Urine stone positioning and crushing path optimization method based on AI ultrasonic image analysis

Abstract

The application relates to the field of ultrasonic diagnosis, in particular to a urinary calculus positioning and fragmentation path optimization method based on AI ultrasonic image analysis, which comprises the following steps: acquiring a multimodal image feature set, calculating registration parameters and generating a calculus image feature, extracting an anatomical relationship graph and an energy curve, searching and optimizing a shock wave path, and dynamically adjusting a shock wave parameter.In the application, residual network is used to extract multimodal image structure features, a dense optical flow method is used to model registration error displacement, the accuracy of a registration parameter matrix is improved, three-dimensional convolution is used to enhance calculus image volume and density expression, a convolutional neural network is used to extract topological node structure differences, a high-precision anatomical relationship graph is formed, finite element analysis is used to calculate the energy distribution of a shock wave incident angle and tissue impedance, a genetic algorithm is used to search a path combination, a graph search algorithm is used to output a generation value sequence, a step angle is adjusted to optimize path nodes and direction connection, and the output result has adaptability of incident direction and energy control parameters.

Description

Technical Field

[0001] This invention relates to the field of ultrasound diagnostic technology, and in particular to a method for locating urinary stones and optimizing their fragmentation path based on AI ultrasound image analysis. Background Technology

[0002] The field of ultrasound diagnostic technology includes image acquisition and analysis based on the reflection characteristics of sound waves, used to identify, assess, and locate lesions or abnormal structures in the body. The core content is to use the reflected echoes generated by ultrasound waves at different tissue interfaces to perform real-time imaging through imaging equipment, so that medical personnel can perform clinical diagnosis under non-invasive conditions. Ultrasound diagnosis is widely used in cardiovascular, obstetrics and gynecology, abdominal, and urinary system fields. Relying on the penetration and reflection characteristics of high-frequency sound waves, combined with image processing and analysis, it realizes the detection and identification of tissue structures. With the introduction of multimodal image fusion and artificial intelligence, ultrasound diagnosis has gradually developed into a comprehensive technology system with high-precision identification and decision support functions.

[0003] Among them, the method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis involves extracting the location, volume, and hardness features of stones from multimodal images using a deep learning convolutional neural network model. A shock wave-tissue interaction model is then established using finite element analysis. Based on the relationship between energy attenuation rate and stone distribution, a heuristic search algorithm is used to dynamically plan the shock wave path, achieving precise stone location and path optimization. The technical aspects covered include grayscale normalization of multimodal image data, noise reduction, multi-scale feature enhancement, and spatial calibration. An adaptive weighted fusion is used to generate a fused image with anatomical and functional information. Subsequently, a three-dimensional convolutional neural network is applied to extract stone features, and a non-rigid transformation is constructed based on Lie group theory to achieve multimodal image registration, ultimately completing the optimized planning of the shock wave path.

[0004] Current ultrasound imaging analysis relies on the reflection characteristics of a single mode of sound waves for image acquisition. However, due to the varying penetration capabilities of ultrasound waves in different tissues, there are issues with insufficient image resolution and positioning errors. This is especially true when the stone is small or located near complex anatomical structures, making high-precision identification difficult. Image registration based on traditional experience often results in accumulated spatial errors, leading to deviations in stone positioning. Image feature extraction relies heavily on manual or semi-automatic boundary recognition, lacking deep feature learning capabilities. Shockwave path planning uses fixed incident angles and energy intensities, failing to fully consider individual tissue differences and energy attenuation patterns. This results in energy waste or accidental injury during treatment, increasing the incidence of postoperative complications and reducing treatment efficiency. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method for locating urinary stones and optimizing their fragmentation path based on AI ultrasound image analysis. The technical solution is as follows:

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis, comprising the following steps: S1: By acquiring multimodal medical image data from CT, MRI, and PET, feature extraction is performed on the medical images using a residual network. Gray-level histogram normalization is performed on kidney and ureter images, and edge detection is performed using the Laplacian operator to generate a multimodal image feature set. The method of using residual networks to extract features from medical images includes the following training process: CT, MRI, and PET multimodal medical images are uniformly processed to a standard size, and label maps of the kidney and ureter regions are constructed based on expert annotations as supervision signals. During the training phase, ResNet-34 is used as the base model, and ImageNet pre-trained weights are used for transfer learning initialization. The network output and the real labels are supervised and trained using the cross-entropy loss function or Dice Loss. The Adam optimizer is used, with an initial learning rate of 0.001, 100–200 training epochs, and a batch size of 8–16. Data augmentation strategies such as rotation, scaling, and contrast enhancement are used to improve the robustness of the network to different organ morphologies and grayscale changes. During training, the learning rate is dynamically adjusted based on the IoU and Dice coefficients of the validation set, and an EarlyStopping strategy is set to prevent overfitting. After training, the model is exported in ONNX format, input into medical images, and high-dimensional feature maps are extracted. S2: Based on the multimodal image feature set, the registration error is calculated using the dense optical flow method, and the registration parameter matrix is ​​obtained by combining cosine similarity comparison. Spatial affine transformation is then performed, and the stone image is processed using three-dimensional convolution operation to generate a stone image feature set. S3: Based on the stone image feature set, a densely connected convolutional neural network is called to extract the topology, obtain the stone boundary point set, blood vessel location point set, and nerve location point set, construct a stone anatomical relationship map, and use finite element analysis to calculate the energy attenuation value to generate a shock wave energy curve set. The method involves using a densely connected convolutional neural network (DCNN) to extract the topology. The training process of the DCNN includes: collecting and organizing a medical image dataset with stone annotation information. Data sources include stone edges, blood vessel locations, and nerve distribution areas marked in preoperative CT / MRI scans. Corresponding masked images are used as training supervision signals. The network structure uses the DenseNet-121 framework as its base. All convolutional layers achieve multi-layer feature reuse through feature map cascading, enhancing information flow and gradient propagation efficiency. During training, the input is a stone image block after registration and normalization, and the output is the three-dimensional coordinates and heatmap labels for each type of structural point, including stone edges, blood vessels, and nerves. The optimizer used is Adam with a learning rate of 0.0005, a batch size of 4, and a multi-channel Dice loss function. The combination of Loss and position error penalty term is used for 150 training rounds. Dropout and L2 regularization are introduced to avoid overfitting. After training, the stone image feature set is input, the model performs feature map mapping, normalization and ReLU activation, and outputs the three-dimensional spatial coordinate set of stone edge point set, blood vessel location point set and nerve location point set to obtain the topological structure coordinate matrix. S4: Based on the shock wave energy curve set, a genetic algorithm is used to perform multi-objective path search, the Dijkstra algorithm is called to calculate the optimal shock wave path and perform step angle adjustment to generate an optimized shock wave path set.

[0007] As a further embodiment of the present invention, the multimodal image feature set includes a unified gray level mapping value, organ edge response value, and spatial structure expression; the stone image feature set includes position coordinate coefficients, tissue density range, and morphological feature encoding; the shock wave energy curve set includes attenuation trend, path conduction characteristics, and nerve location interference weight; and the optimized shock wave path set includes angle offset range, path continuity index, and incident direction set.

[0008] As a further aspect of the present invention, step S1 specifically comprises: S101: By acquiring multimodal medical image data from CT, MRI, and PET, slice layering is performed on each set of image data using a uniform sampling interval. The image boundaries of the urinary structure region are identified and assigned numbered indices. The pixel matrix of the corresponding kidney and ureter images in the indexed images is extracted and mapped to a standard coordinate space to obtain a unified image index set. The standard coordinate space establishes a three-dimensional rectangular coordinate system with the highest point of the iliac crest as the origin; S102: Based on the unified image index set, the backbone residual block structure in the residual network is called to extract image region features. Combined with the pixel matrix, deformable convolution kernels are used for multimodal feature alignment, cross-channel normalization is performed, and features are updated through skip connections. Block-level channel statistics are used for feature value normalization analysis. Combined with the peak and mean values ​​in the same resolution group, feature contrast coefficients are generated to obtain the region feature contrast coefficient matrix. S103: Calculate the normalized difference rate between the region feature contrast coefficient matrix and the standard grayscale mapping histogram, and use the difference rate to call the Laplacian operator to perform edge detection and obtain a multimodal image feature set.

[0009] As a further aspect of the present invention, step S2 specifically includes: S201: Extract the grayscale change vector field between adjacent pixels of the image feature map in the multimodal image feature set using the dense optical flow method, perform pixel position displacement calculation on the vector field, convert the pixel displacement amount into the actual displacement value in millimeters, generate a global displacement vector matrix, calculate the modal image registration error value, and obtain the registration error value matrix. S202: Based on the registration error value matrix, compare the cosine similarity of each group of registration error value vectors with the standard reference vector, calculate the cosine value of the angle between the registration errors, and generate a matching parameter sequence by matching and combining the cosine values ​​to establish a registration parameter matrix; The standard reference vector is selected based on the characteristics of the intraoperative imaging modality, and the standard reference vector adopts the registration error vector corresponding to the intraoperative real-time ultrasound image with high real-time performance and anatomical accuracy as the benchmark reference modality.

[0010] S203: Based on the registration parameter matrix, perform spatial affine transformation to perform coordinate mapping and registration of modal images, construct the registration parameter matrix and dynamically adjust the channel attention weights of the three-dimensional convolution kernel, use anisotropic convolution kernels to extract spatial features of each axis, combine the convolution kernel weights to generate a feature space mapping matrix, and obtain the stone image feature set.

[0011] As a further aspect of the present invention, step S3 specifically comprises: S301: Based on the stone image feature set, a densely connected convolutional neural network is invoked to sequentially perform feature mapping, normalization and activation operations, perform multi-dimensional channel feature enhancement processing on the convolutional feature map, extract edge points, blood vessel location points and branch points in the topology, and perform geometric coordinate calibration on the extraction results to obtain the topology coordinate matrix. S302: Based on the topological coordinate matrix, perform Euclidean distance calculation on the stone boundary points, blood vessel location points and nerve location points, and construct a stone anatomical relationship atlas by combining the spatial topological relationships between each point set; S303: Based on the anatomical relationship atlas of the stone, the Young's modulus and Poisson's ratio of the stone tissue are obtained by preoperative CT value markers. Using finite element analysis, the incident angle and energy parameters are set, the energy attenuation value is calculated, and a set of shock wave energy curves is generated. In the finite element analysis, the ureteral wall is set as a fixed constraint boundary, the contact area between the stone and soft tissue is set as a free boundary, a shock wave excitation is applied to the incident surface, and the energy decay process is simulated by combining tissue damping parameters and material properties.

[0012] As a further aspect of the present invention, the specific formula for setting the incident angle and energy parameters and calculating the energy attenuation value is as follows: ; Calculate the energy decay value; in, This represents the attenuation value of the shock wave energy per unit volume. Representing the The angle between the incident direction of the shock wave and the normal. Representing the The initial energy of the incident shock wave, Representing the The propagation depth of the path, Representing the stone tissue in the Equivalent absorption coefficient under the path, Representing the The relative propagation volume of the stone in a local area along the path, Represents the total energy of the reference shock wave. This represents the index of the number of paths in a multi-path propagation scenario. This is a discrete index variable used for path numbering in the multipath propagation model of shock waves.

[0013] As a further aspect of the present invention, step S4 specifically comprises: S401: Based on the shock wave energy curve set, using the energy attenuation rate, incident angle and tissue impedance parameters, call the genetic algorithm to perform multi-target path search, screen path gene sequences, and obtain a set of path gene sequences; S402: Based on the set of path gene sequences, the Dijkstra algorithm is called to establish a weighted directed graph model, the node weights are set, the cost value of the path is calculated, the path region with the cost value between the energy consumption and tissue penetration equilibrium threshold is selected, and the shock wave path cost value matrix is ​​obtained. The set node weight is the sum of the shock wave energy consumption coefficient and the tissue penetration depth coefficient; The tissue penetration equalization threshold is set according to tissue impedance and thickness distribution, and is used to limit the penetration depth of shock waves to ensure tissue safety. S403: Based on the shock wave path cost matrix, the incident angle step is adjusted on the selected path, the average angle change of the path coordinate point set is extracted, the step threshold is set in combination with the tissue penetration direction gradient, the incident direction sequence is obtained, the adjusted incident direction sequence and the path coordinate point set are integrated, and the optimized shock wave path set is output. The step threshold is set based on the gradient of stone density distribution in the preoperative CT images.

[0014] As a further aspect of the present invention, the method further includes: S5: Based on the optimized shock wave path set, real-time acquisition of intraoperative ultrasound elastography monitoring images, image elastic modulus detection of the images, analysis of the relationship between elastic modulus change rate and output energy intensity, analysis of stone fragmentation degree based on image elastic modulus detection, dynamic adjustment of shock wave frequency and energy intensity based on stone fragmentation degree, and output of shock wave parameter adjustment set. The shock wave parameter adjustment set includes frequency control parameters, energy release level, and pulse period.

[0015] As a further aspect of the present invention, step S5 specifically includes: S501: Based on the optimized shock wave path set, intraoperative ultrasound elastography monitoring images are acquired in real time, and the elastic modulus of the stone region in the monitoring images is detected. By detecting the rebound coefficient and strain gradient of the pixel region, the overall elastic modulus distribution curve of the stone is calculated, and the elastic modulus distribution value of the stone is obtained. The elasticity coefficient was measured using the optical flow method. By performing inter-frame cross-correlation calculations on the radio frequency signal frames before and after the force was applied, the displacement of the local tissue under longitudinal compression was extracted, and its elastic recovery characteristics were further evaluated. S502: Based on the elastic modulus distribution value of the stone, perform gradient change analysis on the average elastic modulus of multiple regions, identify abnormal areas of stone morphology, and evaluate the degree of stone fragmentation by analyzing the relationship between the rate of change of elastic modulus and the output energy intensity and according to the amplitude and range of the elastic change interval, and generate stone fragmentation state parameters. S503: Based on the stone fragmentation state parameters, according to the current degree of stone fragmentation, the voltage parameters of the shock wave transmitter are adjusted in real time by using a PID controller to dynamically adjust the frequency and energy intensity of the shock wave and output the shock wave parameter adjustment set.

[0016] As a further aspect of the present invention, the specific formula for calculating the overall elastic modulus distribution curve of the stone by detecting the rebound coefficient and strain gradient of the pixel region is as follows: ; Calculate the elastic modulus distribution of the gallstone; in, This represents the overall elastic modulus distribution value of the stone region. For the first The rebound coefficient per pixel For the first Local strain gradient at each pixel For the first The normal elastic modulus direction angle of each pixel For all The mean, For the first Shear modulus of each pixel For the first Local volume change of each pixel In order to be with the first The equivalent elastic transformation coefficients related to the volume of each pixel This represents the total number of valid pixels in the stone area. For pixel index.

[0017] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By unifying the acquisition and processing of multimodal medical image data and combining residual networks to extract high-dimensional features from kidney and ureter images, the ability to express details in grayscale histogram normalization and multi-scale edge detection is enhanced. The introduction of dense optical flow methods effectively improves the dynamic capture of registration errors. Through precise comparison of cosine similarity, spatial affine transformation is performed after obtaining the registration parameter matrix, enhancing the integrity of stone image features. The application of densely connected convolutional neural networks in topology extraction strengthens the fine identification of stone boundary points, blood vessels, and nerves. The constructed stone anatomical relationship map, combined with finite element analysis, enables quantitative prediction of shock wave energy attenuation. The combination of genetic algorithm-driven multi-target path search and Dijkstra algorithm-calculated path optimization improves the flexibility and adaptability of shock wave paths. Step angle adjustment further matches individual anatomical features, forming an adaptive mechanism for path planning. The overall processing logic achieves systematic improvements in multimodal image processing, registration accuracy, structure recognition, energy calculation, and path optimization, enhancing the accuracy of stone localization and the targeting of shock wave effects, and reducing the risk of accidental damage to healthy tissues. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the workflow of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0021] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0022] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0023] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0024] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0025] Please see Figure 1 This invention provides a technical solution: a method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis, comprising the following steps: S1: By acquiring multimodal medical image data from CT, MRI, and PET, feature extraction is performed on the medical images using a residual network. Gray-level histogram normalization is performed on kidney and ureter images, and edge detection is performed using the Laplacian operator to generate a multimodal image feature set. The method of using residual networks to extract features from medical images includes the following training process: CT, MRI, and PET multimodal medical images are uniformly processed to a standard size, and label maps of the kidney and ureter regions are constructed based on expert annotations as supervision signals. During the training phase, ResNet-34 is used as the base model, and ImageNet pre-trained weights are used for transfer learning initialization. The network output and the real labels are supervised and trained using the cross-entropy loss function or Dice Loss. The Adam optimizer is used, with an initial learning rate of 0.001, 100–200 training epochs, and a batch size of 8–16. Data augmentation strategies such as rotation, scaling, and contrast enhancement are used to improve the robustness of the network to different organ morphologies and grayscale changes. During training, the learning rate is dynamically adjusted based on the IoU and Dice coefficients of the validation set, and an EarlyStopping strategy is set to prevent overfitting. After training, the model is exported in ONNX format, input into medical images, and high-dimensional feature maps are extracted. S2: Based on the multimodal image feature set, the registration error is calculated using the dense optical flow method. Combined with cosine similarity comparison, the registration parameter matrix is ​​obtained, and spatial affine transformation is performed. The stone image is then processed using three-dimensional convolution operation to generate a stone image feature set. S3: Based on the stone image feature set, a densely connected convolutional neural network is called to extract the topology, obtain the stone boundary point set, blood vessel location point set, and nerve location point set, construct the stone anatomical relationship map, and use finite element analysis to calculate the energy attenuation value and generate a shock wave energy curve set. The method involves using a densely connected convolutional neural network (DCNN) to extract the topology. The training process of the DCNN includes: collecting and organizing a medical image dataset with stone annotation information. Data sources include stone edges, blood vessel locations, and nerve distribution areas marked in preoperative CT / MRI scans. Corresponding masked images are used as training supervision signals. The network structure uses the DenseNet-121 framework as its base. All convolutional layers achieve multi-layer feature reuse through feature map cascading, enhancing information flow and gradient propagation efficiency. During training, the input is a stone image block after registration and normalization, and the output is the three-dimensional coordinates and heatmap labels for each type of structural point, including stone edges, blood vessels, and nerves. The optimizer used is Adam with a learning rate of 0.0005, a batch size of 4, and a multi-channel Dice loss function. The combination of Loss and position error penalty term is used for 150 training rounds. Dropout and L2 regularization are introduced to avoid overfitting. After training, the stone image feature set is input, the model performs feature map mapping, normalization and ReLU activation, and outputs the three-dimensional spatial coordinate set of stone edge point set, blood vessel location point set and nerve location point set to obtain the topological structure coordinate matrix. S4: Based on the shock wave energy curve set, a genetic algorithm is used to perform multi-objective path search, the Dijkstra algorithm is called to calculate the optimal shock wave path and perform step angle adjustment to generate an optimized shock wave path set; S5: Based on the optimized shock wave path set, real-time acquisition of intraoperative ultrasound elastography monitoring images is performed. The image elastic modulus is detected, the relationship between the rate of change of elastic modulus and the output energy intensity is analyzed, and the degree of stone fragmentation is analyzed based on the image elastic modulus detection. The shock wave frequency and energy intensity are dynamically adjusted according to the degree of stone fragmentation, and the shock wave parameter adjustment set is output.

[0026] The multimodal image feature set includes uniform gray level mapping value, organ edge response value, and spatial structure expression; the stone image feature set includes location coordinate coefficient, tissue density range, and morphological feature encoding; the shock wave energy curve set includes attenuation trend, path conduction characteristics, and nerve location interference weight; the optimized shock wave path set includes angle offset range, path continuity index, and incident direction set; and the shock wave parameter adjustment set includes frequency control parameters, energy release level, and pulse action period.

[0027] Please see Figure 1 The specific steps of S1 are as follows: S101: By acquiring multimodal medical image data from CT, MRI, and PET, slice layering is performed on each set of image data using a uniform sampling interval. The image boundaries of the urinary structure region are identified and assigned numbered indices. The pixel matrix of the corresponding kidney and ureter images in the indexed images is extracted and mapped to a standard coordinate space to obtain a unified image index set. By acquiring multimodal medical image data from CT, MRI, and PET, the process begins by acquiring their DICOM data formats and decoding them into raw image matrices. For each type of image data, spatial coordinates are simultaneously sampled at a uniform sampling interval (e.g., 1.0 mm). Axial slicing is then used to decompose each group of three-dimensional images layer by layer at this sampling interval, generating two-dimensional slices. Subsequently, the edge contours of the kidneys and ureters in the urinary system are identified. In each slice, a Region of Interest (ROI) is constructed using edge detection combined with a region growing strategy. Each valid slice is then numbered to form an index identifier. All slices are mapped to a unified three-dimensional spatial coordinate system to form a standard index set. During the mapping process, the pixel position coordinates are translated, scaled, and rotated using affine transformation matrices. For example, for a 512×512 resolution image slice from a CT image... With a pixel pitch of 0.9 mm, if the kidney region has 1485 pixels, its shape can be reconstructed on a two-dimensional plane. By matching the spatial geometric distance between the geometric center and edge points, the coordinates of the pixel matrix can be transformed into position coordinates in standard three-dimensional space. Similarly, the same index number can be used for slices at the same location in MRI and PET, and the extracted ureter region pixel set can be uniformly processed to obtain a standard image index set. Each index set includes the slice number, the kidney pixel coordinate set, the ureter pixel coordinate set, and its registration error with the original image. If the error is controlled within 0.5 mm, it is considered qualified. Among them, the data set numbered 3 has a pixel pitch of 0.9 mm, 1485 kidney pixels, 412 ureter pixels, and a mapping error of 0.42 mm. This data set is included in the unified image index set.

[0028] A three-dimensional rectangular coordinate system is established in the standard coordinate space with the highest point of the iliac crest as the origin.

[0029] S102: Based on the unified image index set, the backbone residual block structure in the residual network is called to extract image region features. Combined with the pixel matrix, deformable convolution kernels are used for multimodal feature alignment, cross-channel normalization is performed, and features are updated through skip connections. Block-level channel statistics are used for feature value normalization analysis. Combined with the peak and mean values ​​in the same resolution group, feature contrast coefficients are generated to obtain the region feature contrast coefficient matrix. Based on a unified image index set, for each group of slice images, the corresponding image is first retrieved from the image database according to its number. Image region features are extracted through convolution operations. A 3×3 window slides across each pixel, recording pixel values ​​within the window region and centering them by mean. During deep feature extraction, a residual backbone module containing three convolutional layers is called level by level. The output of the l-th layer feature map is weighted and summed to form a residual mapping value, which is then added to the input feature map to complete the feedforward update. Simultaneously, during feature extraction, the pixel matrices of CT, MRI, and PET images under the same number are spatially aligned using a 3×3 deformable convolution kernel. The convolution sampling point position is dynamically adjusted according to the displacement field of the reference image. In practical applications… In a CT image, if a pixel at a certain location is (256, 320), and its displacement in the corresponding MRI image due to patient position is (258, 318), the sampling position is adjusted to the new coordinates to achieve registration. After registration, the modal feature maps undergo cross-channel normalization. The pixel value of each channel is subtracted from the channel mean and then divided by the standard deviation to achieve distribution uniformity. Subsequently, skip connections are performed to cascade and merge shallow and deep features. In the block-level statistical stage, the response values ​​of all channels within each feature block are statistically analyzed, and the difference between the maximum and average channel values ​​is calculated to measure the regional intensity difference. Assuming that the maximum channel response value of a certain feature block is 0.92 and the average is 0.61, the generated feature contrast coefficient is... By analogy, the feature contrast coefficients of each image region are calculated and combined into a region feature contrast coefficient matrix.

[0030] S103: Calculate the normalized difference rate between the regional feature contrast coefficient matrix and the standard grayscale mapping histogram, and use the difference rate to call the Laplacian operator to perform edge detection and obtain a multimodal image feature set; Based on the regional feature contrast coefficient matrix, the normalized difference rate is calculated for each coefficient value and the gray value distribution curve of the corresponding standard gray-level mapping histogram. The calculation process is as follows: first, the feature coefficients are normalized to the interval of 0 to 1, and then the difference rate is calculated based on the frequency probability value of their corresponding positions in the gray-level histogram. The formula for calculating the difference rate is: ,in For the normalized value of regional characteristics, This represents the probability density value of the region in the grayscale histogram, such as when , At that time, Then, the difference rate is used as a weighting coefficient for the image edge response to construct an edge Laplacian response map. The location with the largest difference in the three channels is taken as the center point, and a Laplacian template is applied to it for convolution. For example, if the center point pixel value is 142, and its surrounding pixels are [140, 143, 144, 139, 142, 141, 138, 145], the Laplacian response is calculated as follows. The response value is superimposed with the difference rate to obtain the multimodal fusion edge feature map, which is then used to construct the overall multimodal image feature set as the input for subsequent 3D reconstruction and classification.

[0031] Please see Figure 1 The specific steps of S2 are as follows: S201: The dense optical flow method is used to extract the gray-level change vector field of the image feature map in the multimodal image feature set between adjacent pixels, and the pixel position displacement is calculated on the vector field. The pixel displacement is converted into the actual displacement value in millimeters, a global displacement vector matrix is ​​generated, the modal image registration error value is calculated, and the registration error value matrix is ​​obtained. The dense optical flow method is used to extract the gray-level change vector of adjacent pixels in the feature set of multimodal images. First, based on CT and MRI images of the same location, the gray-level difference and gradient direction of each pixel in the time frame or sequence of images are calculated. The displacement estimate of each pixel in the x and y directions is recorded. For example, if the gray level of pixel (120, 200) in an image is 142, and the gray level of the same location in a neighboring image is 145, then the gray-level change is +3. The movement direction and magnitude of this pixel are estimated using the gray-level difference and gradient direction (Δx = 2.4 pixels, Δy = 3.1 pixels). Combined with the pixel spacing parameter of the image, such as 0.9 mm, the actual displacement is... Repeat the above operations to construct a complete vector field for all pixels in the image, generating a global displacement vector matrix. Then, compare this matrix with the modal registration reference model, calculate the displacement error for each matching point, and calculate the absolute value of the difference between the actual displacement and the reference position. Assume that the reference position error range is controlled within 0.5mm. When the actual displacement of a pixel in a certain area is 4.68mm and the standard value is 4.35mm, the error is 0.33mm. Summarize each error point into a registration error value matrix, and finally complete the goal of this step. The specific data is shown in Table 1. Table 1. Registration Error Calculation Table 1 45 5.2 0.90 4.680 0.32 2 38 4.8 1.00 4.800 0.41 3 52 6.1 0.85 5.185 0.38 4 47 5.7 0.95 5.415 0.35 As shown in Table 1, the error values ​​are all controlled below 0.5 mm, which meets the requirements for generating the registration error matrix.

[0032] S202: Based on the registration error value matrix, compare the cosine similarity of each group of registration error value vectors with the standard reference vector, calculate the cosine value of the angle between the registration errors, and generate a matching parameter sequence by matching and combining the cosine values ​​to establish the registration parameter matrix; Based on the aforementioned registration error value matrix, the components of each vector pair are extracted item by item and compared with the directionality of the standard reference vector. Let the standard vector be... Each registration vector is First, calculate the magnitude of each of the two vectors. , Then calculate the inner product. Finally, the cosine value is obtained. If the standard vector is (3,4,0) and the first registration vector is (2.7,4.2,0), then... , The inner product is The cosine value is When the cosine value falls above 0.95, it is judged as a high similarity match. All matching results are arranged in descending order of cosine value to form a matching parameter sequence. Each record in the sequence corresponds to a number, matching cosine value and its direction vector index. Finally, a registration parameter matrix is ​​generated, with each row recording a matching relationship. The smaller the angle (the closer the cosine value is to 1), the better the matching quality.

[0033] S203: Based on the registration parameter matrix, perform spatial affine transformation to perform coordinate mapping and registration of modal images, construct the registration parameter matrix and dynamically adjust the channel attention weights of the three-dimensional convolution kernel, use anisotropic convolution kernels to extract spatial features of each axis, combine convolution kernel weights to generate a feature space mapping matrix, and obtain the stone image feature set. Based on the above registration parameter matrix, the registration vector group corresponding to each modal image is selected sequentially, and the six parameters of the affine transformation matrix are calculated. Let the input coordinates be (x, y), and the affine transformation result be (x...y ... y ), executed in the following mapping form: , , where the coefficient Based on the cosine value weighting settings in the registration parameter matrix, the closer the cosine value is to 1, the higher the corresponding parameter weight. For example, when the cosine value is 0.997, its weight is set to 0.98. This weight is then substituted into the affine matrix parameters for mapping. After coordinate system alignment, attention weights are assigned to the channels of the 3D convolution kernel. The convolution response along each axis is adjusted according to directional characteristics. For instance, if the structural variation rate of the registered image is detected to be higher than 0.35 in the x-axis direction, the weight of that channel is reduced from the default 1.0 to 0.85 to control... In the spatial convolution process, anisotropic convolution kernels are used to determine the convolution response intensity. This involves setting different kernel sizes and strides in the x, y, and z axes. Specifically, kernel sizes of 3×1×1 in the x-axis, 1×3×1 in the y-axis, and 1×1×3 in the z-axis are used to extract spatial structural features in the horizontal, vertical, and layer directions, respectively. Finally, a spatial mapping matrix that fuses the three-axis features is generated by combining the convolution kernel weights. The image feature set processed by this matrix is ​​the candidate feature set for the stone region, which serves as the input for the next stage of classification.

[0034] Please see Figure 1 The specific steps of S3 are as follows: S301: Based on the feature set of stone images, a densely connected convolutional neural network is invoked to perform feature mapping, normalization and activation operations in sequence, and multi-dimensional channel feature enhancement processing is performed on the convolutional feature map to extract edge points, blood vessel location points and branch points in the topology, and the extraction results are geometrically calibrated to obtain the topology coordinate matrix. Based on a densely connected convolutional neural network (CNN) inputting a feature set from gallstone images, the 3D feature input is first mapped to a multi-channel feature map through an initial convolutional layer. Each convolutional kernel is set to 3×3×3 with a stride of 1 and a padding of 1. Then, the dense block module's internal structure is entered. Each convolutional layer has 32 output channels, and each dense block contains 4 convolutional layers. The output channels of the previous layer are concatenated with the input feature map of the current layer to form a fused input before entering the next layer. If the first layer has 64 input channels, the second layer has 64 + 32 = 96 input channels, and so on, resulting in a feature map stacking and channel growth. During execution, batch normalization is performed on the output of each layer. The normalization process involves subtracting the channel mean from the pixel value within each channel and then dividing by the standard deviation. For example, if the pixel mean in a channel is 128, the standard deviation is 20, and the current pixel value is 150, then the normalized value is... The normalized value is then processed by the ReLU activation function and output to the next layer. In the entire network structure, the edge response and texture distribution at different scales and directions are continuously strengthened. Finally, a 64-channel multidimensional feature map is generated at the output. When extracting edge points, blood vessel points and branch points, the maximum response channel index method is used to select the point with the maximum gradient change and the coordinate point is adjusted by combining the position offset field. Let the point with the maximum response in the image be (128,240,25) in edge point extraction. Then this point is marked as an edge point. Similarly, response analysis and offset fine-tuning are performed on blood vessel points and branch points. The coordinate points (135,250,25) and (136,248,25) are recorded in blood vessel points. Finally, all extracted coordinate points are organized into triplet (x,y,z) to construct the topological structure coordinate matrix. Some results are shown in Table 2. Table 2 Topological Structure Coordinate Table 1 (128,240,25) (135,250,25) (132,245,25) 2 (130,243,25) (136,248,25) (129,247,25) 3 (126,239,25) (134,251,25) (131,246,25) As shown in Table 2, each group of numbers corresponds to a set of three-dimensional coordinates of three types of structural points.

[0035] S302: Based on the topological coordinate matrix, Euclidean distance is calculated for the boundary points, blood vessel locations, and nerve locations of the stones. Combined with the spatial topological relationships between each set of points, a stone anatomical relationship atlas is constructed. Based on the coordinates of edge points, blood vessel points, and nerve points extracted from the topological coordinate matrix, point sets A, B, and C are constructed respectively. Euclidean distance is calculated for any two points in each set. Taking point pair P1=(x1,y1,z1) and P2=(x2,y2,z2) as an example, the Euclidean distance calculation formula is as follows: If the edge point is (128, 240, 25) and the blood vessel point is (135, 250, 25), then the distance between the two points is... This calculation is performed on all combinations of edge points with blood vessel points, edge points with nerve points, and blood vessel points with nerve points. The results form a distance matrix. Then, it is determined whether each pair of structural points constitutes a valid topological relationship. The criteria are: if the distance difference between the three points is less than 10%, it is considered a collinear structure; if the ratio of the distance from a point to two points falls between 0.9 and 1.1, it is defined as an equidistant distribution structure. These rules are used to construct the connection relationship between each node and its neighboring structures in the spatial topology graph. For example, if the distances between an edge point and two blood vessel points are 12.21 mm and 11.95 mm respectively, then the distance ratio is... If the stones fall within the above-mentioned range, they are determined to form equidistant connections between the stone nodes and blood vessel branches. Finally, the connection relationships and spatial coordinates between all nodes are sorted out to construct a stone anatomical relationship atlas.

[0036] S303: Based on the anatomical relationship atlas of stones, the Young's modulus and Poisson's ratio of stone tissue are obtained through preoperative CT value markers. Finite element analysis is used to set the incident angle and energy parameters, calculate the energy attenuation value, and generate a set of shock wave energy curves. The specific formula for calculating the energy attenuation value, given the incident angle and energy parameters, is as follows: ; Calculate the energy decay value; in, This represents the attenuation value of the shock wave energy per unit volume. Representing the The angle between the incident direction of the shock wave and the normal. Representing the The initial energy of the incident shock wave, Representing the The propagation depth of the path, Representing the stone tissue in the Equivalent absorption coefficient under the path, Representing the The relative propagation volume of the stone in a local area along the path, Represents the total energy of the reference shock wave. This represents the index of the number of paths in a multi-path propagation scenario. This is a discrete index variable used for path numbering in the multipath propagation model of shock waves; formula: ; Detailed explanation of the formula and its calculation derivation: The formula is used to calculate the attenuation value of shock wave energy per unit volume, and the result is used to generate a set of shock wave energy curves. Parameter meanings and settings: For the first The angle factor between the incident direction of the shock wave and the normal direction of the tissue surface in the path is set to . ; For the first The initial input energy of the shock wave in the path is set to... ; For the first The actual linear distance the shock wave travels within the tissue along the path is set as... ; For the first The equivalent absorption coefficient corresponding to the stone tissue in the path is set as follows: ; For the first The relative volume of the local stone tissue participating in the shock wave energy transfer along the path is set as follows: ; The total energy of the baseline shock wave used for comparison calculations is set as follows: ; Let the total number of paths of the shock wave under multipath propagation conditions be defined as . ; Substitute the parameters into the formula to calculate: ; ; Calculate the ratio: ; Calculate the difference between the energy and the reference energy: ; result This indicates the attenuation value of shock wave energy per unit volume. This result is used to generate a set of shock wave energy curves, reflecting the energy attenuation of shock waves during propagation in stone tissue.

[0037] In the finite element analysis, the ureteral wall is set as a fixed constraint boundary, the contact area between the stone and soft tissue is set as a free boundary, a shock wave excitation is applied to the incident surface, and the energy decay process is simulated by combining tissue damping parameters and material properties.

[0038] Please see Figure 1 The specific steps of S4 are as follows: S401: Based on the shock wave energy curve set, using energy attenuation rate, incident angle and tissue impedance parameters, a genetic algorithm is invoked to perform multi-target path search, screen path gene sequences, and obtain a set of path gene sequences. Based on the energy parameters of each energy propagation path in the shock wave energy curve set, the initial input energy of each path was first set to 100J. The corresponding energy attenuation rates were measured to be 15% for path 1, 18% for path 2, 12% for path 3, and 20% for path 4. Therefore, their terminal energies were 85J, 82J, 88J, and 80J, respectively. Subsequently, the incident angles formed by the shock wave and the tissue interface in the corresponding paths were collected. The actual measured angles were 45° for path 1, 60° for path 2, 30° for path 3, and 75° for path 4. These angle values ​​were converted into a path propagation coefficient influence factor, denoted as [factor name missing]. Among them, the maximum value of path 3 is The minimum value for path 4 is Then, tissue impedance Z is introduced as a parameter for the reflection and absorption of energy by the medium, and is set sequentially as follows: path 1: 1.5, path 2: 2.0, path 3: 1.2, path 4: 2.3 (unit: Based on the above three participating factors: attenuation rate, cosine of incident angle, and tissue impedance, a path gene factor group is formed. After randomly initializing the population, the fitness function value of each path combination is calculated, and its total remaining energy and propagation efficiency comprehensive index are evaluated. Individuals with fitness higher than 0.75 are retained to enter the next round of screening. It is assumed that path 3 performs best under the three parameters, with the lowest attenuation, the best incident angle, and the lowest impedance, corresponding to a fitness value of 0.89, forming gene sequence P3. The path number and parameters after screening are shown in Table 3. Table 3. Pathway Gene Screening Parameters 1 100 15 45 1.5 0.82 2 100 18 60 2.0 0.76 3 100 12 30 1.2 0.89 4 100 20 75 2.3 0.68 As shown in Table 3, Pathway 3 was selected into the final path gene sequence set because of its highest genetic fitness value.

[0039] S402: Based on the path gene sequence set, a weighted directed graph model is established by calling the Dijkstra algorithm, setting node weights, calculating the cost value of the path, screening the path region where the cost value is between the energy consumption and tissue penetration equilibrium threshold, and obtaining the shock wave path cost value matrix. A directed graph model is constructed based on the path gene sequence set. Each node corresponding to a path is treated as an edge in the graph. A weight is assigned to each edge, which is the product of the energy consumption per unit length of the path and the tissue impedance. For example, if the energy consumption per unit length of path 1 is 15 J and the tissue impedance is 1.5, then its path weight is... Similarly, path 2 is calculated to be 36, path 3 to be 14.4, and path 4 to be 46. All weight values ​​are assigned to the corresponding edges in the graph. When performing path cost calculation, the path graph is traversed in depth first, and the total weight values ​​from the starting point to the end point are accumulated respectively. The starting point is the source emission point and the end point is the stone target point, forming a path cost set. Then, all path costs are screened. The screening rule is that the cost must be between the "energy consumption threshold (25)" and the "penetration impedance threshold (40)". The corresponding paths that meet the conditions are path 1 and path 3. Finally, the shock wave path cost matrix is ​​obtained, and its values ​​are path 1: 22.5 and path 3: 14.4 respectively. The two are candidate sets of paths to be optimized. The node weights are set as the sum of the shock wave energy consumption coefficient and the tissue penetration depth coefficient; The tissue penetration equalization threshold is set based on tissue impedance and thickness distribution to limit the penetration depth of shock waves and ensure tissue safety.

[0040] S403: Based on the shock wave path cost matrix, the incident angle step is adjusted on the selected path, the average angle change of the path coordinate point set is extracted, the step threshold is set in combination with the tissue penetration direction gradient, the incident direction sequence is obtained, the adjusted incident direction sequence is integrated with the path coordinate point set, and the optimized shock wave path set is output. After obtaining the cost matrix, the propagation direction of the screening path is fine-tuned. First, the node coordinates between each path segment are set. For example, path 3 consists of coordinate points P0 to P5, with corresponding coordinate points (0,0,0), (1,1,0.3), (2,1.7,0.8), (3,2.1,1.3), (4,2.2,2.1), and (5,2.4,2.9). The direction vectors between adjacent points are extracted sequentially, and then the included angle is calculated. The angle change is recorded in degrees. The average angle change is calculated as the ratio of the angle of deflection to the gradient angle difference of the tissue penetration direction. For example, if the included angle of the i-th segment is 28°... If the included angle of segment i+1 is 31°, then the change is 3°. If the average change value in five consecutive segments exceeds 10°, then the direction step size is adjusted to the original step size × 0.85 to reduce the direction jump. Through the mapping of direction gradient and spatial distribution, the adjusted direction set is recorded as a new incident direction sequence. The new direction sequence corresponding to P0 to P5 is θ1=45°, θ2=42°, θ3=39°, θ4=35°, θ5=32°. Each direction is combined with the corresponding coordinate to form a path control pair, and finally an optimized shock wave path set is generated. This path set has a stable direction sequence and minimum cost attribute, which is used as the final treatment path output. The step threshold is set based on the gradient of stone density distribution in the preoperative CT images.

[0041] Please see Figure 1 The specific steps of S5 are as follows: S501: Based on the optimized shock wave path set, the intraoperative ultrasound elastography monitoring image is acquired in real time. The elastic modulus of the stone area in the monitoring image is detected. By detecting the rebound coefficient and strain gradient of the pixel area, the overall elastic modulus distribution curve of the stone is calculated, and the elastic modulus distribution value of the stone is obtained. The specific formula for calculating the overall elastic modulus distribution curve of the stone by detecting the rebound coefficient and strain gradient of the pixel region is as follows: ; Calculate the elastic modulus distribution of the gallstone; in, This represents the overall elastic modulus distribution value of the stone region. For the first The rebound coefficient per pixel For the first Local strain gradient at each pixel For the first The normal elastic modulus direction angle of each pixel For all The mean, For the first Shear modulus of each pixel For the first Local volume change of each pixel In order to be with the first The equivalent elastic transformation coefficients related to the volume of each pixel This represents the total number of valid pixels in the stone area. For pixel index; formula: ; Detailed explanation of the formula and its calculation derivation: The formula is used to calculate the overall elastic modulus distribution value of the stone region, and the result is used to reflect the mechanical equilibrium state and shock wave response characteristics of the stone region under ultrasound elastography. Parameter meanings and settings: ; Dimensionless; ; ; ; ; ; ; Substitute the parameters into the formula to calculate: ; result This indicates the overall elastic modulus characteristic value of the multi-point mechanical response within the stone region under directional equilibrium conditions, reflecting the comprehensive elastic trend of the structural region in the monitoring image. This value can be used as a benchmark elastic reference value in the setting of shock wave energy control strategy, and a distribution curve trend map is formed by repeated calculation at different time periods for dynamic evaluation of the fragmentation state evolution. The elasticity coefficient was measured using the optical flow method. By performing inter-frame cross-correlation calculations on radio frequency signal frames before and after stress, the displacement of local tissue under longitudinal compression was extracted, and its elastic recovery characteristics were further evaluated.

[0042] S502: Based on the distribution value of the elastic modulus of the stone, gradient change analysis is performed on the mean elastic modulus of multiple regions to identify abnormal areas of stone morphology. By analyzing the relationship between the rate of change of elastic modulus and the output energy intensity, and according to the amplitude and range of the elastic change interval, the degree of stone fragmentation is evaluated, and stone fragmentation state parameters are generated. Based on the distribution of the elastic modulus of gallstones, multiple regions were first scanned and imaged, and the mean and standard deviation of the elastic modulus of each region were extracted. Regions 1 to 4 were designated as representing four different morphological regions, with mean elastic moduli of 20.5 MPa, 18.0 MPa, 12.5 MPa, and 7.2 MPa, respectively. The modulus variation between adjacent regions was analyzed using a sequential difference method. The rate of change was calculated as the difference between the current region's value and the previous region's value, divided by the previous region's value. For example, the rate of change between region 2 and region 1 was... The actual calculation took an absolute value of 12%. Then, combined with the output energy intensity recorded during shock wave processing, the energy values ​​of regions 1 to 4 were set to 80J, 75J, 65J, and 55J, respectively. The elastic change rate of each region was compared with the energy intensity. The threshold for abnormal morphological feature coding was set as energy intensity reduction greater than 20% and modulus reduction greater than 35%. Regions 3 and 4 both met the judgment conditions, and these two regions were assigned the high fracture sensitivity attribute label. Then, all regions were stratified according to the elastic modulus reduction range. Regions with modulus <15MPa and change rate >40% were judged as fracture level III and above. For example, region 3 has a modulus of 12.5MPa and a change rate of 44%, and is classified into category III. Region 4 has a modulus of 7.2MPa and a change rate of 63%, and is classified into category IV. Finally, the region numbers and parameters are shown in Table 4. The changes of each region were integrated to generate a set of stone fracture state parameters. Table 4 Elastic modulus and fragmentation state of gallstones 1 20.5 12 80 Ⅰ 2 18.0 30 75 Ⅱ 3 12.5 44 65 Ⅲ 4 7.2 63 55 Ⅳ As shown in Table 4, the degree of fragmentation increases sequentially from I to IV, indicating an increasing tendency for fragmentation.

[0043] S503: Based on the stone fragmentation state parameters, according to the current degree of stone fragmentation, the voltage parameters of the shock wave transmitter are adjusted in real time by using a PID controller to dynamically adjust the frequency and energy intensity of the shock wave and output the shock wave parameter adjustment set. Based on the set of fragmentation state parameters, the voltage of the shock wave emitter is adjusted according to the current fragmentation level of the region to control the emission frequency and intensity. The initial voltage is set to 2.5kV, corresponding to a frequency of 0.5Hz. If the current fragmentation level is I, the original voltage output is maintained. If the level is II, the voltage is reduced by 10%, adjusted to 2.25kV, and the frequency is increased to 0.6Hz. If the level is III, the voltage is reduced to 2.0kV, and the frequency is increased to 0.7Hz. If the level is IV, the voltage is adjusted to 1.8kV, and the frequency is controlled at 0.8Hz. During execution, a proportional-integral-derivative (PID) controller is used to calculate in real time the difference between the target fragmentation threshold and the current fragmentation state, denoted as error e(t). For example, in region 4, the current fragmentation level is IV, corresponding to a modulus of 7.2MPa. The difference between this modulus and the preset termination fragmentation state modulus of 6.5MPa is 0.7MPa. The proportional-integral-derivative (PID) controller is used to calculate the error e(t). With gain coefficient Kp=1.2, integral gain Ki=0.05, and differential gain Kd=0.01, the instantaneous adjustment u(t)=1.2×e(t)+0.05×∫e(t)dt+0.01×de(t) / dt. If e(t) is 0.7, 0.5, and 0.3 in three consecutive measurements, the integral term is 1.5, and the derivative is -0.2. Therefore, u(t)=1.2×0.3+0.05×1.5+0.01×(-0.2)=0.36+0.075-0.002=0.433. This value is used to reduce voltage and increase frequency. Assuming the maximum allowable variation is ±0.5kV and ±0.3Hz, the current adjustment should be 2.067kV and 0.833Hz. This voltage and frequency value is combined to form the shock wave parameter adjustment set, and the controller reference value is updated to achieve continuous dynamic adjustment.

[0044] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0045] It should be understood that the term "and / or" in this article merely describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it may also indicate an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0046] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0047] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0048] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0049] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0050] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0051] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0052] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0053] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0054] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for locating urinary stones and optimizing fragmentation paths based on AI ultrasound image analysis, characterized in that, The method includes: S1: By acquiring multimodal medical image data from CT, MRI, and PET, feature extraction is performed on the medical images using a residual network. Gray-level histogram normalization is performed on kidney and ureter images, and edge detection is performed using the Laplacian operator to generate a multimodal image feature set. The method of using residual networks to extract features from medical images includes the following training process: CT, MRI, and PET multimodal medical images are uniformly processed to a standard size, and label maps of the kidney and ureter regions are constructed based on expert annotations as supervision signals. During the training phase, ResNet-34 is used as the base model, and ImageNet pre-trained weights are used for transfer learning initialization. The network output and the real labels are supervised and trained using the cross-entropy loss function or Dice Loss. The Adam optimizer is used, with an initial learning rate of 0.001, 100–200 training epochs, and a batch size of 8–16. Data augmentation strategies such as rotation, scaling, and contrast enhancement are used to improve the robustness of the network to different organ morphologies and grayscale changes. During training, the learning rate is dynamically adjusted based on the IoU and Dice coefficients of the validation set, and an Early Stopping strategy is set to prevent overfitting. After training, the model is exported in ONNX format, input into medical images, and high-dimensional feature maps are extracted. S2: Based on the multimodal image feature set, the registration error is calculated using the dense optical flow method, and the registration parameter matrix is ​​obtained by combining cosine similarity comparison. Spatial affine transformation is then performed, and the stone image is processed using three-dimensional convolution operation to generate a stone image feature set. S3: Based on the stone image feature set, a densely connected convolutional neural network is called to extract the topology, obtain the stone boundary point set, blood vessel location point set, and nerve location point set, construct a stone anatomical relationship map, and use finite element analysis to calculate the energy attenuation value to generate a shock wave energy curve set. The method involves using a densely connected convolutional neural network (DCNN) to extract the topology. The training process of the DCNN includes: collecting and organizing a medical image dataset with stone annotation information. Data sources include stone edges, blood vessel locations, and nerve distribution areas marked in preoperative CT / MRI scans. Corresponding masked images are used as training supervision signals. The network structure uses the DenseNet-121 framework as its base. All convolutional layers achieve multi-layer feature reuse through feature map cascading, enhancing information flow and gradient propagation efficiency. During training, the input is a stone image block after registration and normalization, and the output is the three-dimensional coordinates and heatmap labels for each type of structural point, including stone edges, blood vessels, and nerves. The optimizer used is Adam with a learning rate of 0.0005, a batch size of 4, and a multi-channel Dice loss function. The combination of Loss and position error penalty term is used for 150 training rounds. Dropout and L2 regularization are introduced to avoid overfitting. After training, the stone image feature set is input, the model performs feature map mapping, normalization and ReLU activation, and outputs the three-dimensional spatial coordinate set of stone edge point set, blood vessel location point set and nerve location point set to obtain the topological structure coordinate matrix. S4: Based on the shock wave energy curve set, a genetic algorithm is used to perform multi-objective path search, the Dijkstra algorithm is called to calculate the optimal shock wave path and perform step angle adjustment to generate an optimized shock wave path set.

2. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 1, characterized in that, The multimodal image feature set includes a unified gray level mapping value, organ edge response value, and spatial structure expression; the stone image feature set includes location coordinate coefficients, tissue density range, and morphological feature encoding; the shock wave energy curve set includes attenuation trend, path conduction characteristics, and nerve location interference weight; and the optimized shock wave path set includes angle offset range, path continuity index, and incident direction set.

3. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: By acquiring multimodal medical image data from CT, MRI, and PET, slice layering is performed on each set of image data using a uniform sampling interval. The image boundaries of the urinary structure region are identified and assigned numbered indices. The pixel matrix of the corresponding kidney and ureter images in the indexed images is extracted and mapped to a standard coordinate space to obtain a unified image index set. The standard coordinate space establishes a three-dimensional rectangular coordinate system with the highest point of the iliac crest as the origin; S102: Based on the unified image index set, the backbone residual block structure in the residual network is called to extract image region features. Combined with the pixel matrix, deformable convolution kernels are used for multimodal feature alignment, cross-channel normalization is performed, and features are updated through skip connections. Block-level channel statistics are used for feature value normalization analysis. Combined with the peak and mean values ​​in the same resolution group, feature contrast coefficients are generated to obtain the region feature contrast coefficient matrix. S103: Calculate the normalized difference rate between the region feature contrast coefficient matrix and the standard grayscale mapping histogram, and use the difference rate to call the Laplacian operator to perform edge detection and obtain a multimodal image feature set.

4. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Extract the grayscale change vector field between adjacent pixels of the image feature map in the multimodal image feature set using the dense optical flow method, perform pixel position displacement calculation on the vector field, convert the pixel displacement amount into the actual displacement value in millimeters, generate a global displacement vector matrix, calculate the modal image registration error value, and obtain the registration error value matrix. S202: Based on the registration error value matrix, compare the cosine similarity of each group of registration error value vectors with the standard reference vector, calculate the cosine value of the angle between the registration errors, and generate a matching parameter sequence by matching and combining the cosine values ​​to establish a registration parameter matrix; The standard reference vector is selected based on the characteristics of the intraoperative imaging modality, and the standard reference vector adopts the registration error vector corresponding to the intraoperative real-time ultrasound image with high real-time performance and anatomical accuracy as the benchmark reference modality. S203: Based on the registration parameter matrix, perform spatial affine transformation to perform coordinate mapping and registration of modal images, construct the registration parameter matrix and dynamically adjust the channel attention weights of the three-dimensional convolution kernel, use anisotropic convolution kernels to extract spatial features of each axis, combine the convolution kernel weights to generate a feature space mapping matrix, and obtain the stone image feature set.

5. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Based on the stone image feature set, a densely connected convolutional neural network is invoked to sequentially perform feature mapping, normalization and activation operations, perform multi-dimensional channel feature enhancement processing on the convolutional feature map, extract edge points, blood vessel location points and branch points in the topology, and perform geometric coordinate calibration on the extraction results to obtain the topology coordinate matrix. S302: Based on the topological coordinate matrix, perform Euclidean distance calculation on the stone boundary points, blood vessel location points and nerve location points, and construct a stone anatomical relationship atlas by combining the spatial topological relationships between each point set; S303: Based on the anatomical relationship atlas of the stone, the Young's modulus and Poisson's ratio of the stone tissue are obtained by preoperative CT value markers. Using finite element analysis, the incident angle and energy parameters are set, the energy attenuation value is calculated, and a set of shock wave energy curves is generated. In the finite element analysis, the ureteral wall is set as a fixed constraint boundary, the contact area between the stone and soft tissue is set as a free boundary, a shock wave excitation is applied to the incident surface, and the energy decay process is simulated by combining tissue damping parameters and material properties.

6. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 5, characterized in that, The specific formula for setting the incident angle and energy parameters and calculating the energy attenuation value is as follows: ； Calculate the energy decay value; in, This represents the attenuation value of the shock wave energy per unit volume. Representing the The angle between the incident direction of the shock wave and the normal. Representing the The initial energy of an incident shock wave, Representing the The propagation depth of the path, Representing the stone tissue in the Equivalent absorption coefficient under the path, Representing the The relative propagation volume of the stone in a local area along the path, Represents the total energy of the reference shock wave. This represents the index of the number of paths in a multi-path propagation scenario. This is a discrete index variable used for path numbering in the multipath propagation model of shock waves.

7. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 5, characterized in that, The specific steps for S4 are as follows: S401: Based on the shock wave energy curve set, using the energy attenuation rate, incident angle and tissue impedance parameters, call the genetic algorithm to perform multi-target path search, screen path gene sequences, and obtain a set of path gene sequences; S402: Based on the set of path gene sequences, the Dijkstra algorithm is called to establish a weighted directed graph model, the node weights are set, the cost value of the path is calculated, the path region with the cost value between the energy consumption and tissue penetration equilibrium threshold is selected, and the shock wave path cost value matrix is ​​obtained. The set node weight is the sum of the shock wave energy consumption coefficient and the tissue penetration depth coefficient; The tissue penetration equalization threshold is set according to tissue impedance and thickness distribution, and is used to limit the penetration depth of shock waves to ensure tissue safety. S403: Based on the shock wave path cost matrix, the incident angle step is adjusted on the selected path, the average angle change of the path coordinate point set is extracted, the step threshold is set in combination with the tissue penetration direction gradient, the incident direction sequence is obtained, the adjusted incident direction sequence and the path coordinate point set are integrated, and the optimized shock wave path set is output. The step threshold is set based on the gradient of stone density distribution in the preoperative CT images.

8. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 1, characterized in that, The method further includes: S5: Based on the optimized shock wave path set, real-time acquisition of intraoperative ultrasound elastography monitoring images, image elastic modulus detection of the images, analysis of the relationship between elastic modulus change rate and output energy intensity, analysis of stone fragmentation degree based on image elastic modulus detection, dynamic adjustment of shock wave frequency and energy intensity based on stone fragmentation degree, and output of shock wave parameter adjustment set. The shock wave parameter adjustment set includes frequency control parameters, energy release level, and pulse period.

9. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 8, characterized in that, The specific steps of S5 are as follows: S501: Based on the optimized shock wave path set, intraoperative ultrasound elastography monitoring images are acquired in real time, and the elastic modulus of the stone region in the monitoring images is detected. By detecting the rebound coefficient and strain gradient of the pixel region, the overall elastic modulus distribution curve of the stone is calculated, and the elastic modulus distribution value of the stone is obtained. The rebound coefficient is measured using the optical flow method. By performing inter-frame cross-correlation calculations on the radio frequency signal frames before and after the force is applied, the displacement of the local tissue under longitudinal compression is extracted, and its elastic recovery characteristics are further evaluated. S502: Based on the elastic modulus distribution value of the stone, perform gradient change analysis on the average elastic modulus of multiple regions, identify abnormal areas of stone morphology, and evaluate the degree of stone fragmentation by analyzing the relationship between the rate of change of elastic modulus and the output energy intensity and according to the amplitude and range of the elastic change interval, and generate stone fragmentation state parameters. S503: Based on the stone fragmentation state parameters, according to the current degree of stone fragmentation, the voltage parameters of the shock wave transmitter are adjusted in real time by using a PID controller to dynamically adjust the frequency and energy intensity of the shock wave and output the shock wave parameter adjustment set.

10. The method for locating and optimizing the fragmentation path of urinary stones based on AI ultrasound image analysis according to claim 9, characterized in that, The specific formula for calculating the overall elastic modulus distribution curve of the stone by detecting the rebound coefficient and strain gradient of the pixel region is as follows: ； Calculate the elastic modulus distribution of the gallstone; in, This represents the overall elastic modulus distribution value of the stone region. For the first The rebound coefficient per pixel For the first Local strain gradient at each pixel For the first Normal elastic modulus direction angle of each pixel For all The mean, For the first Shear modulus of each pixel For the first Local volume change of each pixel In order to be with the first The equivalent elastic transformation coefficients related to the volume of each pixel This represents the total number of valid pixels in the stone area. For pixel index.

Images