3D hemodynamic atlas construction method for ultrasonic examination

Abstract

The invention relates to the field of intelligent medicine, in particular to a 3D hemodynamic atlas construction method for ultrasonic examination, and designs a method for modeling a global topological structure of a blood vessel through linear complexity of a state space model based on an optimized Unet model, capturing continuity of the blood vessel and a spatial relationship between branches, and constructing a 3D hemodynamic atlas for ultrasonic examination. According to the method, local features are extracted through non-overlapping windows, blood vessel edge response is enhanced in combination with space attention, through dynamic receptive field adjustment, the recognition and segmentation capacity for different blood vessels is enhanced, the spatial relation of blood vessel branches is scanned and modeled in multiple directions, the problems of information loss and wrong segmentation of one-way scanning are avoided, and accurate three-dimensional blood vessel modeling is achieved; according to the method, macroscopic and microcosmic parameters are quantified and fused into the atlas, artery information in the atlas is enriched, individual blood vessels are geometrically mapped to an ATLAS coordinate system through B spline registration, parameter deviation caused by aortic branch angle variation is solved, the registration precision is improved, and the atlas quality is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent medicine, and specifically relates to a 3D hemodynamic atlas construction method for ultrasonic examination. BACKGROUND

[0002] The construction background of the 3D hemodynamic atlas mainly involves the fields of medical image processing and hemodynamic simulation. This technology can help the medical field better understand and evaluate the blood flow in the human body, providing important information for disease diagnosis and treatment and providing a more intuitive display of blood vessels and blood flow for patients. The existing algorithms and technologies still have limitations in constructing the 3D hemodynamic atlas, and it is difficult to accurately construct 3D blood vessels according to medical images, and the quality of atlas construction based on dynamics is low. SUMMARY

[0003] In view of the above problems, in order to overcome the defects of the prior art, the application provides a 3D hemodynamic atlas construction method for ultrasonic examination. In view of the problem that the existing technology is difficult to accurately construct 3D blood vessels according to medical images, and the quality of atlas construction based on dynamics is low, the application designs an optimized Unet model. The linear complexity of the state space model (SSM) is used to model the global topological structure of the blood vessels, capture the continuity of the blood vessels and the spatial relationship between the branches, reduce the computational complexity, extract local features through non-overlapping windows, combine spatial attention to enhance the edge response of the blood vessels, adjust the window size through dynamic receptive field, improve the compatibility of thick and thin blood vessels, enhance the recognition and segmentation ability of different blood vessels, and explicitly model the spatial relationship of the blood vessel branches through multi-directional scanning, avoid information loss caused by one-way scanning, solve the problem of missegmentation at the intersection and bifurcation of the blood vessels in the traditional technology, and finally realize accurate 3D blood vessel segmentation modeling. The application quantifies macro (flow velocity, kinetic energy) and micro parameters (wall shear stress, vorticity) at the same time and integrates them into the 3D atlas, increases the atlas information, makes the aneurysm rupture risk prediction more sensitive, maps the individual blood vessel geometry to the ATLAS coordinate system through B-spline registration, solves the parameter deviation caused by the angle variation of the aortic branches, increases the accuracy of registration, and improves the atlas quality.

[0004] The application provides a 3D hemodynamic atlas construction method for ultrasonic examination, which specifically comprises the following steps:

[0005] Step S1: data acquisition, acquiring three-dimensional ultrasonic images of the part to be constructed, including blood flow velocity vector field;

[0006] Step S2: data preprocessing, preprocessing the three-dimensional ultrasonic images and slicing them into two-dimensional images, preprocessing the two-dimensional images and converting them into grayscale images, copying the grayscale images to three channels and converting them into three-channel images;

[0007] Furthermore, step S2 specifically includes the following steps:

[0008] Step S21: removing labels of irrelevant structures and small objects in the 3D ultrasound image;

[0009] Step S22: normalizing the voxel intensity values ​​of the three-dimensional ultrasound image using a window setting of [400, 40];

[0010] Step S23: converting the 3D ultrasound image into a 2D slice, i.e., a 2D image;

[0011] Step S24: resample all slices to a uniform spatial resolution of 1 mm × 1 mm;

[0012] Step S25: retain non-zero slices and perform intensity normalization on the two-dimensional image;

[0013] Step S26: resizing the processed two-dimensional image to 1024×1024 pixels;

[0014] Step S27: converting the two-dimensional image into a grayscale image, and converting the grayscale image into a three-channel image by copying the grayscale image into three channels (1024×1024×3);

[0015] Step S3: segmenting the three-channel image based on the optimized Unet model to obtain a blood vessel segmentation map;

[0016] Step S31: Image segmentation and linear projection: input a three-channel image, segment it into 4×4 non-overlapping blocks, and obtain a feature map by linearly projecting the non-overlapping blocks;

[0017] Step S32: Encoder encoding, using a visual state space module based on receptive field optimization to encode the feature map to obtain a primary encoding result;

[0018] Step S321: Global modeling, using the visual state space model VSSM based on residual optimization to perform global modeling on the feature map to obtain global features: ;

[0019] Where, represents the visual state space model, represents the normalization operation, is a learnable scaling factor, Represents the feature map, adjusts the residual connection strength, is a global feature;

[0020] Step S322: two-dimensional selective scanning, expanding the feature map into a one-dimensional sequence in four directions, respectively processing Mamba, and then merging into a 2D feature: ;

[0021] In the formula, is a feature expansion operation, represents the expansion direction, represents a merging operation, represents a 2D feature;

[0022] Step S323: local feature enhancement based on non-overlapping windows, extracting a receptive field through a non-overlapping window and combining spatial attention to perform local feature enhancement on the 2D feature to obtain a local enhanced feature: ;

[0023] In the formula, represents a grouped convolution with a convolution kernel size of k*k, represents an average pooling process, represents a standardization process, and represents an activation function, is a local enhanced feature;

[0024] Step S324: dynamic fusion of features, dynamically fusing the global feature and the local enhanced feature to obtain a primary encoding result;

[0025] Step S33: down-sampling, down-sampling the primary encoding result to obtain a primary down-sampled result;

[0026] Step S34: multi-layer down-sampling encoding, repeating steps S32 to S33 three times to obtain a down-sampling encoding;

[0027] Step S35: multi-layer up-sampling encoding, using one of the above visual state space modules based on receptive field optimization and Patch Expanding to perform 4-layer up-sampling on the down-sampling encoding to obtain a decoding feature;

[0028] Step S36: restoring the decoding feature to the original three-channel image size through a linear layer, and outputting a blood vessel segmentation map;

[0029] Step S4: 3D reconstruction, stitching the blood vessel segmentation maps of all two-dimensional images to obtain a 3D blood vessel segmentation map;

[0030] Step S5: blood flow velocity field reconstruction, reconstructing the blood flow velocity vector field to obtain blood flow dynamics parameters and blood flow paths:

[0031] Step S51: Doppler angle correction, combined with the geometric parameters of the three-dimensional ultrasound probe, the blood flow velocity vector field is recovered: ;

[0032] wherein, represents the nodal velocity of the blood flow velocity vector field, is the blood flow direction vector, is the acoustic beam direction vector, is the included angle, wherein is estimated by three-dimensional probe posture data , represents the recovered velocity;

[0033] Step S52: Radial basis function interpolation is used to fill in the missing values of the blood flow velocity vector field;

[0034] Step S53: Blood flow dynamics parameter calculation, the wall shear stress WSS and the vorticity of the blood vessel surface are calculated: , ;

[0035] wherein, represents the blood viscosity, represents the tangential velocity in the blood flow velocity vector field, represents the wall shear stress, represents the spatial gradient of the blood flow velocity vector field;

[0036] Step S54: Blood flow path construction, the streamlines are generated by integrating along the velocity field to describe the blood flow path;

[0037] Step S6: Atlas construction based on B-spline transformation, the blood flow dynamics parameters and the blood flow path are mapped to the surface of the 3D blood vessel segmentation graph to generate a 3D blood flow dynamics atlas: ;

[0038] wherein, represents the B-spline basis function, represents the control point, is the result of mapping the wall shear stress WSS, the velocity and the vorticity of the blood flow dynamics parameters to the 3D blood vessel segmentation graph.

[0039] The beneficial results achieved by the application using the above scheme are as follows:

[0040] (1) The application designs an optimized Unet model, models the global topological structure of blood vessels through the linear complexity of the state space model (SSM), captures the continuity of blood vessels and the spatial relationship between branches, reduces the computational complexity, extracts local features through non-overlapping windows, combines spatial attention to enhance the response of blood vessel edges, adjusts the window size according to the diameter of the blood vessel through dynamic receptive field adjustment, improves the compatibility of thick and thin blood vessels, enhances the recognition and segmentation ability of different blood vessels, and explicitly models the spatial relationship of blood vessel branches through multi-directional scanning, avoids the information loss of one-way scanning, solves the problem of easy missegmentation at the intersection and bifurcation of blood vessels in the prior art, and finally realizes accurate 3D blood vessel segmentation modeling;

[0041] (2) The application quantifies macro (flow rate, kinetic energy) and micro parameters (wall shear stress, vorticity) at the same time and integrates them into a 3D atlas, increases the atlas information, makes it more sensitive to aneurysm rupture risk prediction, maps individual blood vessel geometry to the ATLAS coordinate system through B-spline registration, solves the parameter deviation caused by the angle variation of the aortic branch, increases the accuracy of registration, and improves the atlas quality. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A flowchart of a 3D hemodynamic atlas construction method for ultrasonic examination provided by the application is provided.

[0043] Figure 2 A data processing flowchart of a visual state space module based on receptive field optimization provided by the application.

[0044] The accompanying drawings are used to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0046] Embodiment one, refer to Figure 1 A 3D hemodynamic atlas construction method for ultrasonic examination, specifically comprising the following steps:

[0047] Step S1: Collecting three-dimensional ultrasonic images of the required atlas construction site, including blood flow velocity vector field;

[0048] Step S2: pre-process the three-dimensional ultrasound image and slice it into two-dimensional images, pre-process the two-dimensional images and convert them into grayscale images, and copy the grayscale images to three channels to convert them into three-channel images;

[0049] Step S3: segment the three-channel images using an optimized Unet model to obtain a blood vessel segmentation map;

[0050] Step S4: stitch the blood vessel segmentation maps of all two-dimensional images to obtain a 3D blood vessel segmentation map;

[0051] Step S5: reconstruct the blood flow velocity vector field to obtain blood flow dynamics parameters and blood flow paths;

[0052] Step S6: map the blood flow dynamics parameters and blood flow paths to the surface of the 3D blood vessel segmentation map to generate a 3D blood flow dynamics atlas, wherein the blood flow dynamics parameters are 3D mapped based on B-splines: ;

[0053] In the formula, represents the B-spline basis function, represents the control point, is the result of mapping the wall shear stress WSS, velocity, and vorticity of the blood flow dynamics parameters to the 3D blood vessel segmentation map.

[0054] In Example Two, based on the above-mentioned example, step S3 specifically includes the following steps:

[0055] Step S31: input the three-channel image and divide it into 4x4 non-overlapping blocks, and obtain a feature map by linearly projecting the non-overlapping blocks;

[0056] Step S32: encoder encoding, using a receptive field optimization-based visual state space module to encode the feature map to obtain a primary encoding result;

[0057] Step S33: downsample the primary encoding result to obtain a primary downsampled result;

[0058] Step S34: multi-layer downsampled encoding, repeating steps S32 to S33 three times to obtain a downsampled encoding;

[0059] Step S35: multi-layer upsampled encoding, using the above-mentioned receptive field optimization-based visual state space module and Patch Expanding to perform 4-layer upsampled encoding on the downsampled encoding to obtain a decoded feature;

[0060] Step S36: restore the decoded feature to the original three-channel image size through a linear layer to output a blood vessel segmentation map;

[0061] The above scheme is applied to brain vessel segmentation:

[0062] Task background: Early diagnosis of intracranial aneurysm relies on high-precision brain vessel segmentation, but MRA (magnetic resonance angiography) data usually has low contrast and small vessel rupture problems;

[0063] 3D improvement: A 3D improved design of visual state space module based on receptive field optimization:

[0064] Expand 2D four-direction scanning to 3D sequence scanning of six directions (x / y / z axis bidirectional), model the spatial continuity of blood vessels: ;

[0065] Use 3D grouped convolution (kernel size 3x3x3) to extract local blood vessel features, and change the attention map calculation to 3D average pooling;

[0066] Network structure:

[0067] Input: 3D MRA volume data (256x256x128);

[0068] Encoder: 4-stage downsampling (3D Patch Merging), channel number of each stage [96, 192, 384, 768];

[0069] Skip connection: Concatenate the 3D features of the encoder with the corresponding layers of the decoder to preserve the spatial information of the blood vessel branches.

[0070] Example three, see Figure 2 , the embodiment is based on the above embodiment, step S32, specifically comprising the following steps:

[0071] Step S321: Use a visual state space model VSSM based on residual optimization to globally model the feature map to obtain global features: ;

[0072] In the formula, represents the visual state space model, represents the normalization operation, is a learnable scaling factor, represents the feature map, and adjusts the residual connection strength, is the global feature;

[0073] Step S322: Expand the feature map into a one-dimensional sequence in four directions, respectively process it with Mamba, and then merge it into a 2D feature: ;

[0074] In the formula, for feature expansion operation, representing the direction of expansion, representing the merge operation, representing 2D features;

[0075] Step S323: local feature enhancement is performed on the 2D features by extracting the receptive field through a non-overlapping window and combining the spatial attention, to obtain local enhanced features: ;

[0076] In the formula, representing group convolution with a convolution kernel size of k*k, representing average pooling processing, representing standardization processing, and representing an activation function, local enhanced features;

[0077] Step S324: dynamic fusion is performed on the global features and the local enhanced features to obtain a primary encoding result;

[0078] The above scheme is applied to coronary CTA segmentation (quantification of stenosis lesions)

[0079] Task background: In coronary CTA, calcified plaques and vascular stenosis cause local gray level mutations, and traditional methods are prone to over-segmentation;

[0080] (1) Dynamic receptive field adjustment:

[0081] Variable kernel group convolution: adaptively select the convolution kernel size (1x1x1 to 5x5x5) according to the vessel radius, use small kernel to retain details for narrow blood vessels, and use large kernel to suppress noise for wide blood vessels:

[0082] Original formula: ;

[0083] Optimization: ; representing the local vessel diameter, which needs to be standardized before calculation to adapt to the log function;

[0084] (2) Multi-modal input fusion:

[0085] Dual-path encoder: parallel processing of CTA original data and calcification enhanced map (generated by threshold segmentation), and fusion of features in a visual state space module based on receptive field optimization.

[0086] Embodiment four, based on the above embodiment, an ultrasound-based aortic 3D hemodynamic atlas is constructed, which is used for early identification of vulnerable plaques and prediction of acute events (such as plaque rupture):

[0087] Data acquisition

[0088] Device: 3D ultrasound (TEE) probe, frequency 5 MHz, scan from ascending to descending aortic segment;

[0089] Parameters: Acquire B-mode and Doppler data for 10 consecutive cardiac cycles, spatial resolution 0.4x0.4x0.5mm³, frame rate 25Hz;

[0090] Preprocessing and segmentation

[0091] Filtering: Perona-Malik model (κ=0.1) is used to suppress blood flow speckle noise and preserve vessel wall structure;

[0092] Segmentation: Aorta lumen is initialized based on optimized Unet model (step S3) (Dice=0.94), and level set evolution is used to correct calcified plaque boundary (50 iterations);

[0093] Hemodynamic calculation

[0094] Key parameters:

[0095] Low shear stress region: Mark the vulnerable plaque area;

[0096] Oscillatory shear index OSI: Calculate the change of time-varying shear direction, OSI>0.3 indicates the risk of inflammation;

[0097] Blood flow residence time: Simulate by particle tracking (blood flow velocity vector field), areas with residence time>300ms indicate the tendency of thrombosis;

[0098] Atlas construction and verification

[0099] Registration: Register 50 patient data to the standard aortic ATLAS (B-spline control point interval 10mm);

[0100] Statistical atlas: Generate probability distribution heat map of low WSS and high OSI;

[0101] Clinical output:

[0102] Risk atlas: Visualize 3D "high-risk red area" (low WSS+high OSI+long residence time), guide intervention decision.

[0103] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it is intended to be limited only by the claims.

[0104] While the embodiments of the application have been shown and described, it is to be understood that the embodiments described are only by way of example and that changes can be made to the embodiments without departing from the spirit and scope of the application. The scope of the application is limited only by the claims and the equivalents thereof.

[0105] The above description of the application and its embodiments is not intended to limit the application, as described by the appended claims, and their equivalents.

Claims

1. A method for constructing a 3D hemodynamic atlas for ultrasound examination, characterized by: The specific steps include: Step S1: Acquire a three-dimensional ultrasound image of the area where the atlas is to be constructed, including a blood flow velocity vector field; Step S2: pre-processing the three-dimensional ultrasound image and slicing it into a two-dimensional image, pre-processing the two-dimensional image and converting it into a grayscale image, and copying the grayscale image into three channels to convert it into a three-channel image; Step S3: segmenting the three-channel image using the optimized Unet model to obtain a blood vessel segmentation map; Step S4: splicing the blood vessel segmentation maps of all two-dimensional images to obtain a 3D blood vessel segmentation map; Step S5: reconstructing the blood flow velocity vector field to obtain hemodynamic parameters and blood flow path; Step S6: Map the hemodynamic parameters and blood flow paths to the surface of the 3D blood vessel segmentation map to generate a 3D hemodynamic atlas, wherein the hemodynamic parameters are 3D mapped based on B-spline: ； Where, represents the B-spline basis function, represents the control point, This is the result of mapping the hemodynamic parameters of wall shear stress WSS, velocity, and vorticity into the 3D blood vessel segmentation map.

2. The method for constructing a 3D hemodynamic atlas for ultrasound examination according to claim 1, characterized in that: Step S3 specifically includes the following steps: Step S31: Input a three-channel image, divide it into 4×4 non-overlapping blocks, and obtain a feature map by linearly projecting the non-overlapping blocks; Step S32: Encoder encoding, using a visual state space module based on receptive field optimization to encode the feature map to obtain a primary encoding result; Step S33: downsampling the primary encoding result to obtain a primary downsampling result; Step S34: multi-layer downsampling coding, repeating steps S32 to S33 three times to obtain downsampling coding; Step S35: multi-layer upsampling encoding, using the above-mentioned visual state space module based on receptive field optimization and Patch Expanding to perform 4-layer upsampling on the downsampled code to obtain decoding features; Step S36: Restore the decoded features to the original three-channel image size through a linear layer and output a blood vessel segmentation map.

3. The method for constructing a 3D hemodynamic map for ultrasound examination according to claim 2, characterized in that: Step S32 specifically includes the following steps: Step S321: Use the visual state space model VSSM based on residual optimization to perform global modeling on the feature map to obtain global features: ； Where, represents the visual state space model, represents the normalization operation, is a learnable scaling factor, Represents the feature map, adjusts the residual connection strength, is a global feature; Step S322: Expand the feature map into one-dimensional sequences in four directions, perform Mamba processing on each, and then merge them into 2D features: ； Where, To expand the feature, Represents the direction of expansion, Represents a merge operation, Represents 2D features; Step S323: Extract the receptive field through non-overlapping windows and perform local feature enhancement on the 2D features in combination with spatial attention to obtain local enhanced features: ； Where, Represents a grouped convolution with a convolution kernel size of k*k, represents average pooling processing, Stands for standardized processing, and represents the activation function, It is a local enhancement feature; Step S324: Dynamically fuse the global features and the local enhanced features to obtain a primary coding result.

Images