Non-contrast super-resolution ultrasonic micro blood flow dynamic imaging method and system

Abstract

The invention discloses a non-radiography super-resolution ultrasonic micro blood flow dynamic imaging method and system, and belongs to the technical field of ultrasonic blood flow imaging. The method comprises the steps that a data pair including an ultrafast power Doppler image and an ultrasonic positioning microscope image is acquired; the method comprises the following steps: constructing a U-MambaIR based on Mama; defining a joint loss function, and setting a weight parameter of each loss item; training to generate a reconstructed super-resolution power Doppler image; the ultra-fast power Doppler image is input into the trained ultrasonic blood vessel image recovery model, low-frequency features are extracted through the shallow feature extraction module, high-frequency features are extracted through the deep feature extraction module, the low-frequency features and the high-frequency features are aggregated through the super-resolution image reconstruction module, and a non-contrast super-resolution ultrasonic micro blood flow dynamic imaging result is output. The non-contrast super-resolution micro blood flow dynamic imaging method has strong feature extraction and space detail recovery capabilities, is strong in model generalization performance, realizes non-contrast super-resolution micro blood flow dynamic imaging, and has an application prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of ultrasonic blood flow imaging, and in particular relates to a contrast-free super-resolution ultrasonic micro-blood flow dynamic imaging method and system. Background Art

[0002] Ultrasound localization microscopy (ULM) breaks the acoustic diffraction limit by isolating sparsely located microbubbles as subwavelength sources within each image. It then locates these microbubbles with micron-level precision and tracks their trajectory across tens of thousands of image frames, enabling micron-scale reconstruction of microvascular structures. The development of ULM technology has significant clinical value in the diagnosis and treatment of vascular diseases, including cardiovascular disease, stroke, and chronic kidney disease.

[0003] However, the use of microbubble contrast agents, the long acquisition and processing times of ULM, and the low acoustic pressure and stable microbubble concentrations limit the practical application of ULM technology. Specifically, clinically used microbubble contrast agents are costly and have limited applicability. They are not suitable for patients allergic to microbubble contrast agents, neonates, pregnant women, and those with renal insufficiency. Furthermore, due to the random nature of microbubble location and the limitations of microbubble localization techniques, ULM imaging generally requires long data acquisition times (≥10 seconds) to accumulate sufficient microbubble signal to obtain a complete and reliable image of the vascular system. This significantly reduces the temporal resolution of ULM imaging and imposes a significant computational burden. Furthermore, to prevent microbubble rupture, low acoustic pressure is preferred during ULM imaging, which limits the imaging signal-to-noise ratio and penetration depth, reducing the practicality of deep imaging in clinical practice. Furthermore, maintaining a relatively stable microbubble concentration is crucial for achieving optimal image quality but is challenging in practice, especially during the commonly used bolus injection. Therefore, developing a contrast-agent-free super-resolution imaging method is of clinical value.

[0004] However, it is still a huge challenge to break through the diffraction limit of the ultrasound system and achieve micron-level micro-blood flow imaging under contrast-free conditions. In recent years, deep learning-based methods have been expanded and integrated into the field of contrast-free ultrasound micro-blood flow imaging. Related research uses the basic neural network U-Net to extract blood flow signal features from uPD (ultrafast power Doppler) spatiotemporal data to reconstruct super-resolution images, which has become an effective solution for achieving contrast-free super-resolution ultrasound micro-blood flow imaging. However, in ultrasound blood flow imaging, the spatiotemporal data training strategy adopted will greatly increase the training cost, and the receptive field of the basic U-Net network is small, the feature extraction ability is relatively poor, and the performance improvement of the reconstructed image is relatively limited. Summary of the Invention

[0005] The present invention provides a method and system for contrast-free super-resolution ultrasonic micro-blood flow dynamic imaging, aiming to solve the current great challenge of breaking through the diffraction limit of the ultrasound system to achieve microscopic blood flow imaging under contrast-free conditions. The spatiotemporal data training strategy adopted by the existing U-Net-based contrast-free super-resolution blood flow imaging technology will greatly increase the training cost, and the receptive field of the basic U-Net network is small, the feature extraction ability is relatively poor, and the performance improvement of the reconstructed image is relatively limited.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: The present invention provides a method for contrast-free super-resolution ultrasound micro-blood flow dynamic imaging, comprising the following steps: S1. Acquire data pairs including ultrafast power Doppler images and ultrasound positioning microscope images through simulation and experiments, preprocess and enhance the data pairs to form training sets, validation sets, and test sets; S2, based on Mamba, build U-MambaIR, which includes shallow feature extraction module, deep feature extraction module and super-resolution image reconstruction module; Among them, the deep feature extraction module consists of multiple residual-connected omnidirectional state space groups; S3. Define a joint loss function that includes pixel loss, perceptual loss, and adversarial loss, and set the weight parameters of each loss term in the joint loss function; S4, using the training set to input the ultrafast power Doppler image, optimizing the network parameters through the joint loss function to generate the reconstructed super-resolution power Doppler image; S5. Input the ultrafast power Doppler image to be processed into the trained U-MambaIR, extract low-frequency features through the shallow feature extraction module, extract high-frequency features through the deep feature extraction module, and aggregate low-frequency features and high-frequency features through the super-resolution image reconstruction module to output the contrast-free super-resolution ultrasound microblood flow dynamic imaging results.

[0007] In some embodiments, in S1, preprocessing includes: reconstructing the ultrasound positioning microscope image into a preset high-resolution size through a positioning algorithm and a tracking algorithm, and downsampling the corresponding ultrafast power Doppler image into a preset low-resolution size; and normalizing the data pair.

[0008] In some embodiments, in S1, data augmentation includes performing image flipping, image cropping, and random rotation operations on the data pair.

[0009] In some embodiments, in S2, the shallow feature extraction module uses a convolutional layer to extract low-frequency features of the input image and directly passes the low-frequency features to the super-resolution image reconstruction module.

[0010] In some embodiments, in S2, the omnidirectional state space group is composed of multiple residually connected omnidirectional state space blocks and convolutional layers, and the omnidirectional state space block includes an omnidirectional selective scanning module and a channel attention mechanism.

[0011] Furthermore, in S2, the omnidirectional selective scanning module processes the input features in the following manner: The input feature is expanded and divided into two parallel branches; the first branch is processed in sequence by depth convolution, activation function, omnidirectional selective scanning layer and layer normalization; the second branch is processed by activation function; the features output by the first branch and the second branch are merged through aggregation operation and projected back to the original number of channels through linear layer.

[0012] Furthermore, in S2, the omnidirectional selective scanning layer in the omnidirectional state space block performs forward and reverse selective scanning on the input image block along the horizontal, vertical, diagonal and anti-diagonal directions to generate multiple sequences, which are then merged after capturing the long-range dependencies of each sequence using the state space model, thereby fusing the global modeling information from multiple directions.

[0013] In some embodiments, in S2, the super-resolution image reconstruction module aggregates low-frequency features and high-frequency features by element-by-element summation, and uses an upsampling module to perform multiple amplification to generate a high-resolution super-resolution power Doppler image.

[0014] In some implementations, in S3, the joint loss function adopts the following formula (11): (11); in, and is the weight parameter, is the joint loss function, is the perceptual loss function, To counter the loss function, is the pixel loss function.

[0015] The present invention also provides a non-contrast super-resolution ultrasound micro-blood flow dynamic imaging method system, which includes a data integration module, a model construction module, a parameter definition module, a model training module and an ultrasound blood flow imaging module, wherein: Data integration module: used to obtain data pairs including ultrafast power Doppler images and ultrasound positioning microscope images through simulation and experiments, preprocess and enhance the data pairs to form training sets, validation sets, and test sets; Model construction module: used to build a shallow feature extraction module, a deep feature extraction module and a super-resolution image reconstruction module U-MambaIR based on Mamba; Among them, the deep feature extraction module consists of multiple residual-connected omnidirectional state space groups; Parameter definition module: used to define the joint loss function including pixel loss, perceptual loss and adversarial loss, and set the weight parameters of each loss term in the joint loss function; Model training module: used to input ultrafast power Doppler images using the training set, optimize network parameters through the joint loss function, and generate reconstructed super-resolution power Doppler images; Ultrasonic blood flow imaging module: The ultrafast power Doppler image to be processed is input into the trained U-MambaIR. The shallow feature extraction module extracts low-frequency features, the deep feature extraction module extracts high-frequency features, and the super-resolution image reconstruction module aggregates low-frequency and high-frequency features to output contrast-free super-resolution ultrasonic microblood flow dynamic imaging results.

[0016] Compared with the prior art, the present invention provides a method and system for contrast-free super-resolution ultrasound micro-blood flow dynamic imaging, which has the following beneficial effects: The present invention discloses a contrast-free super-resolution ultrasonic micro-blood flow dynamic imaging method. Mamba is introduced into the field of ultrasonic blood flow imaging and an ultrasonic vascular image restoration model U-MambaIR is constructed. By utilizing the advantages of Mamba's global receptive field and linear complexity, blood flow features can be effectively extracted from context-rich and complex uPD images, and microvascular images with complete structure and highly consistent with the ULM true value can be reconstructed. The reconstructed time resolution is the same as that of uPD imaging, thus realizing contrast-free super-resolution micro-blood flow dynamic imaging.

[0017] Furthermore, the present invention uses an omnidirectional selective scanning mechanism, that is, forward and reverse selective scanning along the horizontal, vertical, diagonal and anti-diagonal directions, which is conducive to better extraction of spatial features across large spatial scales and multiple directions in ultrasound vascular images. Mamba's selective scanning mechanism uses kernel fusion and recalculation strategies to significantly reduce a large number of redundant convolution kernel parameters, achieve fast parallel scanning and efficient use of memory, so that U-MambaIR exhibits stronger feature extraction and spatial detail recovery capabilities while maintaining a small model size. The present invention established a complete ultrasound vascular image dataset containing uPD images and ULM images through simulation, animal and clinical experiments to train, verify and test the performance of the model, and also made up for the current lack of datasets in this field. The ultrasound vascular image restoration model U-MambaIR proposed in the present invention has strong generalization performance and can robustly reconstruct CS-PD images in various organs and tissues, showing a wide range of application prospects in complex clinical human imaging environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings in the specification are used to provide further understanding of the present invention and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0019] Figure 1 Schematic diagram of the network architecture of the Mamba-based ultrasound vascular image restoration model U-MambaIR in a contrast-free super-resolution ultrasound micro-blood flow dynamic imaging method of the present invention; Figure 2 Schematic diagram of the architecture of the OSSB module in the U-MambaIR model of a contrast-free super-resolution ultrasound micro-blood flow dynamic imaging method of the present invention; Figure 3 Schematic diagram of the architecture of the OSSM module in the U-MambaIR model of a contrast-free super-resolution ultrasound micro-blood flow dynamic imaging method of the present invention; Figure 4 Schematic diagram of the architecture of the omnidirectional selective scanning mechanism in a contrast-free super-resolution ultrasound micro-blood flow dynamic imaging method of the present invention; Figure 5 Schematic diagram of the comparison of traditional uPD image, CS-PD image reconstructed by U-MambaIR and ULM image in an embodiment of a contrast-free super-resolution ultrasound micro-blood flow dynamic imaging method of the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0022] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0023] In the description of the embodiments of the present invention, it should be noted that if the terms "upper," "lower," "horizontal," "inner," etc. appear, the orientation or positional relationship indicated is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the inventive product is typically placed when in use. These terms are merely for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0025] In the description of the embodiments of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0026] like Figure 1 As shown, the present invention provides a non-contrast super-resolution ultrasound micro-blood flow dynamic imaging method, comprising the following steps: S1. Acquire data pairs including ultrafast power Doppler images and ultrasound positioning microscope images through simulation and experiments, preprocess and enhance the data pairs to form training sets, validation sets, and test sets; S2, based on Mamba, build U-MambaIR, which includes shallow feature extraction module, deep feature extraction module and super-resolution image reconstruction module; Among them, the deep feature extraction module consists of multiple residual-connected omnidirectional state space groups; S3. Define a joint loss function that includes pixel loss, perceptual loss, and adversarial loss, and set the weight parameters of each loss term in the joint loss function; S4, using the training set to input the ultrafast power Doppler image, optimizing the network parameters through the joint loss function to generate the reconstructed super-resolution power Doppler image; S5. Input the ultrafast power Doppler image to be processed into the trained U-MambaIR, extract low-frequency features through the shallow feature extraction module, extract high-frequency features through the deep feature extraction module, and aggregate low-frequency features and high-frequency features through the super-resolution image reconstruction module to output the contrast-free super-resolution ultrasound microblood flow dynamic imaging results.

[0027] In current state-of-the-art technologies, SSMs (State Space Models), particularly the improved Mamba, have advantages in global receptive field and linear complexity, making them an efficient backbone for constructing deep networks for image processing tasks. However, research on the application of Mamba-based models in ultrasound vascular image super-resolution has not yet begun. Currently, there is no method that can efficiently reconstruct super-resolution blood flow images from low-resolution uPD (ultrafast Power Doppler) images. The present invention introduces a Mamba-based neural network to effectively utilize the multidimensional information embedded in the visualized ultrasound data stream for super-resolution image reconstruction, which is conducive to promoting the clinical application of contrast-free super-resolution ultrasound micro-blood flow imaging. The method of the present invention constructs a deep neural network U-MambaIR with Mamba as the backbone. U-MambaIR leverages the advantages of Mamba's global receptive field and linear complexity to effectively extract blood flow features from low-resolution uPD images and reconstruct CS-PD (Contrast-Free Super-Resolution Power Doppler) images. Specifically: The present invention uses Mamba as the backbone to build an ultrasonic vascular image restoration model U-MambaIR.

[0028] An ultrasound vascular image dataset containing uPD images and corresponding ULM (Ultrasound localization microscopy) images obtained by simulation and experiments was established.

[0029] Design a joint loss function to optimize network parameters according to task requirements.

[0030] The established ultrasound vascular image dataset is used to train, test and optimize the ultrasound vascular image restoration model U-MambaIR, so as to ultimately achieve contrast-free super-resolution ultrasound microblood flow dynamic imaging.

[0031] The ultrasonic vascular image restoration model U-MambaIR of the present invention consists of three parts: shallow feature extraction, deep feature extraction and super-resolution image reconstruction. Convolution extracts shallow features that mainly contain low-frequency information and passes them directly to the reconstruction module to preserve the low-frequency information. The deep feature extraction part consists of several stacked residual OSSGs (Omnidirectional State Space Groups) and a The convolutional layer is composed of multiple residual connected OSSB (Omnidirectional State Space Blocks) and a The network consists of 10 convolutional layers, with OSSB being the core module. This module includes the Omnidirectional Selective Scan Module (OSSM) and channel attention. The convolutional layer at the end of each OSSG introduces the inductive bias of the convolution operation into the network, facilitating the aggregation of shallow and deep features later. The super-resolution image reconstruction component aggregates shallow and deep features using an element-by-element summation approach to generate the output image.

[0032] Furthermore, the residual connection OSSB in the deep feature extraction part of the present invention is mainly composed of OSSM and channel attention. The input deep features are normalized using layers and then enter OSSM to capture spatial long-term dependencies and are learned by the scale factor. To control the jump connection. Then, local convolution is introduced to recover local features. In addition, in order to more effectively utilize the information of different channels, channel attention is introduced after local convolution so that SSMs can select key channels. Finally, another adjustable scaling factor is used in the residual connection. To get the output of OSSB.

[0033] The residual connected OSSM in the OSSB of the present invention uses the state space equation to capture long-range dependencies. The input features of OSSM are first processed by a linear layer to convert the feature channels Expand to ( is a predefined channel expansion factor), generating two parallel branches. One branch is refined through depthwise convolution, SiLU activation function, OSS (Omnidirectional Selective Scan) layer, and layer normalization to capture complex patterns; the other branch is processed by SiLU activation function. Afterwards, the features of the two branches are aggregated through Hadamard product. Finally, a linear layer is used to project the number of channels back to .

[0034] The OSS layer in the OSSB of our invention employs an omnidirectional selective scanning mechanism. First, the input image block is scanned forward and backward along the horizontal, vertical, diagonal, and anti-diagonal directions to obtain eight image block sequences. The selective scanning mechanism of the SSM is then used to capture long-range dependencies within each sequence and perform global modeling in specific directions. Finally, all sequences are combined using summation to fuse global modeling information from multiple directions.

[0035] In the specific implementation of the super-resolution image reconstruction part of the present invention, a sub-pixel upsampling module with PixelShuffle is used to upsample the features, and the upsampling ratio is 4.

[0036] This ultrasound vascular image dataset contains uPD images and ULM images obtained through simulation and experimentation. The uPD images are 128×128 pixels in size, and the corresponding ULM images are 512×512 pixels in ground truth. The dataset contains 1500 data pairs, 80% of which are used as training sets, 10% as validation sets, and 10% as test sets. All data pairs are saved as .mat files.

[0037] The joint loss function of the present invention consists of three parts: L1 pixel loss , Perceptual Loss and Fighting Losses The L1 pixel loss function focuses only on pixel-level errors, which can promote the spatial sparsity of the underlying microvascular structure, making the restored image clearer. The perceptual loss function captures higher-level structural and texture information by minimizing the Euclidean distance between the input image and the target image in the feature space extracted by the pre-trained VGG19 network, making the generated image more natural. The adversarial loss function improves realism through adversarial learning, helps to recover high-frequency details, and makes the output image more natural and realistic.

[0038] Furthermore, the uPD images in the ultrasonic vascular image dataset of the present invention are reconstructed using ≥200 frames, and the ULM images are reconstructed using ≥10,000 frames.

[0039] The layer weights of the perceptual loss function are set to {0.1, 0.1, 1, 1, 1}, and the weights of the joint loss function are set to , To effectively balance clarity and visual realism, thereby obtaining a visually optimal reconstructed image.

[0040] A contrast-free super-resolution ultrasound microblood flow dynamic imaging method based on the Mamba neural network was developed using the Pytorch framework.

[0041] The present invention also provides a non-contrast super-resolution ultrasound micro-blood flow dynamic imaging method system, which includes a data integration module, a model construction module, a parameter definition module, a model training module and an ultrasound blood flow imaging module, wherein: Data integration module: used to obtain data pairs including ultrafast power Doppler images and ultrasound positioning microscope images through simulation and experiments, preprocess and enhance the data pairs to form training sets, validation sets, and test sets; Model construction module: used to build U-MambaIR based on Mamba, which includes shallow feature extraction module, deep feature extraction module and super-resolution image reconstruction module; Among them, the deep feature extraction module consists of multiple residual-connected omnidirectional state space groups; Parameter definition module: used to define the joint loss function including pixel loss, perceptual loss and adversarial loss, and set the weight parameters of each loss term in the joint loss function; Model training module: used to input ultrafast power Doppler images using the training set, optimize network parameters through the joint loss function, and generate reconstructed super-resolution power Doppler images; Ultrasonic blood flow imaging module: The ultrafast power Doppler image to be processed is input into the trained U-MambaIR. The shallow feature extraction module extracts low-frequency features, the deep feature extraction module extracts high-frequency features, and the super-resolution image reconstruction module aggregates low-frequency and high-frequency features to output contrast-free super-resolution ultrasonic microblood flow dynamic imaging results.

[0042] The following is a detailed description of a method and system for contrast-free super-resolution ultrasound micro-blood flow dynamic imaging of the present invention through specific embodiments.

[0043] This paper proposes a Mamba-based ultrasound vascular image restoration model, U-MambaIR, which realizes super-resolution ultrasound micro-blood flow dynamic imaging without contrast imaging. This is achieved through the following steps: An ultrasound vascular image dataset was established. The ultrasound vascular image dataset contains uPD data and ULM data obtained by simulation and experiment. In data preprocessing, the original ULM data was reconstructed using the Gaussian fitting positioning and Kuhn-Munkres tracking algorithm in the PALA toolbox, and the reconstructed ULM image was resized to 512×512 and saved as a mat file. The corresponding uPD image was processed using the combined angle singular value filtering method to calculate the power, and downsampled by bilinear interpolation to obtain a low-resolution image of size 128×128, which was saved as a mat file with the same name. The two mat files are paired to form a data pair. All data pairs are normalized from 0-1 minimum to maximum to standardize the data range and improve the training effect and stability of the model. In addition, in order to increase the diversity of training samples and improve the generalization ability of the model, horizontal and vertical flipping, image cropping and random rotation ( and ) and other methods to perform image data enhancement, and finally formed an ultrasound microvascular image dataset with 1500 data pairs, of which 80% were used as training sets, 10% as validation sets, and 10% as test sets.

[0044] like Figure 1 As shown, the U-MambaIR network is built. The U-MambaIR network of the present invention consists of three parts: shallow feature extraction, deep feature extraction and super-resolution image reconstruction. For an input low-resolution uPD image ,in 、 and 1 are the length, width and number of channels of the image respectively. First, use a The convolutional layer To extract shallow features that mainly contain low-frequency information , and passed to the deep feature extraction part and the super-resolution image reconstruction part. The deep feature extraction part consists of several residual connected OSSGs and a The convolutional layer consists of Deep layer features The extraction process is shown below: (1); in, For the OSSG, For the Deep layer features.

[0045] like Figure 2 As shown, OSSG consists of multiple residual connected OSSBs and a The convolutional layer of OSSB is the core module. In OSSB, the deep features of the input After using layer normalization, it enters the OSSM module to capture spatial long-term dependencies. In addition, a learnable scaling factor is used to control the jump connection, and the output of OSSM is finally obtained. . This process can be expressed as: (2); in, is layer normalization, is a learnable scaling factor.

[0046] After OSSM, local convolution is introduced to recover local features, and channel attention is introduced so that SSMs can focus on learning different channel representations, and then select key channels through subsequent channel attention. Finally, another adjustable scaling factor is used in the residual connection to obtain the output of OSSB . This process is expressed as: (3); in, is the channel attention, is local convolution, is a learnable scaling factor.

[0047] like Figure 3 As shown, the OSSM module uses the state space equation to capture long-range dependencies. The input features of OSSM are First, a linear layer is used to expand the feature channel to ( is a predefined channel expansion factor), generating two parallel branches. One branch is refined through depthwise convolution, SiLU activation function, OSS layer, and layer normalization to capture complex patterns; the other branch is processed by SiLU activation function. Afterwards, the features of the two branches are aggregated through Hadamard product. Finally, a linear layer is used to project the number of channels back to The above process can be expressed as: (4); (5); (6); in, is the depthwise convolution, is a linear layer, is the input feature of OSSM, is the Hadamard product.

[0048] like Figure 4As shown in FIG, the OSS layer of the present invention adopts an omnidirectional selective scanning mechanism. Specifically, the input image block is first scanned along eight directions to obtain eight image block sequences. Then, the selective scanning mechanism of SSM is used to capture the long-range dependencies of each sequence and perform global modeling in a specific direction. Finally, all sequences are merged using summation to fuse global modeling information from multiple directions. In this way, large spatial features can be extracted from all directions in the ultrasound vascular image. Finally, in the super-resolution image reconstruction part, Reconstruct the CS-PD image by aggregating shallow and deep features using element-wise summation , as shown below: (7); in, It is a shallow feature. For the Deep features, Element-wise sum.

[0049] Design a joint loss function to optimize network parameters. The joint loss function consists of three parts: L1 pixel loss function, perceptual loss function and adversarial loss function. L1 pixel loss function The formula is as follows: (8); in, For The corresponding ULM image is used as the true value for model training.

[0050] Perceptual loss function The formula is as follows: (9); in, For VGG19 network The feature map output by the layer, , as well as They are VGG19 network The dimensions of the layer output.

[0051] Adversarial Loss Function The formula is as follows: (10); in, Expectations for CS-PD images, As the discriminator, a three-scale patch discriminator is used here.

[0052] The formula of the joint loss function is as follows: (11); in, and is the weight parameter.

[0053] Training parameter settings. The initial learning rate of the model generator is 0.001, and the initial learning rate of the discriminator is 0.0001. The training of the generator and the discriminator uses a multi-step decay strategy. After a certain number of iterations, the learning rate is reduced to 1 / 2 of the original. The model uses the Adam iterative optimizer to update the model parameters. The parameters of the Adam iterative optimizer are , Set them to 0.9 and 0.999 respectively. The batch_size of training is set to 1.

[0054] Training begins. The training set data is fed into the U-MambaIR network. After several iterations, the network model saves the parameter information from each stage and generates the final optimal weight model, which is used to test the model's reconstruction and generalization performance.

[0055] The present invention proposes a Mamba-based ultrasound vascular image restoration model, U-MambaIR. This network utilizes the advantages of Mamba's global receptive field and linear complexity to effectively extract blood flow features from uPD images with rich context and complex content, and reconstruct microvascular images with complete structure and highly consistent with the ULM true value. The reconstructed temporal resolution is the same as that of uPD imaging, and contrast-free super-resolution dynamic imaging can be achieved. In addition, Mamba's selective scanning mechanism significantly reduces a large number of redundant convolution kernel parameters by utilizing kernel fusion and recalculation strategies, achieving fast parallel scanning and efficient use of memory, allowing U-MambaIR to exhibit stronger feature extraction and spatial detail recovery capabilities while maintaining a small model size. U-MambaIR also has excellent generalization capabilities, and can robustly reconstruct complete vascular network structures with more details in various organs and tissues, and effectively improve the spatial resolution of imaging, showing broad application prospects in complex clinical human imaging environments.

[0056] like Figure 5 As shown, for comparison, Figure 5The following table compares uPD images, CS-PD images reconstructed using U-MambaIR, and their corresponding ULM ground truth images using test set data. Compared with traditional uPD images, the U-MambaIR of our invention robustly reconstructs the complete microvascular structure using Mamba's powerful global receptive field, significantly improving spatial resolution. The generated CS-PD images are highly consistent with the ULM images. To quantitatively evaluate the performance of U-MambaIR, we calculated the SSIM (structure similarity index measure), PSNR (peak signal-to-noise ratio), and FWHM (full width at half maxima) ratio (CS-PD / ULM). SSIM and PSNR are calculated based on all data in the test set, while the FWHM ratio is calculated from the cross-sectional profiles of 12 manually selected local blood vessels in the CS-PD and ULM images of the test set. The SSIM and PSNR of the CS-PD images reconstructed using U-MambaIR were 0.805±0.030 and 23.72±0.95, respectively, and the FWHM ratio (CS-PD / ULM) was 1.08±0.21. These results demonstrate that the CS-PD images reconstructed using U-MambaIR have a high degree of consistency in vascular structure with the true ULM image and achieve similar spatial resolution to that of ULM.

[0057] Finally, it should be noted that the above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any ordinary technician in this industry can smoothly implement the present invention as shown in the specification and described above. Any equivalent changes, modifications and evolutions made by using the technical content disclosed above are all equivalent embodiments of the present invention. At the same time, any equivalent changes, modifications and evolutions made to the above embodiments based on the essential technology of the present invention are still within the scope of protection of the technical solution of the present invention.

Claims

1. A method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging, characterized in that: The steps include: S1. Acquire data pairs including ultrafast power Doppler images and ultrasound positioning microscope images through simulation and experiments, preprocess and enhance the data pairs to form training sets, validation sets, and test sets; S2, based on Mamba, build U-MambaIR, which includes shallow feature extraction module, deep feature extraction module and super-resolution image reconstruction module; Among them, the deep feature extraction module consists of multiple residual-connected omnidirectional state space groups; S3. Define a joint loss function that includes pixel loss, perceptual loss, and adversarial loss, and set the weight parameters of each loss term in the joint loss function; S4, using the training set to input the ultrafast power Doppler image, optimizing the network parameters through the joint loss function to generate the reconstructed super-resolution power Doppler image; S5. Input the ultrafast power Doppler image to be processed into the trained U-MambaIR, extract low-frequency features through the shallow feature extraction module, extract high-frequency features through the deep feature extraction module, and aggregate low-frequency features and high-frequency features through the super-resolution image reconstruction module to output the contrast-free super-resolution ultrasound microblood flow dynamic imaging results.

2. The method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging according to claim 1, characterized in that: In said S1, the preprocessing includes: reconstructing the ultrasound positioning microscope image into a preset high-resolution size through a positioning algorithm and a tracking algorithm, and processing the corresponding ultrafast power Doppler image into a preset low-resolution size through downsampling; and normalizing the data pairs.

3. The method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging according to claim 1, characterized in that: In S1, data enhancement includes performing image flipping, image cropping, and random rotation operations on the data pairs.

4. The method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging according to claim 1, characterized in that: In S2, the shallow feature extraction module uses a convolutional layer to extract low-frequency features of the input image, and directly transfers the low-frequency features to the super-resolution image reconstruction module.

5. The method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging according to claim 1, characterized in that: In the S2, the omnidirectional state space group consists of multiple residual-connected omnidirectional state space blocks and convolutional layers, and the omnidirectional state space block includes an omnidirectional selective scanning module and a channel attention mechanism.

6. The method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging according to claim 5, characterized in that: In S2, the omnidirectional selective scanning module processes the input features in the following manner: The input feature is expanded and divided into two parallel branches; the first branch is processed in sequence by depth convolution, activation function, omnidirectional selective scanning layer and layer normalization; the second branch is processed by activation function; the features output by the first branch and the second branch are merged through aggregation operation and projected back to the original number of channels through linear layer.

7. The method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging according to claim 6, characterized in that: In the S2, the omnidirectional selective scanning layer in the omnidirectional state space block scans the input image block forward and backward along the horizontal, vertical, diagonal and anti-diagonal directions to generate multiple sequences, and uses the state space model to capture the long-range dependencies of each sequence and then merge them to fuse the global modeling information from multiple directions.

8. The method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging according to claim 1, characterized in that: In S2, the super-resolution image reconstruction module aggregates low-frequency features and high-frequency features by element-by-element summation, and uses an upsampling module to perform multiple amplification to generate a high-resolution super-resolution power Doppler image.

9. The method for non-contrast super-resolution ultrasound micro-blood flow dynamic imaging according to claim 1, characterized in that: In S3, the joint loss function adopts the following formula (11): （11）； in, and is the weight parameter, is the joint loss function, is the perceptual loss function, To counter the loss function, is the pixel loss function.

10. The system based on the non-contrast super-resolution ultrasound micro-blood flow dynamic imaging method according to any one of claims 1 to 9, characterized in that: The system includes a data integration module, a model construction module, a parameter definition module, a model training module and an ultrasonic blood flow imaging module, wherein: Data integration module: used to obtain data pairs including ultrafast power Doppler images and ultrasound positioning microscope images through simulation and experiments, preprocess and enhance the data pairs to form training sets, validation sets, and test sets; Model construction module: used to build U-MambaIR based on Mamba, which includes shallow feature extraction module, deep feature extraction module and super-resolution image reconstruction module; Among them, the deep feature extraction module consists of multiple residual-connected omnidirectional state space groups; Parameter definition module: used to define the joint loss function including pixel loss, perceptual loss and adversarial loss, and set the weight parameters of each loss term in the joint loss function; Model training module: used to input ultrafast power Doppler images using the training set, optimize network parameters through the joint loss function, and generate reconstructed super-resolution power Doppler images; Ultrasonic blood flow imaging module: The ultrafast power Doppler image to be processed is input into the trained U-MambaIR. The shallow feature extraction module extracts low-frequency features, the deep feature extraction module extracts high-frequency features, and the super-resolution image reconstruction module aggregates low-frequency and high-frequency features to output contrast-free super-resolution ultrasonic microblood flow dynamic imaging results.

Images