A method and system for generating an ultrasound speed of sound image based on a T1-weighted MRI image
By generating CT and ultrasound velocity images through a staged network of a deep learning model, the radiation risk and velocity error caused by CT scanning in existing technologies are solved, and high-quality velocity images based on MRI images are generated, improving the accuracy and efficiency of treatment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENGDONG MEDICAL TECHNOLOGY (WUXI) CO LTD
- Filing Date
- 2025-04-11
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies rely on CT scans to obtain intracranial ultrasound velocity values, which leads to ionizing radiation risks and high costs. Furthermore, generating pseudo-CT images based on MRI results in increased errors in the velocity values.
By constructing a deep learning model, which is divided into a first-stage network to generate CT images and a second-stage network to generate ultrasound velocity images, the model is trained using paired data of T1-weighted MRI images and CT images, avoiding additional CT scans and achieving the fusion of modal information.
Without increasing the radiation dose, the speed of sound in the skull can be accurately obtained, improving the precision and efficiency of sonodynamic therapy and reducing the risk of ionizing radiation.
Smart Images
Figure CN120374769B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical image generation technology, and in particular to a method and system for generating ultrasound velocity images based on T1-weighted MRI images. Background Technology
[0002] Ultrasound velocity images reflect the characteristics of tissue's response to ultrasound waves and are important parameters for applications such as ultrasound imaging and ultrasound therapy. In sonodynamic therapy for brain tumors, accurate intracranial ultrasound velocity distribution is crucial for sound field simulation, energy focusing, and optimization of treatment parameters. Inaccurate velocity information can lead to inaccurate energy focusing, affecting treatment efficacy and even causing adverse reactions.
[0003] Traditional methods for acquiring transcranial ultrasound sound velocity primarily rely on computed tomography (CT) scans of the brain. This involves estimating the sound velocity at the corresponding skull location using the Hue value (HU) of the CT images, and then obtaining the sound velocity value for the corresponding soft tissue in the brain using T1-weighted MRI (Magnetic Resonance Imaging) images. However, this method, dependent on CT scans, exposes patients to unnecessary ionizing radiation and results in high treatment costs. To avoid ionizing radiation, existing deep learning-based methods typically generate pseudo-CT images from MRI scans and then use these pseudo-CT images to obtain the corresponding sound velocity values. However, the error between the pseudo-CT images and the real CT images further increases the error in the sound velocity value.
[0004] Therefore, how to accurately obtain the sound velocity value of the skull from MRI images without increasing the additional radiation dose is a problem that urgently needs to be solved in this field. Summary of the Invention
[0005] To at least partially overcome the problem in related technologies where generating pseudo-CT images based on MRI leads to a further increase in the error of sound velocity values, this application provides a method and system for generating ultrasound sound velocity images based on T1-weighted MRI images.
[0006] The proposed solution is as follows:
[0007] According to a first aspect of the embodiments of this application, a method for generating ultrasound velocity images based on T1-weighted MRI images is provided, comprising:
[0008] T1-weighted MRI and CT images of historical subjects were acquired as sample data;
[0009] Construct a deep learning model, which includes a first-stage network and a second-stage network;
[0010] The first-stage network is trained using the sample data, so that when a T1-weighted MRI image is input, the first-stage network generates a CT image.
[0011] The second-stage network is trained using the sample data and the CT images generated by the first-stage network, so that the second-stage network can generate ultrasound velocity images when T1-weighted MRI images and CT images are input.
[0012] Preferably, the method further includes:
[0013] Acquire T1-weighted MRI images of the subjects;
[0014] The subject's T1-weighted MRI images are input into the deep learning model;
[0015] The first-stage network of the deep learning model generates the subject's CT image based on the subject's T1-weighted MRI image;
[0016] The second-stage network of the deep learning model generates the subject's ultrasound velocity image based on the subject's T1-weighted MRI image and the subject's CT image generated by the first-stage network.
[0017] Preferably, after acquiring T1-weighted MRI and CT images of historical subjects, the method further includes:
[0018] Data anonymization was performed on T1-weighted MRI and CT images of historical subjects.
[0019] Preferably, after acquiring T1-weighted MRI and CT images of historical subjects, the method further includes:
[0020] T1-weighted MRI and CT images of historical subjects were resampled to unify the resolution of T1-weighted MRI and CT images of historical subjects;
[0021] The CT data of historical subjects were rigidly registered to the T1-weighted MRI data of the resampled historical subjects.
[0022] Preferably, it further includes:
[0023] Based on the rigidly registered CT data and T1-weighted MRI data of historical subjects, the ultrasound velocity images of each pixel were determined as the true ultrasound velocity images.
[0024] Preferably, training the first-stage network using the sample data includes:
[0025] Features of T1-weighted MRI images were extracted using a single-branch network;
[0026] Generating simulated CT images based on extracted features;
[0027] The similarity between the generated simulated CT images and real CT images is determined; the real CT images are CT images of historical subjects.
[0028] Training is complete when the similarity between the generated simulated CT image and the real CT image is higher than a preset threshold.
[0029] Preferably, training the second-stage network using the sample data and the CT images generated by the first-stage network includes:
[0030] Features of T1-weighted MRI images are extracted using the first branch of a two-branch fusion network;
[0031] Features of real CT images are extracted using the second branch of a two-branch fusion network;
[0032] The features extracted from the first and second branches are fused using an attention mechanism, and an ultrasonic velocity image is generated based on the fused features.
[0033] Preferably, training the second-stage network using the sample data and the CT images generated by the first-stage network further includes:
[0034] Determine whether the network structures of the first-stage network and the second-stage network are stable;
[0035] If the network structures of the first-stage network and the second-stage network reach the preset stability condition, the CT image generated by the first-stage network is used as the input of the second branch of the two-branch fusion network, so that the second branch of the two-branch fusion network extracts the features of the CT image generated by the first-stage network.
[0036] The features extracted from the first and second branches are fused using an attention mechanism, and an ultrasonic velocity image is generated based on the fused features.
[0037] Distinguish the similarity between the generated ultrasonic velocity image and the real ultrasonic velocity image;
[0038] Training is complete when the similarity between the generated ultrasonic velocity image and the real ultrasonic velocity image is higher than a preset threshold.
[0039] Preferably, it further includes:
[0040] The mean square absolute error index was used to evaluate the overall difference between the CT images generated by the first-stage network and the real CT images.
[0041] The degree of overlap between the skull region in the CT images generated by the first-stage network and the real CT images was evaluated using the Jaccard coefficient and the Dice coefficient.
[0042] The overall difference between the ultrasound velocity images generated by the second-stage network and the real ultrasound velocity images was evaluated using MAE.
[0043] According to a second aspect of the embodiments of this application, an ultrasound velocity image generation system based on T1-weighted MRI images is provided, comprising:
[0044] Processor and memory;
[0045] The processor and memory are connected via a communication bus:
[0046] The processor is used to call and execute the program stored in the memory;
[0047] The memory is used to store a program, which is at least used to execute a method for generating ultrasound velocity images based on T1-weighted MRI images as described in any of the preceding claims.
[0048] The technical solution provided in this application may include the following beneficial effects:
[0049] This technical solution utilizes historical T1-weighted MRI and CT images of subjects as sample data to construct a training dataset. By using existing paired data, additional CT scans are avoided, thus reducing the risk of ionizing radiation, while ensuring that the training data includes accurate acoustic velocity characteristics of the skull and soft tissues. The deep learning model is divided into a first-stage network and a second-stage network. The first-stage network generates CT images, and the second-stage network generates ultrasound acoustic velocity images. The first stage generates CT images using T1-weighted MRI, fully utilizing the soft tissue information of MRI and the advantages of CT in bone imaging to provide accurate anatomical references for subsequent acoustic velocity mapping. The second stage generates ultrasound acoustic velocity images based on T1-weighted MRI and CT images, achieving the fusion of information from both modalities. This effectively generates high-quality ultrasound acoustic velocity images, avoiding ionizing radiation from CT scans, and can be used to plan the control of ultrasound energy in sonodynamic therapy, thereby improving the accuracy and efficiency of treatment.
[0050] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0051] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0052] Figure 1This is a schematic flowchart of a method for generating ultrasound velocity images based on T1-weighted MRI images according to an embodiment of this application;
[0053] Figure 2 This is a schematic diagram of the deep learning model network structure in an ultrasound velocity image generation method based on T1-weighted MRI images provided in one embodiment of this application;
[0054] Figure 3 This is a schematic diagram of the GF structure of a deep learning model in an ultrasound velocity image generation method based on T1-weighted MRI images provided in one embodiment of this application;
[0055] Figure 4 This is a schematic diagram of the second-stage network structure of a deep learning model in an ultrasound velocity image generation method based on T1-weighted MRI images provided in one embodiment of this application;
[0056] Figure 5 This application provides a method for generating ultrasound velocity images based on T1-weighted MRI images using a deep learning model to generate CT images and ultrasound velocity images, as provided in one embodiment of the application.
[0057] Figure 6 This is a sound field simulation diagram for acoustic dynamics therapy in an ultrasound velocity image generation method based on T1-weighted MRI images provided in one embodiment of this application;
[0058] Figure 7 This is a schematic diagram of the structure of an ultrasound velocity image generation system based on T1-weighted MRI images provided in one embodiment of this application.
[0059] Reference numerals: Processor-21; Memory-22. Detailed Implementation
[0060] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0061] Example 1
[0062] A method for generating ultrasound velocity images based on T1-weighted MRI images, comprising:
[0063] S11: Acquire T1-weighted MRI and CT images of historical subjects as sample data;
[0064] It should be noted that after obtaining the T1-weighted MRI and CT images of historical subjects, the method also includes:
[0065] Data anonymization was performed on T1-weighted MRI and CT images of historical subjects.
[0066] In practice, T1-weighted MRI and CT data of the brain are collected from multiple (e.g., 100) subjects. This data is obtained in a clinical setting using standard scanning protocols. Furthermore, this data undergoes anonymization to ensure patient privacy.
[0067] T1-weighted MRI of the brain utilizes magnetic resonance imaging (MRI) technology to highlight differences in T1 relaxation times among different tissues by selecting specific echo times (TE) and repetition times (TR). This imaging method primarily reflects the recovery rates of fat and water molecules in tissues, thus clearly distinguishing soft tissue structures. It is mainly used for detailed imaging of brain structures, disease diagnosis (such as tumors, brain atrophy, etc.), and surgical planning.
[0068] The T1-weighted MRI images in this embodiment were acquired using a 3.0T magnetic resonance scanner and a T1-weighted sequence. The scanning parameters included: TR (repetition time) = 500ms, TE (echo time) = 10ms, field of view = 220mm x 220mm, matrix size = 256x 256, and slice thickness = 1mm.
[0069] CT imaging uses X-ray projection data from multiple angles, which is then processed by a computer to reconstruct a three-dimensional image. It is commonly used in examinations of craniocerebral trauma, fractures, calcified lesions, and angiography.
[0070] The CT images in this embodiment were acquired using a multi-slice spiral CT scanner. The scanning parameters included: tube voltage = 120kVp, tube current = 200mA, and slice thickness = 1mm.
[0071] It should be noted that after obtaining the T1-weighted MRI and CT images of historical subjects, the method also includes:
[0072] T1-weighted MRI and CT images of historical subjects were resampled to unify the resolution of T1-weighted MRI and CT images of historical subjects;
[0073] The CT data of historical subjects were rigidly registered to the T1-weighted MRI data of the resampled historical subjects.
[0074] In practice, linear interpolation or trilinear interpolation is used for resampling of MRI and CT images to unify them to a voxel size of 1mm x 1mm x 1mm, so that they have the same resolution. The size of all slice images is 256x256.
[0075] CT data were rigidly registered to resampled T1-weighted MRI data. Mutual information was used as the similarity measure for registration, and registration was performed using either the Iterative Closest Point (ICP) algorithm or a gradient descent-based optimization algorithm.
[0076] Preprocessing T1-weighted MRI and CT data to ensure they have the same resolution, and rigidly registering CT data to the resampled T1-weighted MRI data, are primarily to ensure complete spatial alignment of the two modalities, thereby achieving accurate pixel-level mapping and multimodal information fusion. The specific reasons are as follows:
[0077] MRI and CT scans may differ in acquisition parameters, pixel size, and resolution. Without a unified resolution, the same anatomical structures across different modalities may appear as different sizes or scales in images, leading to a lack of one-to-one pixel correspondence and impacting subsequent deep learning model training and image fusion. When the resolution is consistent, deep learning models can capture the same structural information at the same scale when extracting features, helping the model better learn the mapping relationship between the two modalities. Unifying the resolution is a crucial step in data preprocessing, reducing interference from differences in data sources, acquisition equipment, and parameters, thereby improving model robustness.
[0078] Different imaging modalities (MRI and CT) often exhibit spatial geometric differences such as translation and rotation due to factors like patient position and equipment angle during acquisition. Rigid registration can correct these translational and rotational errors, ensuring accurate overlap of identical anatomical structures in the two images. Rigid registration guarantees spatial correspondence between bone and calcified regions in CT images and soft tissue structures in MRI images, providing accurate anatomical references for subsequent pseudo-CT generation and ultrasound velocity mapping using CT information. In deep learning tasks, especially pixel-level image generation and conversion, accurate alignment of input data is a prerequisite for ensuring model performance. Registered data significantly reduces noise and errors, helping the model learn the mapping relationships between different modalities more effectively.
[0079] It should be noted that this also includes:
[0080] Based on the rigidly registered CT data and T1-weighted MRI data of historical subjects, the ultrasound velocity images of each pixel were determined as the true ultrasound velocity images.
[0081] By combining information from registered CT images and T1-weighted MRI images, the ultrasound velocity value corresponding to each pixel is accurately determined, providing real comparison data for the validation stage in subsequent model training.
[0082] S12: Construct a deep learning model, which includes a first-stage network and a second-stage network;
[0083] Figure 2 This is a schematic diagram of the deep learning model network structure in an ultrasound velocity image generation method based on T1-weighted MRI images provided in one embodiment of this application; Figure 3 This is a schematic diagram of the GF structure of a deep learning model in an ultrasound velocity image generation method based on T1-weighted MRI images provided in one embodiment of this application; Figure 4 This is a schematic diagram of the second-stage network structure of a deep learning model in an ultrasound velocity image generation method based on T1-weighted MRI images provided in one embodiment of this application.
[0084] like Figure 2 As shown, the first-stage network reference Figure 2 The Stage One diagram in the image shows the second-stage network reference. Figure 2 The Stage Two diagram.
[0085] The specific structure of the second-stage network is referenced. Figure 4 .
[0086] S13: Train the first-stage network using sample data so that the first-stage network can generate CT images when T1-weighted MRI images are input;
[0087] Specifically, the first-stage network is trained using sample data, including:
[0088] Features of T1-weighted MRI images were extracted using a single-branch network;
[0089] Generating simulated CT images based on extracted features;
[0090] The similarity between the generated simulated CT images and real CT images is determined; the real CT images are CT images of historical subjects.
[0091] Training is complete when the similarity between the generated simulated CT image and the real CT image is higher than a preset threshold.
[0092] Specifically, the first-stage network generates simulated CT images (pseudo-CT images). It employs a U-Net structure to construct the generator, including skip connections to preserve local anatomical features. Specifically, it includes a generator G and a discriminator D. The generator G uses an encoder-decoder structure. The encoder consists of multiple convolutional and pooling layers to extract features from the MRI image. The decoder consists of multiple deconvolutional and upsampling layers to generate the pseudo-CT image. The generator uses ReLU as the activation function and a Sigmoid function in the last layer to restrict the pixel values of the pseudo-CT image to [0,1]. The discriminator D consists of five convolutional layers and activation functions to determine whether the input image is synthetic or real, distinguishing between the generated pseudo-CT image and the real CT image. It uses a convolutional neural network structure and outputs a probability value representing the probability that the input image is a real CT image.
[0093] The generated pseudo-CT images and real CT images of historical subjects are fed into the discriminator D of the first-stage network. A loss function is set, and the loss function value is output. Based on the loss function value, gradient descent is performed, and multiple iterations are conducted to optimize and obtain the optimal generator for synthesized CT images.
[0094] S14: Train the second-stage network using sample data and CT images generated by the first-stage network, so that the second-stage network can generate ultrasound velocity images when T1-weighted MRI images and CT images are input.
[0095] Specifically, the second-stage network is trained using sample data and CT images generated by the first-stage network, including:
[0096] Features of T1-weighted MRI images are extracted using the first branch of a two-branch fusion network;
[0097] Features of real CT images are extracted using the second branch of a two-branch fusion network;
[0098] The features extracted from the first and second branches are fused using an attention mechanism, and an ultrasonic velocity image is generated based on the fused features.
[0099] Determine whether the network structures of the first-stage network and the second-stage network are stable;
[0100] If the network structures of the first-stage network and the second-stage network reach the preset stability condition, the CT image generated by the first-stage network is used as the input of the second branch of the two-branch fusion network, so that the second branch of the two-branch fusion network extracts the features of the CT image generated by the first-stage network.
[0101] The features extracted from the first and second branches are fused using an attention mechanism, and an ultrasonic velocity image is generated based on the fused features.
[0102] Distinguish the similarity between the generated ultrasonic velocity image and the real ultrasonic velocity image;
[0103] Training is complete when the similarity between the generated ultrasonic velocity image and the real ultrasonic velocity image is higher than a preset threshold.
[0104] The second stage generates ultrasound velocity images. The second-stage network uses a U-Net structure to construct a two-branch fusion network generator. Each branch employs an independent downsampling encoder, and a unified upsampling decoder is used after the bottom-level fusion. Features of the T1-weighted MRI image are extracted through the first downsampling branch of the two-branch fusion network; features of the generated CT image are extracted through the second downsampling branch of the two-branch fusion network.
[0105] In the U-Net structure generator, features extracted from the first and second branches are fused using a cross-attention mechanism in skip connections and at the bottom layer of the network. The calculation method is as follows: attention-weighted CT features are added element-wise with the original MRI features to achieve intermodal feature calibration and fusion, and finally generate ultrasound velocity images.
[0106] The generator GF structure in this stage is a two-branch fusion structure, such as... Figure 3 As shown in the diagram. The encoder structure for the MRI branch is used to extract features from MRI images, while the pseudo-CT branch structure is used to extract features from pseudo-CT images. The two encoders are similar but have different parameters. Following the encoders, an attention mechanism is used to fuse the features from the MRI and pseudo-CT branches. The structure of the attention mechanism is shown below. Figure 4 As shown, the features after virtual-real fusion are connected to the decoder, which outputs the generated ultrasonic velocity image. The discriminator (DF) in this part is used to distinguish the generated ultrasonic velocity image from the real ultrasonic velocity image.
[0107] The second-stage discriminator is a patch discriminator, consisting of five convolutional layers and activation functions, used to determine whether the input image is a synthetic image or a real image.
[0108] The generated ultrasonic velocity image and the real ultrasonic velocity image are fed into the second-stage discriminator. A loss function is set, and the loss function value is output. Based on the loss function value, gradient descent is performed, and multiple iterations are conducted to optimize the generator and obtain the optimal ultrasonic velocity image generator.
[0109] It should be noted that in this technical solution, after the network structure reaches stability in both stages, serial fine-tuning is performed. The CT images generated in the first stage are replaced with the real CT images from the second stage training process.
[0110] In this technical solution, the deep learning model is divided into a first-stage network and a second-stage network. The first-stage network is used to generate CT images, and the second-stage network is used to generate ultrasound velocity images.
[0111] The goal of the first-stage network is to make the generated simulated CT images as similar as possible to real CT images.
[0112] The goal of the second-stage network is to generate ultrasonic velocity images based on the input MRI images of the subject and the simulated CT images generated by the first-stage network.
[0113] During training, the standard loss function for generative adversarial networks is used, including generator loss and discriminator loss.
[0114] In the first-stage network, the overall difference between the CT images generated by the first-stage network and the real CT images is evaluated by the mean square absolute error index.
[0115] The degree of overlap between the CT images generated by the first-stage network and the skull region in the real CT images was evaluated using the Jaccard coefficient and Dice coefficient.
[0116] In the second-stage network, MAE is used to evaluate the difference between the generated ultrasound velocity images and the real ultrasound velocity images.
[0117] This technical solution utilizes historical T1-weighted MRI and CT images of subjects as sample data to construct a training dataset. By using existing paired data, additional CT scans are avoided, thus reducing the risk of ionizing radiation, while ensuring that the training data includes accurate acoustic velocity characteristics of the skull and soft tissues. The deep learning model is divided into a first-stage network and a second-stage network. The first-stage network generates CT images, and the second-stage network generates ultrasound acoustic velocity images. The first stage generates CT images using T1-weighted MRI, fully utilizing the soft tissue information of MRI and the advantages of CT in bone imaging to provide accurate anatomical references for subsequent acoustic velocity mapping. The second stage generates ultrasound acoustic velocity images based on T1-weighted MRI and CT images, achieving the fusion of information from both modalities. This effectively generates high-quality ultrasound acoustic velocity images, avoiding ionizing radiation from CT scans, and can be used to plan the control of ultrasound energy in sonodynamic therapy, thereby improving the accuracy and efficiency of treatment.
[0118] It should be noted that the model can be used immediately after training. Based on this, the method also includes:
[0119] Acquire T1-weighted MRI images of the subjects;
[0120] The subjects' T1-weighted MRI images were input into the deep learning model;
[0121] The first-stage network of the deep learning model generates CT images of the subject based on the subject's T1-weighted MRI images;
[0122] The second-stage network of the deep learning model generates the subject's ultrasound velocity image based on the subject's T1-weighted MRI image and the subject's CT image generated by the first-stage network.
[0123] Figure 5 This application provides an embodiment of a method for generating ultrasound velocity images based on T1-weighted MRI images, which uses a deep learning model to generate CT images and ultrasound velocity images. Figure 6 This is a simulation diagram of the sound field for acoustic dynamics therapy in an ultrasound sound velocity image generation method based on T1-weighted MRI images, provided in one embodiment of this application. Figures 5-6 The trained deep learning model can automatically generate ultrasound velocity images of the subject after inputting the subject's T1-weighted MRI images, avoiding the need for additional CT scans and thus reducing the risk of ionizing radiation, while ensuring that the training data contains the true sound velocity characteristics of the skull and soft tissues.
[0124] Example 2
[0125] A system for generating ultrasound velocity images based on T1-weighted MRI images, referring to Figure 7 ,include:
[0126] Processor 21 and memory 22;
[0127] Processor 21 and memory 22 are connected via a communication bus:
[0128] The processor 21 is used to call and execute the program stored in the memory 22;
[0129] The memory 22 is used to store a program, which is at least used to execute a method for generating ultrasound velocity images based on T1-weighted MRI images as described in the above embodiments.
[0130] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.
[0131] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.
[0132] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.
[0133] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0134] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.
[0135] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0136] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.
[0137] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0138] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.
Claims
1. A method for generating ultrasound velocity images based on T1-weighted MRI images, characterized in that, include: T1-weighted MRI and CT images of historical subjects were acquired as sample data; Construct a deep learning model, which includes a first-stage network and a second-stage network; The first-stage network is trained using the sample data, so that when a T1-weighted MRI image is input, the first-stage network generates a CT image. The second-stage network is trained using the sample data and the CT images generated by the first-stage network, so that the second-stage network can generate ultrasound velocity images when T1-weighted MRI images and CT images are input. After acquiring T1-weighted MRI and CT images of historical subjects, the method further includes: T1-weighted MRI and CT images of historical subjects were resampled to unify the resolution of T1-weighted MRI and CT images of historical subjects; The CT data of historical subjects were rigidly registered to the T1-weighted MRI data of the resampled historical subjects; Based on the rigidly registered CT data and T1-weighted MRI data of historical subjects, the ultrasound velocity images of each pixel were determined as the true ultrasound velocity images. Training the first-stage network using the sample data includes: Features of T1-weighted MRI images were extracted using a single-branch network; Generating simulated CT images based on extracted features; The similarity between the generated simulated CT images and real CT images is determined; the real CT images are CT images of historical subjects. Training is complete when the similarity between the generated simulated CT image and the real CT image is higher than a preset threshold. Training the second-stage network using the sample data and the CT images generated by the first-stage network includes: Features of T1-weighted MRI images are extracted using the first branch of a two-branch fusion network; Features of real CT images are extracted using the second branch of a two-branch fusion network; The features extracted from the first and second branches are fused using an attention mechanism, and an ultrasonic velocity image is generated based on the fused features. Determine whether the network structures of the first-stage network and the second-stage network are stable; If the network structures of the first-stage network and the second-stage network reach the preset stability condition, the CT image generated by the first-stage network is used as the input of the second branch of the two-branch fusion network, so that the second branch of the two-branch fusion network extracts the features of the CT image generated by the first-stage network. The features extracted from the first and second branches are fused using an attention mechanism, and an ultrasonic velocity image is generated based on the fused features. Distinguish the similarity between the generated ultrasonic velocity image and the real ultrasonic velocity image; Training is complete when the similarity between the generated ultrasonic velocity image and the real ultrasonic velocity image is higher than a preset threshold.
2. The method according to claim 1, characterized in that, The method further includes: Acquire T1-weighted MRI images of the subjects; The subject's T1-weighted MRI images are input into the deep learning model; The first-stage network of the deep learning model generates the subject's CT image based on the subject's T1-weighted MRI image; The second-stage network of the deep learning model generates the subject's ultrasound velocity image based on the subject's T1-weighted MRI image and the subject's CT image generated by the first-stage network.
3. The method according to claim 1, characterized in that, After acquiring T1-weighted MRI and CT images of historical subjects, the method further includes: Data anonymization was performed on T1-weighted MRI and CT images of historical subjects.
4. The method according to claim 1, characterized in that, Also includes: The mean square absolute error index was used to evaluate the overall difference between the CT images generated by the first-stage network and the real CT images. The degree of overlap between the skull region in the CT images generated by the first-stage network and the real CT images was evaluated using the Jaccard coefficient and the Dice coefficient. The overall difference between the ultrasound velocity images generated by the second-stage network and the real ultrasound velocity images was evaluated using MAE.
5. A system for generating ultrasonic velocity images based on T1-weighted MRI images, characterized in that, include: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute the program stored in the memory; The memory is used to store a program, which is at least used to execute the method for generating ultrasound velocity images based on T1-weighted MRI images as described in any one of claims 1-4.