Method and system for assessing vascular fibrosis burden based on intravascular ultrasound data

Abstract

This invention discloses a method and system for assessing vascular fibrosis burden based on intravascular ultrasound data, belonging to the field of image processing technology. The method involves acquiring ultrasound images, denoising the ultrasound images to obtain denoised ultrasound images, identifying the denoised ultrasound images using a target recognition model, calculating the fibrosis burden score based on the identification result, and providing early warnings based on the fibrosis burden score. By denoising the ultrasound images to improve image quality, and using a target recognition model to accurately analyze the denoised intravascular ultrasound images, the method enhances the model's ability to identify fibrotic plaques and insignificant features, thereby achieving automated, quantitative, and accurate calculation of the vascular fibrosis burden score and providing timely early warnings based on the quantified score.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method and system for assessing vascular fibrosis burden based on intravascular ultrasound data. Background Technology

[0002] Intravascular ultrasound, as an important imaging tool for assessing vascular wall structure and plaque characteristics, has been widely used in clinical practice. However, traditional methods relying on physician visual analysis have limitations such as high subjectivity, low repeatability, and insensitivity to early or diffuse fibrotic lesions.

[0003] Publication No. CN121169788A discloses an ultrasound image processing method, apparatus, electronic device, and storage medium; acquiring an ultrasound image containing at least one vascular region; segmenting the ultrasound image for the vascular region in the ultrasound image to obtain a first segmented image, the first segmented image including the vascular region; determining the center pixel of each vascular region in the first segmented image; correcting the vascular region in the first segmented image based on the center pixel and the ultrasound image to obtain a second segmented image; and synthesizing the second segmented image and the ultrasound image to obtain a tissue attenuation image with the vascular region filtered out.

[0004] In existing technologies, the low signal-to-noise ratio and blurred tissue boundaries caused by noise interference in the collected blood vessels lead to reduced accuracy and efficiency in subsequent identification; traditional assessment of vascular fibrosis burden relies on manual labor, which is highly subjective and prone to missed diagnoses. Summary of the Invention

[0005] The purpose of this invention is to address the following issues mentioned in the background: 1. The low signal-to-noise ratio and blurred tissue boundaries caused by noise interference in the acquired blood vessels, leading to reduced accuracy and efficiency in subsequent identification; 2. The limitations of traditional vascular fibrosis burden assessment, which relies on manual intervention, is highly subjective, and prone to missed diagnoses. Therefore, this invention proposes a method and system for assessing vascular fibrosis burden based on intravascular ultrasound data.

[0006] A first aspect of this invention provides a method for assessing vascular fibrosis burden based on intravascular ultrasound data, the method comprising: Acquire ultrasound images and a stitched image set, and calculate the similarity between the ultrasound images and each stitched image in the stitched image set; the stitched images in the stitched image set are obtained by adding noise to standard ultrasound images; Determine the target time label corresponding to the stitched image with the highest similarity, and convert the target time label to obtain the target time label feature code; The ultrasonic image and the target time-label feature encoding are input into the U-net network to obtain the predicted noise. The ultrasonic image is then de-denoised using the predicted noise to obtain the denoised ultrasonic image. The denoised ultrasonic image is identified using a target recognition model to obtain the recognition result; The fiber load score is calculated based on the identification results, and an early warning is issued based on the fiber load score.

[0007] By denoising the original ultrasound images to improve image quality, and then accurately detecting key features in the images based on an improved target recognition model, the quantitative fiber load score is automatically calculated based on the recognition results. This ultimately achieves an objective and rapid assessment and intelligent early warning of the degree of fibrosis, effectively improving the automation level and accuracy of diagnosis.

[0008] Optionally, before acquiring the ultrasound image, the method further includes: Step 1: Obtain a standard ultrasound image; add multiplicative speckle to the standard ultrasound image using a first formula to obtain a speckled standard ultrasound image; add Gaussian noise to the standard ultrasound image using a second formula to obtain a Gaussian standard ultrasound image; and fuse the speckled standard ultrasound image and the Gaussian standard ultrasound image to obtain a fused standard ultrasound image. Step 2: Obtain the total number of diffusion steps and time stamps. Based on the total number of diffusion steps and time stamps, gradually add perturbations to the fused standard ultrasound image to obtain the final perturbation image and time stamp image set; the time stamp image set contains multiple perturbation images. Step 3: Extract the perturbation images from the time-stamped image set to obtain the target time-stamped image set. Then, stitch the fused standard ultrasonic image with each perturbation image in the target time-stamped image set according to the channel to obtain the stitched image set. Step 4: Obtain the time label corresponding to each stitched image in the stitched image set, convert the time label to obtain time label feature code, input the stitched image and the corresponding time label feature code into the U-net network to obtain prediction noise, and average the prediction noise to obtain the average prediction noise. Step 5: Perform reverse denoising on the fused standard ultrasound image using the average predicted noise to obtain a denoised standard ultrasound image. Calculate the difference between the standard ultrasound image and the denoised standard ultrasound image. If the difference is less than a preset threshold, the denoising is deemed successful; otherwise, the denoising is deemed unsuccessful, and the model parameters of the U-net network are updated. Repeat steps 1 to 5 until the denoising is deemed successful, then determine the current model parameters.

[0009] By constructing a robust denoising framework based on a diffusion model, a fused standard ultrasound image containing simulated speckle and Gaussian mixed noise is generated for training. The noise distribution is learned using multi-time-label perturbation and a U-net network. The effectiveness and generalization ability of the denoising model are ensured through iterative training and reverse denoising verification. The trained model is then applied to real ultrasound images to improve image quality, providing a clear and reliable input foundation for subsequent target recognition and accurate calculation of fiber load fraction.

[0010] Optionally, the denoised ultrasonic image is identified using a target recognition model to obtain a recognition result. The target recognition model, based on improvements to the YOLOv8 model, includes: The improved backbone network is obtained by replacing each C2f module in the backbone network with an improved C2f module, and the target recognition model is obtained by replacing the neck structure with an improved neck structure; the YOLOv8 model includes a backbone network and a neck structure. The working principle of the improved neck structure includes: The output of the improved C2f module at layer 4 in the improved backbone network is used as the first input diagram of the improved neck structure, the output of the improved C2f module at layer 6 in the improved backbone network is used as the second input diagram of the improved neck structure, and the output of the SPPF module at layer 9 in the improved backbone network is used as the third input diagram of the improved neck structure. The first input image, the second input image, and the third input image are input into the MFC module to obtain the first feature image. The first feature image is convolved to obtain the second feature image. The first feature image is upsampled to obtain the third feature image. The second feature image is concatenated with the third input image to obtain the fourth feature image. The third feature image is concatenated with the first input image to obtain the fifth feature image. The fourth and fifth feature maps are input into the improved C2f module to obtain the sixth and seventh feature maps, respectively. The first, sixth, and seventh feature maps are input into the MFC module to obtain the eighth feature map. The eighth feature map is convolved and then concatenated with the second and sixth feature maps to obtain the ninth feature map. The eighth feature map is upsampled and then concatenated with the third and seventh feature maps to obtain the tenth feature map. The 8th feature map, the 9th feature map, and the 10th feature map are used as the output of the improved neck structure.

[0011] By constructing a dense bidirectional multi-scale fusion loop, the MFC module and the improved C2f module are deeply nested. Through two cross-layer MFC fusions and multiple feature concatenation, upsampling, and downsampling operations, a high-strength and cyclically interactive information pathway is established between the shallow, medium, and deep feature layers. This achieves the alignment of features at different scales and the deep aggregation of contextual information, allowing semantic information and detailed features to fully complement and enhance each level of the pyramid. The output feature map is highly complementary and incorporates rich global and local contextual information, improving the model's detection accuracy and robustness for multi-scale targets.

[0012] Optionally, the improved C2f module's workflow includes: Obtain the original tensor, input the original tensor into the Conv module to obtain the first tensor, split the first tensor to obtain the first branch tensor and the second branch tensor, input the first branch tensor into the bottleneck module to obtain the third branch tensor, concatenate the second branch tensor and the third branch tensor to obtain the second tensor, input the second tensor into the Conv module to obtain the third tensor, and use the third tensor as the output of the improved C2f module; The workflow of the MFC module includes: Obtain the first feature tensor, the second feature tensor, and the third feature tensor. Downsample the first feature tensor to obtain the fourth feature tensor. Convolve the second feature tensor to obtain the fifth feature tensor. Upsample the third feature tensor and then convolve it to obtain the sixth feature tensor. The 4th, 5th, and 6th feature tensors are concatenated to obtain the 7th feature tensor. The 7th feature tensor is then input into the target DWConv module to obtain the 8th, 9th, 10th, and 11th feature tensors, respectively. The kernel size of the target DWConv module can range from 5, 7, 9, to 11. The 7th, 8th, 9th, 10th, and 11th feature tensors are concatenated and then convolved to obtain the 12th feature tensor, which is then used as the output of the MFC module.

[0013] The improved C2f module, through branching and bottleneck structure processing, achieves efficient fusion and feature reuse of shallow details and deep semantics within a single feature layer, enhancing the model's representational ability and gradient flow. Meanwhile, the MFC module constructs a powerful multi-scale contextual information capture and fusion mechanism through cross-scale downsampling, upsampling alignment, and multi-path parallel large-kernel depthwise separable convolutions. Working together, these two modules improve the model's feature extraction and integration efficiency, enabling the network to more fully represent and associate target information at different scales, thus enhancing the detection performance of small and detailed targets.

[0014] Optionally, the fiber load fraction is calculated based on the identification result, and the method includes: The area of ​​fibrous tissue and the longitudinal section area of ​​blood vessels are determined based on the identification results, and the fibrous load fraction is calculated based on the area of ​​fibrous tissue and the longitudinal section area of ​​blood vessels. Formula for calculating fibrosis load: *100%; Where F is the fiber load fraction, F0 is the area of ​​fibrous tissue, and F1 is the area of ​​the longitudinal section of the blood vessel.

[0015] A second aspect of the present invention provides a vascular fibrosis burden assessment system based on intravascular ultrasound data, the system comprising: An image acquisition module is used to acquire ultrasonic images and a stitched image set, and to calculate the similarity between the ultrasonic images and each stitched image in the stitched image set; the stitched images in the stitched image set are obtained by adding noise to standard ultrasonic images; The label conversion module is used to determine the target time label corresponding to the stitched image with the highest similarity, and convert the target time label to obtain the target time label feature code; The inverse denoising module is used to input the ultrasonic image and the target time label feature encoding into the U-net network to obtain predicted noise, and to perform inverse denoising on the ultrasonic image using the predicted noise to obtain a denoised ultrasonic image. The image recognition module is used to identify the denoised ultrasonic image using a target recognition model to obtain the recognition result; The image evaluation module is used to calculate the fiber load score based on the recognition result and to issue an early warning based on the fiber load score.

[0016] Optionally, the system further includes: The first operation module is used to acquire a standard ultrasound image, add multiplicative speckle to the standard ultrasound image using a first formula to obtain a speckle standard ultrasound image, add Gaussian noise to the standard ultrasound image using a second formula to obtain a Gaussian standard ultrasound image, and fuse the speckle standard ultrasound image and the Gaussian standard ultrasound image to obtain a fused standard ultrasound image. The second operation module is used to obtain the total number of diffusion steps and time stamps, and to gradually add perturbations to the fused standard ultrasound image based on the total number of diffusion steps and time stamps to obtain the final perturbation image and time stamp image set; the time stamp image set contains multiple perturbation images; The third operation module is used to extract the perturbation images in the time-stamped image set to obtain the target time-stamped image set, and then combine the fused standard ultrasonic image with each perturbation image in the target time-stamped image set according to the channel to obtain the spliced ​​image set. The fourth operation module is used to obtain the time label corresponding to each stitched image in the stitched image set, convert the time label to obtain the time label feature code, input the stitched image and the corresponding time label feature code into the U-net network to obtain the prediction noise, and average the prediction noise to obtain the average prediction noise. The fifth operation module is used to reverse denoise the fused standard ultrasound image by averaging the predicted noise to obtain a denoised standard ultrasound image, calculate the difference between the standard ultrasound image and the denoised standard ultrasound image, and determine that the denoising is qualified if the difference is less than a preset threshold; otherwise, it determines that the denoising has failed and updates the model parameters of the U-net network, and re-executes the fourth to fifth operation modules until the denoising is determined to be qualified, then determines the current model parameters.

[0017] Optionally, the denoised ultrasonic image is identified using a target recognition model to obtain a recognition result. The target recognition model, based on improvements to the YOLOv8 model, includes: The improved backbone network is obtained by replacing each C2f module in the backbone network with an improved C2f module, and the target recognition model is obtained by replacing the neck structure with an improved neck structure; the YOLOv8 model includes a backbone network and a neck structure. The working principle of the improved neck structure includes: The output of the improved C2f module at layer 4 in the improved backbone network is used as the first input diagram of the improved neck structure, the output of the improved C2f module at layer 6 in the improved backbone network is used as the second input diagram of the improved neck structure, and the output of the SPPF module at layer 9 in the improved backbone network is used as the third input diagram of the improved neck structure. The first input image, the second input image, and the third input image are input into the MFC module to obtain the first feature image. The first feature image is convolved to obtain the second feature image. The first feature image is upsampled to obtain the third feature image. The second feature image is concatenated with the third input image to obtain the fourth feature image. The third feature image is concatenated with the first input image to obtain the fifth feature image. The fourth and fifth feature maps are input into the improved C2f module to obtain the sixth and seventh feature maps, respectively. The first, sixth, and seventh feature maps are input into the MFC module to obtain the eighth feature map. The eighth feature map is convolved and then concatenated with the second and sixth feature maps to obtain the ninth feature map. The eighth feature map is upsampled and then concatenated with the third and seventh feature maps to obtain the tenth feature map. The 8th feature map, the 9th feature map, and the 10th feature map are used as the output of the improved neck structure.

[0018] Optionally, the improved C2f module's workflow includes: Obtain the original tensor, input the original tensor into the Conv module to obtain the first tensor, split the first tensor to obtain the first branch tensor and the second branch tensor, input the first branch tensor into the bottleneck module to obtain the third branch tensor, concatenate the second branch tensor and the third branch tensor to obtain the second tensor, input the second tensor into the Conv module to obtain the third tensor, and use the third tensor as the output of the improved C2f module; The workflow of the MFC module includes: Obtain the first feature tensor, the second feature tensor, and the third feature tensor. Downsample the first feature tensor to obtain the fourth feature tensor. Convolve the second feature tensor to obtain the fifth feature tensor. Upsample the third feature tensor and then convolve it to obtain the sixth feature tensor. The 4th, 5th, and 6th feature tensors are concatenated to obtain the 7th feature tensor. The 7th feature tensor is then input into the target DWConv module to obtain the 8th, 9th, 10th, and 11th feature tensors, respectively. The kernel size of the target DWConv module can range from 5, 7, 9, to 11. The 7th, 8th, 9th, 10th, and 11th feature tensors are concatenated and then convolved to obtain the 12th feature tensor, which is then used as the output of the MFC module.

[0019] Optionally, the image evaluation module is further configured to determine the fibrous tissue area and the longitudinal section area of ​​the blood vessel based on the recognition result, and calculate the fibrous load score based on the fibrous tissue area and the longitudinal section area of ​​the blood vessel. Formula for calculating fibrosis load: *100%; Where F is the fiber load fraction, F0 is the area of ​​fibrous tissue, and F1 is the area of ​​the longitudinal section of the blood vessel.

[0020] The beneficial effects of this invention are: This invention proposes a method for assessing vascular fibrosis burden based on intravascular ultrasound data. Intravascular ultrasound images are processed using image denoising technology to suppress noise interference and improve image clarity, providing a high-quality data foundation for subsequent quantitative analysis. A target model is used to accurately identify and analyze the denoised images. This model enhances the detection sensitivity and localization accuracy of fibrotic plaques, especially those with small, edge-shaped, and low-contrast features, by strengthening cross-scale feature interaction and information fusion capabilities. Attached Figure Description

[0021] Figure 1 A flowchart of a method for assessing vascular fibrosis burden based on intravascular ultrasound data provided in an embodiment of the present invention; Figure 2 A model framework diagram of a vascular fibrosis burden assessment method based on intravascular ultrasound data provided in an embodiment of the present invention; Figure 3 A schematic diagram of a vascular fibrosis burden assessment method based on intravascular ultrasound data provided in an embodiment of the present invention. Detailed Implementation

[0022] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0023] This invention provides a method for assessing vascular fibrosis burden based on intravascular ultrasound data. See also... Figure 1 , Figure 1 A flowchart illustrating a method for assessing vascular fibrosis burden based on intravascular ultrasound data, provided in an embodiment of the present invention. The method includes the following steps: S101, acquire the ultrasonic image and the stitched image set, and calculate the similarity between the ultrasonic image and each stitched image in the stitched image set; S102, determine the target time label corresponding to the stitched image with the highest similarity, and convert the target time label to obtain the target time label feature code; S103, input the ultrasonic image and target time label feature encoding into the U-net network to obtain the predicted noise, and use the predicted noise to perform reverse denoising on the ultrasonic image to obtain the denoised ultrasonic image; S104, the target recognition model is used to identify the denoised ultrasonic image to obtain the recognition result; S105 calculates the fiber load score based on the identification results and issues an early warning based on the fiber load score.

[0024] The stitched images in the image set are obtained by adding noise to standard ultrasound images; Based on the vascular fibrosis load assessment method based on intravascular ultrasound data provided in this invention, the image quality is improved by denoising the ultrasound image, and the target recognition model performs accurate analysis on the denoised intravascular ultrasound image; the model's ability to identify fibrotic plaques and insignificant features is enhanced, thereby realizing the automated, quantitative and accurate calculation of vascular fibrosis load score, and providing timely early warning based on the quantified score.

[0025] In one implementation, the similarity between the ultrasonic image and each stitched image in the image set is calculated: that is, the similarity of noise between the ultrasonic image and each stitched image is calculated. The similarity of noise can be quantified by means of noise variance, spectral features, etc. The most similar stitched image is found based on the similarity, and its corresponding time label is applied to the ultrasonic image, thereby reducing the parameter retrieval process in the denoising process.

[0026] In one implementation, intravascular ultrasound data includes: Conventional images (grayscale images): processed cross-sectional images of the blood vessel that are directly visible to the naked eye. These are used to display the three layers of the vessel wall (intima, media, and adventitia), lumen size, plaque location, and overall morphology. Radiofrequency signal / raw echo data: the raw electrical signal data after the probe emits ultrasound waves and receives the echoes. It contains detailed information such as the amplitude and frequency spectrum of the ultrasound waves reflected from the tissue. This is used to assess tissue components such as fibrosis, lipids, and calcification.

[0027] In one implementation, during the fibrosis load assessment of blood vessels, the system acquires hundreds of continuous cross-sectional images. The computer reconstructs these two-dimensional images into a three-dimensional model of the blood vessel, and a specific blood vessel segment can be delineated for load assessment as needed.

[0028] In one implementation, the fiber load score is calculated based on the identification results, and the fiber structure echo ratio is divided into mild (<30%), moderate (30%-60%), and severe (>60%).

[0029] In one embodiment, before acquiring the ultrasound image, the method further includes: Step 1: Obtain a standard ultrasound image. Add multiplicative speckle to the standard ultrasound image using the first formula to obtain a speckled standard ultrasound image. Add Gaussian noise to the standard ultrasound image using the second formula to obtain a Gaussian standard ultrasound image. Fuse the speckled standard ultrasound image and the Gaussian standard ultrasound image to obtain a fused standard ultrasound image. Step 2: Obtain the total number of diffusion steps and time tags. Based on the total number of diffusion steps and time tags, gradually add perturbations to the fused standard ultrasound image to obtain the final perturbation image and time tag image set; the time tag image set contains multiple perturbation images. Step 3: Extract the perturbation images from the time-stamped image set to obtain the target time-stamped image set. Then, stitch the fused standard ultrasound image with each perturbation image in the target time-stamped image set according to the channel to obtain the stitched image set. Step 4: Obtain the time label corresponding to each stitched image in the stitched image set, convert the time label to obtain the time label feature code, input the stitched image and the corresponding time label feature code into the U-net network to obtain the prediction noise, and average the prediction noise to obtain the average prediction noise. Step 5: Perform reverse denoising on the fused standard ultrasound image using the average predicted noise to obtain a denoised standard ultrasound image. Calculate the difference between the standard ultrasound image and the denoised standard ultrasound image. If the difference is less than a preset threshold, the denoising is deemed successful; otherwise, the denoising is deemed unsuccessful, and the model parameters of the U-net network are updated. Repeat steps 1 to 5 until the denoising is deemed successful, then determine the current model parameters.

[0030] In one implementation, standard ultrasound images are collected from raw, clean vascular ultrasound images in clinical settings that are free of noise and artifacts. These uninterrupted, realistic images serve as a reference for all subsequent processing, ensuring that noise modeling and denoising model training have clear objectives.

[0031] A speckled standard ultrasound image is obtained by adding multiplicative speckles to a standard ultrasound image using the first formula: speckle noise is generated by coherent interference between ultrasound waves and tissue microstructures (cells, collagen fibers), and appears as granular particles, multiplying with the image signal; First formula: , among which, I mul I0 is the standard ultrasound image, and n1 is the speckle noise factor, which follows a Rayleigh, K or Gamma distribution to simulate the noise intensity under different clinical scenarios. Gaussian standard ultrasound images are obtained by adding Gaussian noise to standard ultrasound images using the second formula. This type of noise originates from the electronic amplifiers, A / D converters, and other circuits of the ultrasound equipment. It is independent of the superposition interference of the image signal and follows a Gaussian distribution. The second formula is as follows: , among which, Iadd The image is a standard Gaussian ultrasound image, and n2 is a Gaussian noise factor that follows a Gaussian distribution with a mean of 0 and a variance of a specific value (the magnitude of the variance corresponds to the noise intensity, and the variance is larger in low signal-to-noise ratio scenarios).

[0032] In one implementation, perturbations are progressively added to the fused standard ultrasound image based on the total number of diffusion steps and time labels to obtain the final perturbation image and time-labeled image set: the total number of diffusion steps T is determined, for example, 1000 steps, and the noise variance β for each time label t is set using a cosine noise scheduling strategy. t This replaces the linear scheduling of the original denoising diffusion probability model, avoiding the problem of excessive noise in the early stage and insufficient noise in the later stage; each time tag corresponds to a fixed set of noise parameters to control the noise addition intensity in that stage. Through the formula: ,in, Given the previous time-labeled image y t−1 Current time-stamped image y t The conditional probability distribution of y; t For the perturbation image at time label t, y t−1 Here is the perturbation image at time label t−1, where N is a Gaussian distribution and β t Let be the noise variance of the time label t. The mean, β of the current time-stamped image t *I represents the variance of the current time-stamped image, and I is the identity matrix; after T=1000 steps, the fused standard ultrasound image is gradually perturbed into a final perturbed image that is close to pure Gaussian noise, at which point the original structural information of the image is masked by noise.

[0033] In one implementation, the standard ultrasonic image and the perturbation images in the target time-stamped image set are stitched together according to the channel to obtain the stitched image set: the standard ultrasonic image and the perturbation image of any time-stamp t during the forward diffusion process are stitched together along the channel dimension: the image is usually H×W×C, where C is the number of channels, to obtain the stitched image.

[0034] In one implementation, the time stamps are transformed into time stamp feature codes: for each diffusion time stamp t (range 1~T), a Transformer-based sinusoidal positional encoding is used to convert it into a time stamp embedding t. c (A vector with dimensions consistent with the network feature dimensions), and t c Each residual block is passed into the subsequent network. This allows the improved U-net model to clearly identify which stage of inverse denoising it is currently in, thus adaptively adjusting the denoising intensity—higher in the early stages and lower in the later stages.

[0035] In one implementation, the model parameters in the U-net network mainly include: the weights and biases of all convolutional layers in the encoder-decoder path, the transformation layer parameters corresponding to the skip connections used for multi-scale feature fusion, the weights of the embedding layer and attention layer that fuse time-labeled feature encoding with image features, and learnable scaling and offset parameters in possible normalization layers. These parameters work together to learn how to accurately predict noise from noisy images based on the input time-labeled encoding, and are continuously updated through training to approximate the true noise distribution, ultimately achieving effective inverse denoising.

[0036] In one implementation, the reverse denoising process is the inverse process of restoring a clean image from pure Gaussian noise. The time label gradually decreases from T=1000 to 1. At each time label t, the model predicts the noise component corresponding to that step based on the current noisy image and the original noisy image information, and then removes this part of the noise to obtain a clearer image. For example, pure noise is processed at t=1000, and the final denoising is completed at t=1, outputting the final clean image.

[0037] In one implementation, the U-net network includes an input module, an encoder module, a decoder module, and an output module; the input module, encoder module, decoder module, and output module are connected sequentially, and the encoder module and decoder module are connected in skip connections.

[0038] In one implementation, the encoder module further includes: a convolutional residual block (containing two normalized layers + a SiLU activation layer), an SE (Squeeze-and-Excitation) attention module, and a downsampling convolutional submodule (4×4 convolutional kernel, stride 2).

[0039] In one implementation, the decoder module further includes: an upsampled transposed convolution submodule (4×4ConvTrans2D, stride 2), an SE attention module, a convolutional residual block (same as the encoder), and a skip connection submodule.

[0040] In one embodiment, a target recognition model is used to identify the denoised ultrasonic image to obtain the recognition result. The target recognition model is based on improvements to the YOLOv8 model, including: The improved backbone network is obtained by replacing each C2f module in the backbone network with an improved C2f module, and the target recognition model is obtained by replacing the neck structure with an improved neck structure. The YOLOv8 model includes a backbone network and a neck structure. The working principle of improving neck structure includes: The output of the improved C2f module at layer 4 in the improved backbone network is used as the first input graph of the improved neck structure; the output of the improved C2f module at layer 6 in the improved backbone network is used as the second input graph of the improved neck structure; and the output of the SPPF module at layer 9 in the improved backbone network is used as the third input graph of the improved neck structure. The first, second, and third input images are input into the MFC module to obtain the first feature image. The first feature image is convolved to obtain the second feature image. The first feature image is upsampled to obtain the third feature image. The second feature image is concatenated with the third input image to obtain the fourth feature image. The third feature image is concatenated with the first input image to obtain the fifth feature image. The 4th and 5th feature maps are input into the improved C2f module to obtain the 6th and 7th feature maps, respectively. The 1st, 6th, and 7th feature maps are input into the MFC module to obtain the 8th feature map. The 8th feature map is convolved and then concatenated with the 2nd and 6th feature maps to obtain the 9th feature map. The 8th feature map is upsampled and then concatenated with the 3rd and 7th feature maps to obtain the 10th feature map. The 8th, 9th, and 10th feature maps are used as the output for improving the neck structure.

[0041] In one implementation, see [link to implementation details]. Figure 2 , Figure 2 This is a model framework diagram for assessing vascular fibrosis burden based on intravascular ultrasound data provided in this embodiment of the invention. The improved neck structure enhances the integration and flow of multi-scale information by constructing a multi-layered, densely interactive feature fusion path. Features from the shallow, intermediate, and deep layers of the backbone network are input into the MFC module for initial fusion, achieving cross-scale contextual information alignment and convergence. This generates a base feature map rich in multi-dimensional information. By convolving and upsampling this base feature map and concatenating it across layers with the original input features, the deep semantic information is enhanced with detail, and the shallow detail information also receives semantic supplementation.

[0042] In one implementation, an improved C2f module is used to efficiently extract and optimize the aggregated dual-path features, further enhancing their feature representation capabilities. Then, the MFC module is used to perform a secondary fusion of the initial fused features and the features from the two optimized paths, achieving deep interaction and refinement of multi-level features. Through convolution, upsampling, and final concatenation operations, three closely related and complementary feature maps are output. By performing two MFC fusions and multiple cross-layer connections, the diversity and robustness of the feature pyramid are enhanced, improving the model's detection performance for targets at different scales.

[0043] In one embodiment, improving the workflow of the C2f module includes: Obtain the original tensor, input the original tensor into the Conv module to obtain the first tensor, split the first tensor into the first branch tensor and the second branch tensor, input the first branch tensor into the bottleneck module to obtain the third branch tensor, concatenate the second branch tensor and the third branch tensor to obtain the second tensor, input the second tensor into the Conv module to obtain the third tensor, and use the third tensor as the output of the improved C2f module; The workflow of an MFC module includes: Obtain the first feature tensor, the second feature tensor, and the third feature tensor. Downsample the first feature tensor to obtain the fourth feature tensor. Convolve the second feature tensor to obtain the fifth feature tensor. Upsample the third feature tensor and then convolve it to obtain the sixth feature tensor. The 4th, 5th, and 6th feature tensors are concatenated to obtain the 7th feature tensor. The 7th feature tensor is then input into the target DWConv module to obtain the 8th, 9th, 10th, and 11th feature tensors, respectively. The kernel size of the target DWConv module can be 5, 7, 9, or 11. The 7th, 8th, 9th, 10th, and 11th feature tensors are concatenated and then convolved to obtain the 12th feature tensor, which is then used as the output of the MFC module.

[0044] In one implementation, the improved C2f module first performs preliminary feature extraction on the input tensor through a convolutional layer (Conv), obtaining the first tensor. This tensor is then split into two branches. The first branch uses a bottleneck module for cross-layer feature transformation and compression, extracting deep abstract features and reducing the number of parameters. The second branch retains the original features to maintain shallow detail information. Next, the feature tensors from the two branches are concatenated and fused, combining the complementary advantages of deep and shallow features to enhance feature diversity and expressive power. Finally, a convolutional layer is used to integrate and reduce the dimensionality of the fused features, outputting the enhanced third tensor. While maintaining lightweight design, this approach strengthens the cross-layer propagation of gradients and feature reuse capabilities, improving the model's feature capture and representation performance for multi-scale targets, especially small targets.

[0045] In one implementation, the MFC module first performs downsampling, convolution, and upsampling operations on feature maps at different levels to achieve size alignment and preliminary feature enhancement, ensuring that multi-scale information can be fused and interacted. Subsequently, the fused features are input into a depthwise separable convolutional (DWConv) layer containing multiple large convolutional kernels (5, 7, 9, 11) for parallel processing. This efficiently captures rich multi-scale contextual information through different receptive fields, particularly detailed features of small targets and semantic features of distant targets. By concatenating and fusing the outputs of all branches, this module integrates complementary information from multiple scales and receptive fields, enhancing the feature representation capability. Finally, it outputs an enhanced feature tensor that integrates fine local details and global semantic context, improving the model's detection performance for different scales, especially small and distant targets.

[0046] In one embodiment, the fiber load fraction is calculated based on the identification result, the method comprising: The area of ​​fibrous tissue and the longitudinal section area of ​​blood vessels are determined based on the identification results, and the fibrous load fraction is calculated based on the area of ​​fibrous tissue and the longitudinal section area of ​​blood vessels. Formula for calculating fibrosis load: *100%; Where F is the fiber load fraction, F0 is the area of ​​fibrous tissue, and F1 is the area of ​​the longitudinal section of the blood vessel.

[0047] In one implementation, see [link to implementation details]. Figure 3 , Figure 3 This is a schematic diagram of a vascular fibrosis burden assessment method based on intravascular ultrasound data provided in an embodiment of the present invention. 301 represents the longitudinal cross-sectional area of ​​the vessel, 302 represents the area of ​​fibrous tissue, and 303 represents the sampled vessel. The units for both the fibrous tissue area and the longitudinal cross-sectional area of ​​the vessel are mm. 2 The fibrous tissue area refers to the region within a blood vessel where fibrous tissue exists. The longitudinal section area of ​​a blood vessel is the unit length of fibrous tissue. The area of ​​this segment of the blood vessel is separated based on the unit length, and the area of ​​this segment is calculated as the longitudinal section area of ​​the blood vessel, which is also the minimum annular area of ​​this segment of the blood vessel. The fibrous load fraction has no unit; the larger the value, the higher the degree of fibrosis.

[0048] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A method for assessing vascular fibrosis burden based on intravascular ultrasound data, characterized in that, The method includes: Acquire ultrasound images and a stitched image set, and calculate the similarity between the ultrasound images and each stitched image in the stitched image set; the stitched images in the stitched image set are obtained by adding noise to standard ultrasound images; Determine the target time label corresponding to the stitched image with the highest similarity, and convert the target time label to obtain the target time label feature code; The ultrasonic image and the target time-label feature encoding are input into the U-net network to obtain the predicted noise. The ultrasonic image is then de-denoised using the predicted noise to obtain the denoised ultrasonic image. The denoised ultrasonic image is identified using a target recognition model to obtain the recognition result; The fiber load score is calculated based on the identification results, and an early warning is issued based on the fiber load score.

2. The method for assessing vascular fibrosis burden based on intravascular ultrasound data according to claim 1, characterized in that, Before acquiring ultrasound images, the method further includes: Step 1: Obtain a standard ultrasound image; add multiplicative speckle to the standard ultrasound image using a first formula to obtain a speckled standard ultrasound image; add Gaussian noise to the standard ultrasound image using a second formula to obtain a Gaussian standard ultrasound image; and fuse the speckled standard ultrasound image and the Gaussian standard ultrasound image to obtain a fused standard ultrasound image. Step 2: Obtain the total number of diffusion steps and time stamps. Based on the total number of diffusion steps and time stamps, gradually add perturbations to the fused standard ultrasound image to obtain the final perturbation image and time stamp image set; the time stamp image set contains multiple perturbation images. Step 3: Extract the perturbation images from the time-stamped image set to obtain the target time-stamped image set. Then, stitch the fused standard ultrasonic image with each perturbation image in the target time-stamped image set according to the channel to obtain the stitched image set. Step 4: Obtain the time label corresponding to each stitched image in the stitched image set, convert the time label to obtain time label feature code, input the stitched image and the corresponding time label feature code into the U-net network to obtain prediction noise, and average the prediction noise to obtain the average prediction noise. Step 5: Perform reverse denoising on the fused standard ultrasound image using the average predicted noise to obtain a denoised standard ultrasound image. Calculate the difference between the standard ultrasound image and the denoised standard ultrasound image. If the difference is less than a preset threshold, the denoising is deemed successful; otherwise, the denoising is deemed unsuccessful, and the model parameters of the U-net network are updated. Repeat steps 1 to 5 until the denoising is deemed successful, then determine the current model parameters.

3. The method for assessing vascular fibrosis burden based on intravascular ultrasound data according to claim 1, characterized in that, The denoised ultrasonic image is identified using a target recognition model to obtain the recognition result. The target recognition model is an improvement based on the YOLOv8 model, which includes: The improved backbone network is obtained by replacing each C2f module in the backbone network with an improved C2f module, and the target recognition model is obtained by replacing the neck structure with an improved neck structure; the YOLOv8 model includes a backbone network and a neck structure. The working principle of the improved neck structure includes: The output of the improved C2f module at layer 4 in the improved backbone network is used as the first input diagram of the improved neck structure, the output of the improved C2f module at layer 6 in the improved backbone network is used as the second input diagram of the improved neck structure, and the output of the SPPF module at layer 9 in the improved backbone network is used as the third input diagram of the improved neck structure. The first input image, the second input image, and the third input image are input into the MFC module to obtain the first feature image. The first feature image is convolved to obtain the second feature image. The first feature image is upsampled to obtain the third feature image. The second feature image is concatenated with the third input image to obtain the fourth feature image. The third feature image is concatenated with the first input image to obtain the fifth feature image. The fourth and fifth feature maps are input into the improved C2f module to obtain the sixth and seventh feature maps, respectively. The first, sixth, and seventh feature maps are input into the MFC module to obtain the eighth feature map. The eighth feature map is convolved and then concatenated with the second and sixth feature maps to obtain the ninth feature map. The eighth feature map is upsampled and then concatenated with the third and seventh feature maps to obtain the tenth feature map. The 8th feature map, the 9th feature map, and the 10th feature map are used as the output of the improved neck structure.

4. The method for assessing vascular fibrosis burden based on intravascular ultrasound data according to claim 3, characterized in that, The improved C2f module's workflow includes: Obtain the original tensor, input the original tensor into the Conv module to obtain the first tensor, split the first tensor to obtain the first branch tensor and the second branch tensor, input the first branch tensor into the bottleneck module to obtain the third branch tensor, concatenate the second branch tensor and the third branch tensor to obtain the second tensor, input the second tensor into the Conv module to obtain the third tensor, and use the third tensor as the output of the improved C2f module; The workflow of the MFC module includes: Obtain the first feature tensor, the second feature tensor, and the third feature tensor. Downsample the first feature tensor to obtain the fourth feature tensor. Convolve the second feature tensor to obtain the fifth feature tensor. Upsample the third feature tensor and then convolve it to obtain the sixth feature tensor. The 4th, 5th, and 6th feature tensors are concatenated to obtain the 7th feature tensor. The 7th feature tensor is then input into the target DWConv module to obtain the 8th, 9th, 10th, and 11th feature tensors, respectively. The kernel size of the target DWConv module can range from 5, 7, 9, to 11. The 7th, 8th, 9th, 10th, and 11th feature tensors are concatenated and then convolved to obtain the 12th feature tensor, which is then used as the output of the MFC module.

5. The method for assessing vascular fibrosis burden based on intravascular ultrasound data according to claim 1, characterized in that, The fiber load fraction is calculated based on the identification results, and the method includes: The area of ​​fibrous tissue and the longitudinal section area of ​​blood vessels are determined based on the identification results, and the fibrous load fraction is calculated based on the area of ​​fibrous tissue and the longitudinal section area of ​​blood vessels. Formula for calculating fibrosis load: *100%; Where F is the fiber load fraction, F0 is the area of ​​fibrous tissue, and F1 is the area of ​​the longitudinal section of the blood vessel.

6. A vascular fibrosis burden assessment system based on intravascular ultrasound data, characterized in that, The system includes: An image acquisition module is used to acquire ultrasonic images and a stitched image set, and to calculate the similarity between the ultrasonic images and each stitched image in the stitched image set; the stitched images in the stitched image set are obtained by adding noise to standard ultrasonic images; The label conversion module is used to determine the target time label corresponding to the stitched image with the highest similarity, and convert the target time label to obtain the target time label feature code; The inverse denoising module is used to input the ultrasonic image and the target time label feature encoding into the U-net network to obtain predicted noise, and to perform inverse denoising on the ultrasonic image using the predicted noise to obtain a denoised ultrasonic image. The image recognition module is used to identify the denoised ultrasonic image using a target recognition model to obtain the recognition result; The image evaluation module is used to calculate the fiber load score based on the recognition result and to issue an early warning based on the fiber load score.

7. The vascular fibrosis burden assessment system based on intravascular ultrasound data according to claim 6, characterized in that, The system also includes: The first operation module is used to acquire a standard ultrasound image, add multiplicative speckle to the standard ultrasound image using a first formula to obtain a speckle standard ultrasound image, add Gaussian noise to the standard ultrasound image using a second formula to obtain a Gaussian standard ultrasound image, and fuse the speckle standard ultrasound image and the Gaussian standard ultrasound image to obtain a fused standard ultrasound image. The second operation module is used to obtain the total number of diffusion steps and time stamps, and to gradually add perturbations to the fused standard ultrasound image based on the total number of diffusion steps and time stamps to obtain the final perturbation image and time stamp image set; the time stamp image set contains multiple perturbation images; The third operation module is used to extract the perturbation images in the time-stamped image set to obtain the target time-stamped image set, and then combine the fused standard ultrasonic image with each perturbation image in the target time-stamped image set according to the channel to obtain the spliced ​​image set. The fourth operation module is used to obtain the time label corresponding to each stitched image in the stitched image set, convert the time label to obtain the time label feature code, input the stitched image and the corresponding time label feature code into the U-net network to obtain the prediction noise, and average the prediction noise to obtain the average prediction noise. The fifth operation module is used to reverse denoise the fused standard ultrasound image by averaging the predicted noise to obtain a denoised standard ultrasound image, calculate the difference between the standard ultrasound image and the denoised standard ultrasound image, and determine that the denoising is qualified if the difference is less than a preset threshold; otherwise, it determines that the denoising has failed and updates the model parameters of the U-net network, and re-executes the fourth to fifth operation modules until the denoising is determined to be qualified, then determines the current model parameters.

8. The vascular fibrosis burden assessment system based on intravascular ultrasound data according to claim 6, characterized in that, The denoised ultrasonic image is identified using a target recognition model to obtain the recognition result. The target recognition model is an improvement based on the YOLOv8 model, which includes: The improved backbone network is obtained by replacing each C2f module in the backbone network with an improved C2f module, and the target recognition model is obtained by replacing the neck structure with an improved neck structure; the YOLOv8 model includes a backbone network and a neck structure. The working principle of the improved neck structure includes: The output of the improved C2f module at layer 4 in the improved backbone network is used as the first input diagram of the improved neck structure, the output of the improved C2f module at layer 6 in the improved backbone network is used as the second input diagram of the improved neck structure, and the output of the SPPF module at layer 9 in the improved backbone network is used as the third input diagram of the improved neck structure. The first input image, the second input image, and the third input image are input into the MFC module to obtain the first feature image. The first feature image is convolved to obtain the second feature image. The first feature image is upsampled to obtain the third feature image. The second feature image is concatenated with the third input image to obtain the fourth feature image. The third feature image is concatenated with the first input image to obtain the fifth feature image. The fourth and fifth feature maps are input into the improved C2f module to obtain the sixth and seventh feature maps, respectively. The first, sixth, and seventh feature maps are input into the MFC module to obtain the eighth feature map. The eighth feature map is convolved and then concatenated with the second and sixth feature maps to obtain the ninth feature map. The eighth feature map is upsampled and then concatenated with the third and seventh feature maps to obtain the tenth feature map. The 8th feature map, the 9th feature map, and the 10th feature map are used as the output of the improved neck structure.

9. The vascular fibrosis burden assessment system based on intravascular ultrasound data according to claim 8, characterized in that, The improved C2f module's workflow includes: Obtain the original tensor, input the original tensor into the Conv module to obtain the first tensor, split the first tensor to obtain the first branch tensor and the second branch tensor, input the first branch tensor into the bottleneck module to obtain the third branch tensor, concatenate the second branch tensor and the third branch tensor to obtain the second tensor, input the second tensor into the Conv module to obtain the third tensor, and use the third tensor as the output of the improved C2f module; The workflow of the MFC module includes: Obtain the first feature tensor, the second feature tensor, and the third feature tensor. Downsample the first feature tensor to obtain the fourth feature tensor. Convolve the second feature tensor to obtain the fifth feature tensor. Upsample the third feature tensor and then convolve it to obtain the sixth feature tensor. The 4th, 5th, and 6th feature tensors are concatenated to obtain the 7th feature tensor. The 7th feature tensor is then input into the target DWConv module to obtain the 8th, 9th, 10th, and 11th feature tensors, respectively. The kernel size of the target DWConv module can range from 5, 7, 9, to 11. The 7th, 8th, 9th, 10th, and 11th feature tensors are concatenated and then convolved to obtain the 12th feature tensor, which is then used as the output of the MFC module.

10. The method for assessing vascular fibrosis burden based on intravascular ultrasound data according to claim 6, characterized in that, The image evaluation module is further configured to determine the area of ​​fibrous tissue and the longitudinal section area of ​​blood vessels based on the recognition results, and to calculate the fiber load score based on the area of ​​fibrous tissue and the longitudinal section area of ​​blood vessels. Formula for calculating fibrosis load: *100%; Where F is the fiber load fraction, F0 is the area of ​​fibrous tissue, and F1 is the area of ​​the longitudinal section of the blood vessel.

Images