0095] Traditional registration techniques commonly include grayscale-based methods, AI-based flexible registration methods, and motion modeling methods. For grayscale-based methods, traditional grayscale-based registration algorithms such as Demos and LDDMM (large deformation diffeomorphic metric mapping) are transferred to the registration of 3D and 2D images, defining a similarity measure in the 2D image space. This method relies on image grayscale, but intraoperative 2D images lose one dimension of grayscale information, leading to inaccurate registration. For AI-based flexible registration, intraoperative 2D images and preoperative 3D images are directly input into the AI ​​model. Through a series of encoding and decoding processes, a deformation field model is generated, used to deform the preoperative 3D image and calculate the similarity measure with the intraoperative 2D image. This process can be supervised or unsupervised. Supervised registration uses the deformation field as supervision, while unsupervised registration uses the image as supervision, without requiring prior acquisition of accurate deformation fields from the preoperative 3D and intraoperative 2D images. However, this registration method relies on a black-box deep learning registration approach, failing to consider the actual patterns of respiratory motion, leading to inaccurate registration. For motion modeling methods, principal component analysis is used to obtain the patient's motion model, and registration is performed based on this model and intraoperative 2D images. However, this registration method requires preoperative 4D images; registration is impossible in scenarios where preoperative 4D images are unavailable. Therefore, this application provides a registration method that improves registration accuracy.

[0096] The registration method provided in this application can be applied to computer devices, including but not limited to industrial computers, laptops, and tablets, etc., and the internal structure of the computer device can be as follows: Figure 2 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a registration method. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0097] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0098] In one embodiment, such as Figure 3 As shown, a registration method is provided. This embodiment illustrates the application of this method to a computer device. In this embodiment, the method includes the following steps:

[0099] Step 300: Acquire the first medical image and the second medical image.

[0100] The first medical image refers to a three-dimensional medical image of the object acquired preoperatively, while the second medical image refers to an approximately two-dimensional medical image of the object acquired intraoperatively. The modality of the first medical image can be an MR image obtained from a magnetic resonance (MR) scanner, a CT image obtained from a computed tomography (CT) scanner, or a PET image obtained from a positron emission tomography (PET) scanner. The modality of the second medical image can be the same as or different from that of the first medical image. Specifically, the first medical image is a CT image, and the second medical image is a CT image obtained based on real-time CT imaging (CTF, CT Fluoroscopy).

[0101] The computer device can acquire the first and second medical images from the post-processing workstation of the imaging equipment, or from a server such as a PACS (Picture Archiving and Communication Systems). Alternatively, the first and second medical images can be pre-stored in the computer device's memory, and the computer device can directly retrieve them from the memory when registration is required. This embodiment does not limit the specific method for acquiring the first and second medical images, as long as the function can be achieved.

[0102] In an optional embodiment, the acquired second medical image may include respiratory phase information of the second medical image. Typically, the phase of the first medical image is the first phase within a respiratory cycle, which can be represented as phase 00.

[0103] In an optional embodiment, after acquiring the first medical image, the computer device can segment the organs, blood vessels (such as skin, bones, target organs, arteries, veins, etc.) and lesions in the first medical image, and render and display the segmentation results.

[0104] Step 310: Generate low-dimensional deformation data based on the initial network, and input the low-dimensional deformation data into the first target network model to obtain the initial deformation field.

[0105] Low-dimensional deformation data refers to deformation data that can be deformed from a first medical image and then registered with a second medical image. Low-dimensional deformation data can be two-dimensional.

[0106] The initial network can be a convolutional neural network. This embodiment does not limit the structure and type of the initial network model, as long as it achieves its function. The first target network model refers to a pre-trained network model capable of reconstructing low-dimensional deformed data into high-dimensional deformed data. The dimension of the high-dimensional deformed data can be the same as the dimension of the first medical image. The type and kind of the first target network model can be the same as or different from the initial network. This embodiment does not limit the type and kind of the first target network model, as long as it achieves its function.

[0107] Computer equipment can generate low-dimensional deformation data through an initial network. This low-dimensional deformation data is then input into a first target network model. The initial deformation field can be obtained by reconstructing the low-dimensional deformation data through the first target network model.

[0108] Step 320: Register the first medical image and the second medical image based on the initial deformation field to obtain the registration result.

[0109] After obtaining the initial deformation field, the computer device deforms the first medical image based on this initial deformation field, and then registers the deformed first medical image with a second medical image to obtain a registration result. The registration result is used to characterize whether the deformed first and second medical images have been successfully registered. Specifically, the registration result can refer to the degree of registration between the deformed first and second medical images. This embodiment does not limit the specific registration method, as long as it can achieve its function.

[0110] Step 330: Adjust the parameters of the initial network and generate new low-dimensional deformation data based on the adjusted initial network. Return to the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration conditions.

[0111] The registration conditions correspond to the registration results. Registration conditions refer to the conditions for successful registration. If the registration result is the registration degree between the deformed first and second medical images, the registration condition is the registration degree threshold. If the registration degree is greater than the registration degree threshold, it means that the registration result meets the registration conditions. If the registration degree is less than or equal to the registration threshold, it means that the registration result does not meet the registration conditions.

[0112] After acquiring the registration result, the computer device can first compare the registration result with the registration conditions to determine whether the registration result meets the registration conditions. If the computer device determines that the registration result does not meet the registration conditions, it adjusts the parameters of the initial network to generate new low-dimensional deformation data based on the adjusted initial network. After obtaining the new low-dimensional deformation data, the computer device returns to step 310, which involves inputting the low-dimensional deformation data into the first target network model to obtain the initial deformation field, as well as steps 320 and 330. Alternatively, after acquiring the registration result, the computer device can directly adjust the parameters of the initial network to generate new low-dimensional deformation data based on the adjusted initial network, and then return to step 310, which involves inputting the low-dimensional deformation data into the first target network model to obtain the initial deformation field, as well as steps 320 and 330. Specifically, the computer device can adjust the parameters of the initial network based on the registration result. Repeat step 330 until the registration result meets the registration conditions. At this point, the registration between the first medical image and the second medical image is complete, which means that the first medical image is accurately registered.

[0113] The registration method provided in this application involves acquiring a first medical image and a second medical image; generating low-dimensional deformation data based on an initial network; inputting the low-dimensional deformation data into a first target network model to obtain an initial deformation field; registering the first and second medical images based on the initial deformation field to obtain a registration result; adjusting the parameters of the initial network; generating new low-dimensional deformation data based on the adjusted initial network; and returning to the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration conditions. In this embodiment, a pre-trained first target network model is used to transform the low-dimensional deformation data into an initial deformation field. The parameters of the initial network are adjusted based on the registration result, and new low-dimensional deformation data is generated based on the adjusted initial network. Returning to the step of inputting the low-dimensional deformation data into the first target network model, this iterative update of the initial deformation field improves the accuracy of the determined initial deformation field, thereby improving the accuracy of registering the first and second medical images. This reduces the risk of intraoperative puncture and improves the precision and safety of the surgery. Furthermore, the initial deformation field is determined using a pre-trained first target network model, and the initial deformation field is updated iteratively. This improves the efficiency of determining the deformation field corresponding to the registration result that meets the registration conditions.

[0114] In one embodiment, such as Figure 4 As shown, two implementation methods are involved in generating low-dimensional deformable data from the initial network. These implementation methods are as follows:

[0115] Step 400: Input the first medical image and the second medical image into the initial network to generate low-dimensional deformation data.

[0116] The computer device directly inputs the acquired first and second medical images into the initial network. The initial network determines the deformation field between the first and second medical images and compresses the deformation field into a low-dimensional potential space to obtain low-dimensional deformation data.

[0117] In an optional embodiment, the initial network is an encoder. Upon receiving the deformation field, the initial network first converts the high-dimensional deformation field into a low-dimensional deformation field (two-dimensional), and then processes the low-dimensional deformation field to generate low-dimensional deformation data. Specific processing methods may include interpolation, extrapolation, or representing the low-dimensional deformation field using a normal distribution.

[0118] Step 410: Input the first random noise into the initial network to generate low-dimensional deformation data.

[0119] The computer device acquires first random noise, inputs it into an initial network, and processes the random noise through the initial network to generate low-dimensional deformation data. The first random noise can be pre-stored in the computer device or randomly generated by the computer device.

[0120] In an optional embodiment, the initial network is a generator. Random noise can be used as a low-dimensional deformation field, and the initial network processes the random noise to generate low-dimensional deformation data.

[0121] In this embodiment, two implementation methods for generating low-dimensional deformation data from the initial network are provided. Users can choose according to their actual application, which can improve the applicability of the registration method.

[0122] In one embodiment, such as Figure 5 As shown, this relates to an implementation method for registering a first medical image and a second medical image based on an initial deformation field to obtain a registration result. The steps of this implementation method include:

[0123] Step 500: Deform the first medical image according to the initial deformation field to obtain a deformed medical image.

[0124] After obtaining the initial deformation field, the computer equipment deforms the first medical image according to the initial deformation field. In other words, the first medical image is transformed according to the initial deformation field, and the transformed medical image is the deformed medical image.

[0125] Step 510: Determine the similarity between the deformed medical image and the second medical image to obtain the registration result.

[0126] After obtaining a deformed medical image, the computer device calculates the similarity between the deformed medical image and a second medical image to obtain a registration result. The registration result represents the similarity between the deformed medical image and the second medical image. This embodiment does not limit the specific method for calculating the similarity between the deformed medical image and the second medical image, as long as the function can be achieved.

[0127] In this embodiment, the first medical image is deformed using an initial deformation field to obtain a deformed medical image; the similarity between the deformed medical image and the second medical image is calculated to determine the registration result. This method of registering the first and second medical images is quick and easy to implement, improving the efficiency and practicality of the registration method.

[0128] In one embodiment, such as Figure 6 As shown, a method for determining a first target network model includes the following steps:

[0129] Step 600: Obtain training samples; the training samples include image samples from different phases, and the first standard deformation field between image samples from different phases.

[0130] The training samples consist of multi-phase image data acquired over a complete respiratory cycle, from end-inspiratory to end-expiratory and back to end-inspiratory. The training samples are typically randomly acquired and consist of images that clearly characterize the continuous changes in image quality with respiratory motion noise. For example, images of the diaphragm.

[0131] Training samples can be pre-stored in the memory of a computer device, which can then retrieve them directly from the memory when the first target network model needs to be determined.

[0132] In an optional embodiment, after acquiring the training samples, the computer device may preprocess the training samples to ensure the quality and consistency of the training samples. Preprocessing includes, but is not limited to, normalization, alignment, and denoising.

[0133] Step 610: Train the first initial network model based on the training samples; the first initial network model includes an encoder and a decoder.

[0134] In this embodiment, the first initial network model is a variational autoencoder (VAE). The first initial network model includes an encoder and a decoder. After obtaining training samples, the computer device inputs the training samples into the first initial network model to train it.

[0135] In one embodiment, such as Figure 7 As shown, this relates to an implementation method for training an initial network model based on training samples. The steps of this implementation method include:

[0136] Step 700: Input the training samples into the encoder to obtain the first potential deformation data.

[0137] The computer equipment inputs image samples from different phases into the encoder. The encoder can determine the high-dimensional deformation field between different phases and compress the high-dimensional deformation into a low-dimensional potential space to obtain the first potential deformation data.

[0138] In an optional embodiment, after acquiring image samples from different phases, the computer device can first determine the initial deformation field samples between the image samples from different phases, and directly input the initial deformation field samples into the encoder to obtain the first potential deformation data.

[0139] Step 710: Sample the first potential deformation data to generate new first potential deformation data.

[0140] After obtaining the first potential deformation data, the computer device samples the first potential deformation data to generate new first potential deformation data.

[0141] In an optional embodiment, the computer device can express the first potential deformation data using a normal distribution to obtain new first potential deformation data. The computer device can also interpolate or extrapolate the first potential deformation data to generate new first potential deformation data.

[0142] Step 720: Input the new first potential deformation data into the decoder to obtain the first deformation field sample.

[0143] After obtaining new first potential deformation data, the computer device inputs the new first potential deformation data into the decoder. The decoder reconstructs the low-dimensional new first potential deformation data into high-dimensional deformation data, thus obtaining the first deformation field sample.

[0144] Step 730: Update the parameters of the encoder and the decoder based on the first deformation field sample and the first standard deformation field.

[0145] After obtaining the first deformation field sample, the computer device determines a loss function based on the first deformation field sample and the first standard deformation field, and updates the parameters of the encoder and decoder in the first initial network model based on the loss function. This embodiment does not limit the type of loss function, as long as it achieves its function.

[0146] In an alternative embodiment, the loss function can be KL divergence, mean squared error, etc.

[0147] In an optional embodiment, assuming the training samples input to the encoder include an initial deformation field sample x1 and a first standard deformation field X, inputting x1 into the encoder yields first latent deformation data, along with a mean μ and a variance σ. Sampling the first latent deformation data based on the mean μ and variance σ yields new first latent deformation data z1, z1 = μ + σε, where ε is noise sampled from a standard normal distribution. Inputting z1 into the decoder yields the first deformation field sample. According to X and Reconstruction error KL divergence can be expressed as... The total loss can be expressed as L1 = L recon +βL KL , where β is the weighting coefficient. The parameters of the encoder and decoder are updated based on the total loss.

[0148] In an optional embodiment, a sampling optimization algorithm (Adam optimizer) can be used to train the first target network model.

[0149] Step 620: If the first training stopping condition is met, the decoder in the first initial network model after training is determined as the first target network model.

[0150] The first training stopping condition can be either reaching a preset threshold number of iterations or the error function converging. This embodiment does not limit the specific content of the first training stopping condition, as long as its function is achieved.

[0151] During the training process of the first initial network model, if the computer device determines that the first training stopping condition is met, the decoder in the trained first initial network model is identified as the first target network model.

[0152] This embodiment provides a method for obtaining training samples to train the first initial network model when the first initial network model includes an encoder and a decoder, and then determining the decoder in the trained first initial network model as the first target network model. This method for determining the first target network model is easy to implement.

[0153] In one embodiment, such as Figure 8 As shown, another method for determining the first target network model includes the following steps:

[0154] Step 800: Obtain training samples, which include image samples from different phases and the first standard deformation field between image samples from different phases.

[0155] The computer device acquires training samples. A description of the training samples can be found in the specific description in the embodiments, and will not be repeated here.

[0156] Step 810: Train the second initial network model based on the training samples; the second initial network model includes a generator and a discriminator.

[0157] In this embodiment, the second initial network model is a Generative Adversarial Network (GAN). The second initial network model includes a generator and a discriminator. After obtaining training samples, the computer device inputs the training samples into the second initial network model to train it.

[0158] In one embodiment, such as Figure 9 As shown, this relates to an implementation method for training a second initial network model based on training samples. The steps of this implementation method include:

[0159] Step 900: Input the second random noise into the generator to generate the second potential deformation data.

[0160] The second random noise can be pre-stored in the computer device or randomly generated by the computer device. The second random noise can be low-dimensional random data.

[0161] After acquiring the second random noise, the computer device inputs the second random noise into a generator, which then generates the second potential deformation data. The second potential deformation data is also low-dimensional deformation data.

[0162] Step 910: Input the first standard deformation field and the second potential deformation data into the discriminator to obtain the discrimination probability.

[0163] After obtaining the second latent deformation data, the computer device inputs the second latent deformation data and the first standard deformation field into a discriminator. The discriminator can identify whether the second latent deformation data is consistent with the first standard deformation field, and obtain the discrimination probability. The discrimination probability is used to characterize the probability that the second latent deformation data generated by the generator is consistent with the first standard deformation field, and the probability that the second latent deformation data identified by the discriminator is consistent with the first standard deformation field.

[0164] Step 920: Update the parameters of the generator and the discriminator according to the discrimination probability.

[0165] After obtaining the discrimination probability, the computer device determines the loss function based on the discrimination probability, and updates the parameters of the generator and the discriminator based on the loss function.

[0166] In an optional embodiment, assuming the second random noise input to the generator is z2, inputting z2 into the generator yields the second potential deformation data. Will The discrimination probability can be obtained by inputting the first standard deformation field X into the discriminator. The discrimination probability can be expressed as: p real =Discriminator(X) and The loss function of the generator can be expressed as: L G =-log(p fake The loss function of the discriminator can be expressed as L. D =-log(p real )+log(1-p fake The generator parameters are updated based on the generator's loss function, and the discriminator parameters are updated based on the discriminator's loss function.

[0167] Step 820: If the second training stopping condition is met, the generator in the trained second initial network model is determined as the first target network model.

[0168] The description of the second training stop condition can be found in the description of the first training stop condition in the above embodiments, and will not be repeated here.

[0169] During the training process of the second initial network model, if the computer device determines that the second training stopping condition is met, it will identify the generator in the trained second initial network model as the first target network model.

[0170] This embodiment provides a method for training a second initial network model that includes a generator and a discriminator, and then determining the generator in the trained second initial network model as the first target network model. This method for determining the first target network model is easy to implement.

[0171] In one embodiment, such as Figure 10 As shown, another method for determining a first target network model is described, the steps of which include:

[0172] Step 101: Obtain training samples; the training samples include image samples from different phases, and the first standard deformation field between image samples from different phases.

[0173] The computer device acquires training samples. The description of the training samples can be found in the specific description in the above embodiments, and will not be repeated here.

[0174] Step 102: Train the third initial network model based on the training samples. The third initial network model includes a transformation function.

[0175] In this embodiment, the third initial network model is a flow-based model. The third initial network model includes a transformation function. After obtaining training samples, the computer device inputs the training network into the third initial network model to train it.

[0176] In one embodiment, such as Figure 11 As shown, this involves an implementation method for training a third initial network model based on training samples. The steps of this implementation method include:

[0177] Step 111: Input the first standard deformation field into the transformation function to obtain the third potential deformation data.

[0178] After acquiring training samples, the computer device inputs the first standard deformation field from the training samples into the transformation function so that the transformation function converts the high-dimensional standard deformation field into low-dimensional deformation data, namely the third potential deformation data.

[0179] Step 112: Sample the third potential deformation data to generate new third potential deformation data.

[0180] After obtaining the third potential deformation data, the computer device samples the third potential deformation data to generate new third potential deformation data. The description of sampling the third potential deformation data to generate new third potential deformation data can be found in the above embodiment's detailed description of sampling the first potential deformation data to generate new first potential deformation data, and will not be repeated here.

[0181] Step 113: Input the new third potential deformation data into the inverse function of the transformation function to obtain the second deformation field sample.

[0182] After obtaining the new third potential deformation data, the computer device inputs the new third potential deformation data into the inverse function of the transformation function. The inverse function of the transformation function reconstructs the new third potential deformation data into high-dimensional deformation data, namely the second deformation field sample.

[0183] Step 114: Update the parameters of the transformation function based on the second deformation field sample and the first standard deformation field.

[0184] After obtaining the second deformation field sample, the computer device can calculate the probability density based on the second deformation field sample and the first standard deformation field. The probability density is used to characterize the similarity between the second deformation field sample and the standard deformation field. A loss function is determined based on this probability density, and the parameters of the transformation function are updated based on this loss function.

[0185] In an optional embodiment, it is assumed that a first standard deformation field X is input to a transformation function f(X), and the third potential deformation data can be obtained through the transformation function f(X). By sampling the third potential deformation data, a new third potential deformation data z3 can be obtained; z3 is then input into the inverse function f of the transformation function. -1 (X) can be used to obtain the second deformation field sample. The loss function is the log-likelihood estimate. The loss function can be expressed as: L² = -logp(X), where p(X) is the probability density. The parameters of the transformation function are updated based on the loss function.

[0186] Step 103: Under the condition of satisfying the third training condition, determine the inverse function of the transformation function in the trained third initial network model as the first target network model.

[0187] The description of the third training stop condition can be found in the detailed description of the first training stop condition above, and will not be repeated here.

[0188] During the training of the third initial network model, if the computer device determines that the third training stopping condition is met, the inverse function of the transformation function in the trained initial network model is determined as the first target network model.

[0189] This embodiment provides a method for obtaining training samples to train the third initial network model when the third initial network model includes a transformation function, and determining the inverse function of the transformation function in the trained third initial network model as the first target network model. This method for determining the first target network model is easy to implement.

[0190] In an optional embodiment, the first target network model can also be a self-generating model, a variational inference model, a conditional generation model, an energy model, etc.

[0191] In one embodiment, the first target network model includes any one of the target phase model, the full-phase model, and the continuous phase model.

[0192] The target phase model is the phase model corresponding to the target period selected from multiple independent phase models. The independent phase model is used to describe the deformation field between one phase and another; the full phase model is used to describe the deformation field between any phases; and the continuous phase model is used to describe the continuous deformation field between one phase and another.

[0193] For example, such as Figure 12 As shown, taking the diaphragm image as an example, a respiratory cycle includes phases 00 to 90. The phase 00 at the end of inspiration is used as a shift and registered to the remaining phases. This allows us to obtain the deformation field between phase 00 and other phases, and to understand how the state moves from the end-inspiratory image (phase 00) to other phases. Figure 12 The reference lines in the image are two-dimensional medical images acquired during the procedure. Figure 12 The last line represents the deformation field from phase 00 to each phase.

[0194] For independent phase models, the models for each phase are trained using image samples corresponding to phase 00 and other phases, along with a standard deformation field. For example... Figure 13 As shown, based on the corresponding image samples and standard deformation fields of phases 00 and 10, an independent phase model (model 10) for deformation from phase 00 to phase 10 can be trained. Similarly, based on the corresponding image samples and standard deformation fields of phases 00 and 20, an independent phase model (model 20) for deformation from phase 00 to phase 20 can be trained. Using the same method, independent phase models for multiple phases can be obtained, such as model 30 and model 90. The above uses phase 00 as the moving phase. In practical applications, other phases can also be selected as the moving phase. For example, using phase 50 as the moving phase, the obtained training samples can be used to train an independent phase model (model 56) for deformation from phase 50 to phase 60, and an independent phase model (phase 57) for deformation from phase 50 to phase 70, etc.

[0195] In some embodiments, the application of independent phase models can be combined with a respiratory gating system to acquire respiratory status from preoperative and intraoperative images. The respiratory status can include at least one of respiratory phase, respiratory phase duration, and respiratory amplitude. For example, if the respiratory phase of the first medical image (preoperative image) acquired through the respiratory gating system is 00, and the respiratory phase of the second medical image (intraoperative image) is 50, an independent phase model (50 model) showing deformation from phase 00 to phase 50 can be used as the first target network model for initial deformation field calculation.

[0196] For a full-phase model, the deformation field between any two phases can be obtained, and this deformation field can be used to train the full-phase model. For example... Figure 14As shown, phase 1, phase 2, phase 3, ..., phase n represent the deformation field between any two phases. The deformation field between any two phases can be used to train the full phase model.

[0197] In some embodiments of the application of the full-phase model, it is not necessary to strictly obtain the respiratory status of preoperative and intraoperative images. Any preoperative or intraoperative image can call the full-phase model as the first target network model for initial deformation field calculation.

[0198] For continuous phase models, the standard deformation field in the training samples is obtained through state propagation. For example... Figure 15 As shown, 00-10 represents the deformation field from phase 00 to phase 10, 10-20 represents the deformation field between phase 10 and phase 20, 20-30 represents the deformation field between phase 20 and phase 30, and 80-90 represents the deformation field between phase 80 and phase 90. A continuous phase model can be trained using this continuous deformation field.

[0199] In some embodiments, the application of continuous phase models can be integrated with a respiratory gating system to acquire respiratory status data from preoperative and intraoperative images. The respiratory status can include at least one of respiratory phase, respiratory phase duration, and respiratory amplitude. For example, if the respiratory phase of the preoperative image acquired through the respiratory gating system is 00, and the respiratory phase of the intraoperative image is 50, a continuous phase model representing the deformation from phase 00 to phase 50, i.e., the 00-50 model, can be used as the first target network model for initial deformation field calculation. The 00-50 model describes the process of continuous deformation fields from 00-10, 10-20, 20-30, 30-40, and 40-50.

[0200] The target phase model, full-phase model, and continuous phase model can be any of the VAE, GAN, and flow cytometry generation models provided in the above embodiments. When the first target network model is a target phase model or a continuous phase model, the acquired second medical image needs to include respiratory phase information.

[0201] This embodiment describes a first target network model that may include at least one of a target phase model, a full-phase model, and a continuous phase model. This allows the user to choose the appropriate model based on the specific application, improving the applicability of the registration method. When the user selects a target phase model, a full-phase model, or a continuous phase model as the first target network model, multiple initial deformation fields can be obtained. The computer device can select the optimal initial deformation field from these multiple initial deformation fields, or it can average the multiple initial deformation fields to determine the final initial deformation field.

[0202] In one embodiment, Figure 16 and Figure 17 Structural diagrams for two registration methods are provided.

[0203] In one embodiment, when the first target network model is a target phase model, such as Figure 18 As shown, this involves an implementation method for inputting low-dimensional deformation data into a first target network model to obtain an initial deformation field. The steps of this implementation method include:

[0204] Step 181: Based on the target phase of the second medical image, select the target phase model from multiple independent phase models and determine the target phase model as the first target network model.

[0205] The acquired second medical image includes the target phase of the second medical image. After acquiring the target phase of the second medical image, the computer device selects a target phase model from multiple independent phase models based on the target phase and determines the target phase model as the first target network model.

[0206] Step 182: Input the low-dimensional deformation data into the first target network model to obtain the initial deformation field.

[0207] After determining the first target network model, the computer device inputs low-dimensional deformation data into the first target network model to obtain the initial deformation field. A description of inputting low-dimensional deformation data into the first target network model to obtain the initial deformation field can be found in the specific description in the above embodiments, and will not be repeated here.

[0208] Please see Figure 19 This application provides a registration method in one embodiment, which is illustrated by applying the method to a computer device. In this embodiment, the method includes the following steps:

[0209] Step 191: Acquire the first medical image and the second medical image.

[0210] The description of acquiring the first and second medical images can be found in the detailed description of the above embodiments, and will not be repeated here.

[0211] Step 192: Input the first medical image and the second medical image into the second target network model to determine the target deformation field; the second target network model includes a first encoding module and a second encoding module. Specifically, the second target network model is obtained by inputting the first image sample and the second image sample into the first encoding module to obtain the first deformation sample, inputting the second standard deformation field into the second encoding module to obtain the second deformation sample, and training the second target network model based on the first deformation sample and the second deformation sample.

[0212] This embodiment does not limit the type and structure of the second target network model, as long as it can achieve its function. Specifically, the second target network model is an encoder. The second target network model includes a first encoding module and a second encoding module.

[0213] The training process of the second target network model may include: acquiring training samples, which include a first image sample and a second image sample, as well as a second standard deformation field between the first and second image samples. The first image sample may be a preoperative image sample, and the second image sample may be an intraoperative image sample. The first and second image samples are input into a first encoding module, which determines a third deformation field between the first and second image samples and converts this third deformation field into low-dimensional first intermediate deformation data. This low-dimensional intermediate deformation data is then sampled to obtain the first deformation sample. The second standard deformation field is input into a second encoding module, which converts this high-dimensional second standard deformation field into low-dimensional second intermediate deformation data. This second intermediate deformation data is then sampled to obtain the second deformation sample. The parameters of the first and second encoding modules are updated using the first and second deformation samples to train the second target network model. Specifically, a consistency constraint is applied to the first and second deformation samples, ensuring that the first and second deformation samples are approximately identical through training.

[0214] In an optional embodiment, the training block diagram of the second target network model is as follows: Figure 20 As shown.

[0215] After acquiring the first and second medical images, the computer equipment inputs the first and second medical images into the second target network model to obtain the target deformation field.

[0216] Step 193: Register the second medical image and the first medical image according to the target deformation field.

[0217] After obtaining the target deformation field, the computer device deforms the first medical image according to the target deformation field, and then registers the deformed first medical image with the second medical image. This embodiment does not limit the specific registration method, as long as it can achieve its function.

[0218] In this embodiment, a first medical image and a second medical image are acquired; the first and second medical images are input into a second target network model to determine the target deformation field; the second target network model includes a first encoding module and a second encoding module; the second target network model is obtained by inputting the first image sample and the second image sample into the first encoding module to obtain a first deformation sample, and inputting the second standard deformation field into the second encoding module to obtain a second deformation sample; the second target network model is trained based on the first deformation sample and the second deformation sample; and the second medical image and the first medical image are registered according to the target deformation field. In this embodiment, the second target network model is updated based on the first deformation sample and the second deformation sample, that is, the parameters of the first encoding module and the second encoding module are updated according to the consistency constraint of the first deformation sample and the second deformation sample. The second target network model trained in this way can obtain a more accurate target deformation field, thereby enabling accurate registration of the first and second medical images based on the target deformation field, thus reducing the risk of intraoperative puncture and improving the accuracy and safety of the surgery.

[0219] One embodiment of this application provides an intraoperative guidance intervention system, including a processor, the processor being configured to perform the following steps:

[0220] S1. Acquire the first medical image and the second medical image. The first medical image is a preoperative image, and the second medical image is an intraoperative image.

[0221] The descriptions of the first and second medical images can be found in the detailed descriptions of the above embodiments, and will not be repeated here.

[0222] S2, according to the above Figures 2-11 , Figure 18 and Figure 19 The illustrated embodiment registers the first and second medical images to achieve intraoperative three-dimensional guidance.

[0223] The above embodiments can be used to register preoperative and intraoperative images, so that the preoperative images are closer to the actual intraoperative images, thereby enabling more accurate intraoperative three-dimensional guidance based on the preoperative images.

[0224] The intraoperative guided interventional system provided in this application includes a processor, which performs tasks such as acquiring a first medical image and a second medical image. The first medical image is a preoperative image, and the second medical image is an intraoperative image. Figures 2-11 , Figure 18 and Figure 19The steps shown register the first and second medical images to achieve real-time three-dimensional intraoperative guidance. The processor, using the registration method provided in the above embodiments, can improve the accuracy of the determined initial deformation field, thereby improving the accuracy of registration between preoperative and intraoperative images. This reduces the risk of intraoperative puncture and improves the precision and safety of the surgery when guiding intraoperative intervention based on the registered preoperative images.

[0225] In one embodiment, the intraoperative guidance system is used in conjunction with a respiratory gating system, which collects respiratory status data from preoperative and intraoperative images. That is, during the acquisition of preoperative and intraoperative images of the subject using imaging equipment, the respiratory gating system can simultaneously collect the respiratory movements of the subject, thereby determining the respiratory status of the subject corresponding to the preoperative images and the respiratory status of the subject corresponding to the intraoperative images. The respiratory status may include the respiratory phase corresponding to the preoperative and intraoperative images.

[0226] In cases where the first target network model includes a target phase model or a continuous phase model, preoperative and intraoperative images include respiratory phase information.

[0227] The descriptions of the target phase model and the continuous phase model can be found in the specific descriptions of the above embodiments. The selection of the target phase model is related to the phase of the processed image. When the first target network model is the target phase model, the acquired preoperative and intraoperative images carry respiratory phase information. For example, if the respiratory phase information carried by the preoperative image is phase 00 and the respiratory phase information carried by the intraoperative image is phase 20, then the selected target phase model is the phase 20 model.

[0228] The choice of the continuous phase model is also related to the phase of the processed images. When the first target network model is a continuous phase model, the acquired preoperative and intraoperative images carry respiratory phase information. Assuming that the respiratory phase information carried by the preoperative images is phase 00 and the respiratory phase information carried by the intraoperative images is phase 30, after inputting the preoperative and intraoperative images into the continuous phase model, the continuous phase model will first acquire the first deformation field between phase 00 and phase 10, the second deformation field between phase 10 and phase 20, and the third deformation field between phase 20 and phase 30. Based on the first, second, and third deformation fields, the deformation field between phase 00 and phase 30 can be determined, i.e., the initial deformation field.

[0229] In this embodiment, when the first target network model is a target phase model or a continuous phase model, the preoperative and intraoperative images need to include respiratory phase information. This allows the user to select the accurate first target network model based on the actual preoperative and intraoperative images, thereby obtaining an accurate initial deformation field.

[0230] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0231] Based on the same inventive concept, this application also provides a registration apparatus for implementing the registration method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more registration apparatus embodiments provided below can be found in the limitations of the registration method described above, and will not be repeated here.

[0232] In one embodiment, such as Figure 21 As shown, a registration device 10 is provided, including an input module 11, a first registration module 12, and a generation module 13, wherein:

[0233] Input module 11 is used to generate low-dimensional deformation data based on the initial network, and input the low-dimensional deformation data into the first target network model to obtain the initial deformation field;

[0234] The first registration module 12 is used to register the first medical image and the second medical image based on the initial deformation field to obtain the registration result;

[0235] The generation module 13 is used to adjust the parameters of the initial network and generate new low-dimensional deformation data based on the adjusted initial network. It then returns to the step of inputting the low-dimensional deformation data into the first target network model in the execution input module 11 until the registration result meets the registration conditions.

[0236] In one embodiment, the input module 11 is specifically used to input the first medical image and the second medical image into the initial network to generate the low-dimensional deformation data; or to input the first random noise into the initial network to generate the low-dimensional deformation data.

[0237] In one embodiment, the first registration module 12 is specifically used to deform the first medical image according to the initial deformation field to obtain a deformed medical image; determine the similarity between the deformed medical image and the second medical image to obtain a registration result.

[0238] In one embodiment, the registration device 10 further includes a first training module. The first training module is used to acquire training samples; the training samples include image samples from different phases and a first standard deformation field between image samples from different phases; train a first initial network model based on the training samples; the first initial network model includes an encoder and a decoder; and, if a first training stopping condition is met, determine the decoder in the trained first initial network model as the first target network model.

[0239] In one embodiment, the first training module is specifically used to input training samples into the encoder to obtain first latent deformation data; sample the first latent deformation data to generate new first latent deformation data; input the new first latent deformation data into the decoder to obtain first deformation field samples; and update the parameters of the encoder and the decoder according to the first deformation field samples and the first standard deformation field.

[0240] In one embodiment, the registration device 10 further includes a second training module. The second training module is used to acquire training samples, which include image samples from different phases and a first standard deformation field between the image samples from different phases; to train a second initial network model based on the training samples; the second initial network model includes a generator and a discriminator; and, if a second training stopping condition is met, to determine the generator in the trained second initial network model as the first target network model.

[0241] In one embodiment, the second training module is specifically used to input second random noise into the generator to generate second potential deformation data; input the first standard deformation field and the second potential deformation data into the discriminator to obtain the discrimination probability; and update the parameters of the generator and the discriminator according to the discrimination probability.

[0242] In one embodiment, the registration device 10 further includes a third training module. The third training module is used to acquire training samples; the training samples include image samples from different phases and a first standard deformation field between image samples from different phases; train a third initial network model based on the training samples, the third initial network model including a transformation function; and, when a third training stopping condition is met, determine the inverse function of the transformation function in the trained third initial network model as the first target network model.

[0243] In one embodiment, the third training module is specifically used to input the standard deformation field into the transformation function to obtain the third potential deformation data; sample the third potential deformation data to generate new third potential deformation data; input the new third potential deformation data into the inverse function of the transformation function to obtain the second deformation field sample; and update the parameters of the transformation function according to the second deformation field sample and the first standard deformation field.

[0244] In one embodiment, the first target network model includes any one of a target phase model, a full-phase model, and a continuous phase model; the target phase model is a phase model corresponding to the target phase selected from multiple independent phase models; the independent phase model is used to describe the deformation field between one phase and another; the full-phase model is used to describe the deformation field between any phases; and the continuous phase model is used to describe the continuous deformation field between one phase and another.

[0245] In one embodiment, the input module 11 is further configured to select a target phase model from multiple independent phase models based on the target phase of the second medical image, determine the target phase model as the first target network model, and input low-dimensional deformation data into the first target network model to obtain the initial deformation field.

[0246] In one embodiment, such as Figure 22 As shown, a registration device 20 is provided, including: an acquisition module 21, a determination module 22, and a second registration module 23, wherein:

[0247] The acquisition module 21 is used to acquire the first medical image and the second medical image;

[0248] The determination module 22 is used to input the first medical image and the second medical image into the second target network model to determine the target deformation field; the second target network model includes a first encoding module and a second encoding module; wherein, the second target network model obtains a first deformation sample by inputting the first image sample and the second image sample into the first encoding module, obtains a second deformation sample by inputting the second standard deformation field into the second encoding module, and trains the second target network model based on the first deformation sample and the second deformation sample;

[0249] The second registration module 23 is used to register the second medical image and the first medical image according to the target deformation field.

[0250] Each module in the above-mentioned registration device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0251] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0252] Low-dimensional deformation data is generated based on the initial network, and the low-dimensional deformation data is input into the first target network model to obtain the initial deformation field.

[0253] The first and second medical images are registered based on the initial deformation field to obtain the registration result.

[0254] Adjust the parameters of the initial network and generate new low-dimensional deformation data based on the adjusted initial network. Then, return to the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration conditions.

[0255] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0256] Acquire first and second medical images;

[0257] The first medical image and the second medical image are input into the second target network model to determine the target deformation field. The second target network model includes a first encoding module and a second encoding module. The second target network model is obtained by inputting the first image sample and the second image sample into the first encoding module to obtain the first deformation sample, inputting the second standard deformation field into the second encoding module to obtain the second deformation sample, and training the second target network model based on the first deformation sample and the second deformation sample.

[0258] The second medical image and the first medical image are registered based on the target deformation field.

[0259] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0260] Low-dimensional deformation data is generated based on the initial network, and the low-dimensional deformation data is input into the first target network model to obtain the initial deformation field.

[0261] The first and second medical images are registered based on the initial deformation field to obtain the registration result.

[0262] Adjust the parameters of the initial network and generate new low-dimensional deformation data based on the adjusted initial network. Then, return to the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration conditions.

[0263] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0264] Acquire first and second medical images;

[0265] The first medical image and the second medical image are input into the second target network model to determine the target deformation field. The second target network model includes a first encoding module and a second encoding module. The second target network model is obtained by inputting the first image sample and the second image sample into the first encoding module to obtain the first deformation sample, inputting the second standard deformation field into the second encoding module to obtain the second deformation sample, and training the second target network model based on the first deformation sample and the second deformation sample.

[0266] The second medical image and the first medical image are registered based on the target deformation field.

[0267] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0268] Low-dimensional deformation data is generated based on the initial network, and the low-dimensional deformation data is input into the first target network model to obtain the initial deformation field.

[0269] The first and second medical images are registered based on the initial deformation field to obtain the registration result.

[0270] Adjust the parameters of the initial network and generate new low-dimensional deformation data based on the adjusted initial network. Then, return to the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets the registration conditions.

[0271] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0272] Acquire first and second medical images;

[0273] The first medical image and the second medical image are input into the second target network model to determine the target deformation field. The second target network model includes a first encoding module and a second encoding module. The second target network model is obtained by inputting the first image sample and the second image sample into the first encoding module to obtain the first deformation sample, inputting the second standard deformation field into the second encoding module to obtain the second deformation sample, and training the second target network model based on the first deformation sample and the second deformation sample.

[0274] The second medical image and the first medical image are registered based on the target deformation field.

[0275] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0276] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0277] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method of registration, characterized by, The method comprises: generating low-dimensional deformation data according to an initial network, inputting the low-dimensional deformation data into a first target network model to obtain an initial deformation field; registering a first medical image and a second medical image based on the initial deformation field to obtain a registration result; adjusting parameters of the initial network, and generating new low-dimensional deformation data according to the adjusted initial network, and returning to the step of inputting the low-dimensional deformation data into the first target network model until the registration result meets a registration condition.

2. The method of claim 1, wherein, The low-dimensional deformation data generated according to the initial network comprises: inputting the first medical image and the second medical image into the initial network to generate the low-dimensional deformation data; or inputting a first random noise into the initial network to generate the low-dimensional deformation data.

3. The method of claim 1, wherein, The registration of the first medical image and the second medical image based on the initial deformation field to obtain a registration result comprises: deforming the first medical image according to the initial deformation field to obtain a deformed medical image; determining a similarity between the deformed medical image and the second medical image to obtain the registration result.

4. The method according to any one of claims 1 to 3, characterized in that, The method for determining the first target network model comprises: obtaining training samples; the training samples comprise image samples of different phases and a first standard deformation field between the image samples of different phases; training a first initial network model based on the training samples; the first initial network model comprises an encoder and a decoder; in a case where a first training stop condition is met, determining the decoder in the trained first initial network model as the first target network model.

5. The method of claim 4, wherein, The training of the initial network model based on the training samples comprises: inputting the training samples into the encoder to obtain first latent deformation data; sampling the first latent deformation data to generate new first latent deformation data; inputting the new first latent deformation data into the decoder to obtain a first deformation field sample; updating parameters of the encoder and parameters of the decoder according to the first deformation field sample and the first standard deformation field.

6. The method according to any one of claims 1-3, characterized in that, The method for determining the first target network model comprises: obtaining training samples; the training samples comprise image samples of different phases and a first standard deformation field between the image samples of different phases; training a second initial network model based on the training samples; the second initial network model comprises a generator and a discriminator; in a case where a second training stop condition is met, determining the generator in the trained second initial network model as the first target network model.

7. The method of claim 6, wherein, The training of the second initial network model based on the training samples comprises: inputting a second random noise into the generator to generate second latent deformation data; inputting the first standard deformation field and the second latent deformation data into the discriminator to obtain a discrimination probability; updating parameters of the generator and parameters of the discriminator according to the discrimination probability.

8. The method of any one of claims 1-3, wherein, The method for determining the first target network model comprises: obtaining training samples, the training samples including image samples of different phases and first standard deformation fields between the image samples of different phases; training a third initial network model based on the training samples, the third initial network model including a transformation function; in a case where a third training stop condition is met, determining an inverse function of the transformation function in the trained third initial network model as the first target network model.

9. The method of claim 8, wherein, The training of the third initial network model based on the training samples includes: inputting the standard deformation field into the transformation function to obtain third latent deformation data; sampling the third latent deformation data to generate new third latent deformation data; inputting the new third latent deformation data into the inverse function of the transformation function to obtain a second deformation field sample; updating parameters of the transformation function according to the second deformation field sample and the first standard deformation field.

10. The method of any one of claims 1-3, wherein, The first target network model includes any one of a target phase model, a full phase model and a continuous phase model; The target phase model is a phase model corresponding to a target phase selected from a plurality of independent phase models, the independent phase model is used to describe a deformation field between one phase and another phase; the full phase model is used to describe a deformation field between any phases; and the continuous phase model is used to describe a continuous deformation field between one phase and another phase.

11. The method of claim 10, wherein, The inputting of the low-dimensional deformation data into the first target network model to obtain an initial deformation field includes: selecting a target phase model from the plurality of independent phase models according to the target phase of the second medical image, and determining the target phase model as the first target network model; inputting the low-dimensional deformation data into the first target network model to obtain the initial deformation field.

12. A method of registration, characterized by, The method includes: obtaining a first medical image and a second medical image; inputting the first medical image and the second medical image into a second target network model to determine a target deformation field, the second target network model including a first encoding module and a second encoding module; wherein the second target network model obtains a first deformation sample by inputting a first image sample and a second image sample into the first encoding module, obtains a second deformation sample by inputting a second standard deformation field into the second encoding module, and is obtained by training the second target network model according to the first deformation sample and the second deformation sample; registering the second medical image and the first medical image according to the target deformation field.

13. An intraoperative guidance interventional system, characterized by The processor processes the following steps: obtaining a first medical image and a second medical image, the first medical image being a preoperative image, and the second medical image being an intraoperative image; registering the first medical image and the second medical image according to the registration method in any one of claims 1 to 11 to realize intraoperative three-dimensional real-time guidance.

14. The system of claim 13, wherein, For use with a respiratory gating system for acquiring a respiratory state of the preoperative image and the intraoperative image, the preoperative image and the intraoperative image comprising respiratory phase information in case the first target network model comprises a target phase model or a continuous phase model.

15. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, which when executed by a processor implements the steps of the method of any one of claims 1 to 12.

Images