Interventional therapy target accurate positioning method and system based on ultrasonic image

Abstract

The invention relates to an interventional therapy target accurate positioning method and system based on an ultrasonic image. According to the method, a binary mask is generated by extracting an acoustic shadow area in an ultrasonic image, contrast enhancement parameters of different areas are dynamically adjusted based on the binary mask, and target spot edge details are enhanced while shadow area noise is suppressed; a shadow mask attention mechanism is combined to guide a neural network to accurately segment a target point probability graph, a segmentation boundary is evolved and optimized by integrating a level set constrained by a shadow region shape, and local smooth compensation is performed on registration deviation of a three-dimensional reconstruction model and a preoperative image by using a shadow perception deformation field correction strategy; and finally generating a puncture path for avoiding the dangerous structure. The problems of fuzzy target boundary, segmentation error accumulation and deformation compensation misalignment of a traditional method under acoustic shadow interference are effectively solved, and the pertinence of image enhancement, the accuracy of target positioning and the safety of an intervention path are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of ultrasound image processing technology, and in particular relates to a method and system for precise localization of interventional treatment targets based on ultrasound images. Background Technology

[0002] In ultrasound-guided interventional therapy, precise target localization is crucial for surgical success. However, in practice, the presence of acoustic shadows often limits image quality. Hard tissues such as bone and calcifications reflect ultrasound waves, creating shadowed areas in the images, blurring the target boundary and significantly reducing contrast.

[0003] Traditional image enhancement methods, such as histogram equalization, can improve overall contrast, but they tend to over-amplify noise in shadow areas, which can interfere with subsequent segmentation and localization accuracy. Existing segmentation algorithms often rely on single-modality images and do not fully consider the special characteristics of shadow areas, leading to the accumulation of boundary errors during target 3D reconstruction and affecting the accuracy of puncture path planning.

[0004] In addition, the registration between intraoperative ultrasound and preoperative high-resolution images often deviates due to dynamic tissue deformation. Existing technologies mostly adopt a global smoothing strategy to correct the deformation field, which makes it difficult to adaptively adjust to the local characteristics of the shadow area, easily leading to instrument navigation errors.

[0005] To address the aforementioned issues, there is an urgent need for a positioning method that can integrate multi-dimensional information, dynamically suppress shadow interference, and achieve precise deformation compensation in order to improve the safety and success rate of interventional treatment. Summary of the Invention

[0006] Therefore, it is necessary to provide a method and system for precise localization of interventional treatment targets based on ultrasound imaging to address the aforementioned technical problems.

[0007] In a first aspect, this application provides a method for precise localization of interventional therapeutic targets based on ultrasound imaging, including:

[0008] S1. Perform high-frequency feature extraction and morphological processing on the input ultrasound image to generate a binary mask for the acoustic shadow region.

[0009] S2. Based on the binary mask of the acoustic shadow area, dynamically adjust the contrast enhancement parameters, and perform regional enhancement on the ultrasound image according to the contrast enhancement parameters to obtain a regionally enhanced image.

[0010] S3. Input the enhanced image and the binary mask of the acoustic shadow region into the pre-trained neural network model to generate a target probability map;

[0011] S4. Perform shape-constrained level set evolution on the target probability map and output the target segmentation results;

[0012] S5. Based on the target segmentation results, a three-dimensional target model is reconstructed using the moving cube algorithm; the Demons algorithm is used to register the three-dimensional target model with the preoperative images to obtain the initial deformation field; the preoperative images include CT images and MRI images.

[0013] S6. Apply a shadow region smoothing constraint to the initial deformation field to generate a modified deformation field, and generate the final puncture path based on the modified deformation field.

[0014] Secondly, this application also provides a precise target localization system for interventional therapy based on ultrasound imaging, comprising:

[0015] The image preprocessing module is used to extract high-frequency features and perform morphological processing on the input ultrasound images to generate a binary mask for the acoustic shadow region.

[0016] The enhancement processing module is used to dynamically adjust the contrast enhancement parameters based on the binary mask of the acoustic shadow area, and to perform regional enhancement on the ultrasound image based on the contrast enhancement parameters to obtain a regionally enhanced image.

[0017] The probability map generation module is used to input the enhanced image and the binary mask of the acoustic shadow region into the pre-trained neural network model to generate a target probability map;

[0018] The segmentation calculation module is used to perform shape-constrained level set evolution on the target probability map and output the target segmentation results.

[0019] The 3D reconstruction and registration module is used to reconstruct a 3D target model based on the target segmentation results using the moving cube algorithm; the Demons algorithm is used to register the 3D target model with preoperative images to obtain the initial deformation field; the preoperative images include CT images and MRI images.

[0020] The path correction module is used to apply a shadow region smoothing constraint to the initial deformation field, generate a corrected deformation field, and generate the final puncture path based on the corrected deformation field.

[0021] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a method for precise localization of interventional therapeutic targets based on ultrasound imaging as described in the first aspect.

[0022] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for precise localization of interventional therapeutic targets based on ultrasound imaging as described in the first aspect.

[0023] The aforementioned method and system for precise target localization in interventional therapy based on ultrasound imaging extracts acoustic shadow regions from ultrasound images to generate a binary mask. Based on this binary mask, contrast enhancement parameters for different regions are dynamically adjusted to suppress noise in shadow regions while enhancing target edge details. A shadow mask attention mechanism guides a neural network to accurately segment the target probability map. The segmentation boundary is optimized through level set evolution incorporating shape constraints from shadow regions. Simultaneously, a deformation field correction strategy based on shadow perception is used to locally smooth and compensate for registration deviations between the 3D reconstruction model and preoperative images, ultimately generating a puncture path that avoids dangerous structures. This effectively solves the problems of blurred target boundaries, accumulated segmentation errors, and inaccurate deformation compensation in traditional methods under acoustic shadow interference, significantly improving the targeting of image enhancement, the accuracy of target localization, and the safety of the interventional path. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating a method for precise target localization in interventional therapy based on ultrasound imaging, provided by this invention.

[0026] Figure 2 This is a schematic diagram of the process for generating a region-enhanced image in one optional embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the process of regional enhancement in an optional embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of the structure of an interventional therapy target precision localization system based on ultrasound imaging, provided by the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] refer to Figure 1 The document presents a flowchart illustrating a method for precise target localization in interventional therapy based on ultrasound imaging, as provided in this application. This method includes the following steps:

[0031] S1. Perform high-frequency feature extraction and morphological processing on the input ultrasound image to generate a binary mask for the acoustic shadow region.

[0032] Specifically, in ultrasound-guided interventional therapy, the acquired ultrasound image data is first used as input. Ultrasound images are essentially grayscale images formed by the interaction of different tissues with the reflection and absorption of ultrasound waves; the presence of acoustic shadow areas can interfere with subsequent processing.

[0033] The high-frequency feature extraction stage can leverage the specific high-frequency information distribution characteristics of acoustic shadows in ultrasound images. Methods such as wavelet transform and Gabor filtering can be used to process the image, highlighting high-frequency features such as the edge contours and texture details of the acoustic shadow region. For example, wavelet transform can decompose the image into different frequency sub-bands, and the approximate location of the acoustic shadow region can be determined by analyzing the energy distribution in the high-frequency sub-bands.

[0034] Morphological processing can utilize morphological operators (such as dilation, erosion, opening, and closing operations) to further process the extracted high-frequency features, eliminating noise and artifact interference and making the shape of acoustic shadow regions more regular and the boundaries clearer. For example, erosion can remove small isolated noise points in an image, while dilation can make the boundaries of acoustic shadow regions more complete.

[0035] Finally, a binary mask for the acoustic shadow region is generated. This mask is a binary image with the same size as the original ultrasound image, where the pixel value of the acoustic shadow region is 1 (or 0) and the pixel value of the non-shadow region is 0 (or 1), thereby clearly identifying the distribution range of the acoustic shadow in the ultrasound image.

[0036] S2. Based on the binary mask of the acoustic shadow region, dynamically adjust the contrast enhancement parameters, and perform regional enhancement on the ultrasound image according to the contrast enhancement parameters to obtain a regionally enhanced image.

[0037] Specifically, the purpose of contrast enhancement is to address the reduced image contrast caused by acoustic shadows, thereby better highlighting key areas such as targets. First, the grayscale distribution characteristics of acoustic shadow and non-shadow areas are analyzed. In non-shadow areas, because their grayscale information is relatively complete, traditional contrast enhancement methods, such as histogram equalization, can be used. By redistributing the grayscale values ​​of pixels, the grayscale range of this area is expanded, thus improving contrast. However, in acoustic shadow areas, directly applying methods such as histogram equalization may lead to over-enhancement due to the presence of noise, resulting in artifacts and distortion.

[0038] Therefore, this step dynamically adjusts contrast enhancement parameters by analyzing the gray-level distribution and noise characteristics of the acoustic shadow region. For example, adaptive histogram equalization is used to automatically determine the range and intensity of histogram equalization based on the local gray-level distribution of the shadow region, thereby improving contrast while suppressing noise amplification. Alternatively, image segmentation techniques can be used to further divide the shadow region into multiple sub-regions, and contrast enhancement parameters can be calculated separately for each sub-region. For instance, a suitable contrast enhancement factor can be determined based on factors such as the variance of the average gray-level value within the sub-region, and then contrast enhancement processing can be performed on each sub-region through a linear or non-linear gray-level mapping function.

[0039] The enhanced image can more clearly show the boundary and grayscale difference between the target point and the surrounding tissue, providing a better image basis for subsequent localization and segmentation operations.

[0040] S3. Input the enhanced image and the binary mask of the acoustic shadow region into the pre-trained neural network model to generate a target probability map.

[0041] Specifically, during the model training phase, a large dataset of ultrasound images labeled with target locations and acoustic shadow regions is first collected. This training data can be augmented using data augmentation techniques (such as rotation, translation, scaling, and adding noise) to increase the sample size and improve the model's generalization ability. The neural network architecture used can be a Convolutional Neural Network (CNN) or its variants, such as the U-Net network, which performs well in medical image segmentation tasks. The model structure includes multiple convolutional layers, pooling layers, upsampling layers, and skip connections. Convolutional layers extract local features from the images, such as edges, textures, and grayscale patterns; pooling layers reduce the spatial resolution of the feature maps, reducing computation and extracting key features; upsampling layers restore the feature maps to a size similar to the input image for pixel-level classification; and skip connections fuse low-level features with high-level features, helping to preserve detailed information in the images.

[0042] During training, loss functions (such as cross-entropy loss, Dice loss, etc.) are defined to measure the difference between the target probability map predicted by the model and the real annotation. The backpropagation algorithm is used to continuously adjust the weight parameters of the network, so that the model can gradually learn how to accurately identify the position and shape features of the target from the input enhanced image and the binary mask of the acoustic shadow region.

[0043] Once the model training is complete, in practice, the region-enhanced image obtained in step S2 and the binary mask of the acoustic shadow region generated in step S1 are used as inputs. After forward propagation calculation by the neural network model, a target probability map with the same or similar size as the input image is output. Each pixel value in this probability map represents the probability that the location belongs to the target. The higher the pixel value, the greater the probability that it belongs to the target, thus providing a probability-based reference for subsequent target segmentation operations.

[0044] S4. Perform shape-constrained level set evolution on the target probability map and output the target segmentation results.

[0045] Specifically, shape-constrained level set evolution is an image segmentation technique used to accurately obtain target segmentation results from a target probability map. The level set method essentially utilizes curve evolution theory to continuously evolve the initial contour curve within the image until it converges to the boundary of the target object.

[0046] In this step, a shape constraint is first defined based on the prior shape information of the target (such as a circle, ellipse, or regular polygon). For example, if the target typically presents an approximately elliptical shape, the shape constraint can be constructed based on the ellipse equation, with parameters including the lengths of the major and minor axes and the center coordinates. The initial contour curve can be a simple geometric shape (such as a circle or square), and its position and size are roughly determined based on the high-probability regions in the target probability map.

[0047] During the level set evolution, the evolution rate function can be determined by multiple factors: on the one hand, it is driven by the probability gradient information in the target probability map, causing the contour curve to shrink towards regions with higher probability values; on the other hand, the shape constraint term constrains the evolution of the contour curve, preventing it from excessively shrinking or deforming and deviating from the actual shape of the target. For example, when the contour curve tends to deviate from an elliptical shape during evolution, the shape constraint term generates a counterforce, pulling the contour curve back to a trajectory that conforms to the prior shape. Simultaneously, to ensure the accuracy and smoothness of the segmentation results, edge information from the image can be introduced as an auxiliary constraint term, enabling the contour curve to fit more accurately when approaching the true boundary of the target.

[0048] Through repeated iterative level set evolution, a stable contour curve is eventually obtained, and the region enclosed by this curve is the segmentation result of the target point. This segmentation result is presented in the form of a binary image, where the pixel value of the target point region is 1 (or 0), and the pixel value of the background and other non-target point regions is 0 (or 1), thereby achieving accurate segmentation of the target point and providing accurate target point boundary information for subsequent 3D reconstruction and puncture path planning.

[0049] S5. Based on the target segmentation results, a three-dimensional target model is reconstructed using the moving cube algorithm; the Demons algorithm is used to register the three-dimensional target model with the preoperative images to obtain the initial deformation field; the preoperative images include CT images and MRI images.

[0050] Specifically, the Moving Cubes Algorithm (MWA) is an algorithm for extracting isosurfaces from a three-dimensional scalar field. This step uses this algorithm to reconstruct the three-dimensional information of the target point. First, the ultrasound image sequence is arranged according to its actual position and orientation in space to construct a three-dimensional volume data. Each voxel in this three-dimensional volume data corresponds to a pixel in the ultrasound image, and its value is determined by the target point segmentation result at the corresponding position; that is, voxels belonging to the target point have a value of 1 (or 0), otherwise it is 0 (or 1). Then, the MWA algorithm is applied to process the three-dimensional volume data. The core of the algorithm is to divide the three-dimensional volume data into multiple cube units. For each cube unit, the intersection of the cube with the target point boundary is determined based on the voxel values ​​of its eight vertices. According to a predefined lookup table, the possible configurations of each cube unit intersecting with the target point boundary are determined, and corresponding triangular patches are generated to approximate the surface of the target point. By traversing all cube units, a three-dimensional target point model composed of multiple triangular patches is finally obtained. This model can intuitively show the shape, size, and positional relationship of the target point in three-dimensional space.

[0051] Next, the Demons algorithm is used to register the 3D target model with preoperative images. Preoperative images include CT and MRI images, which have high resolution and rich anatomical information. The Demons algorithm is a non-rigid registration algorithm based on grayscale information. Its basic principle is to treat the 3D target model as a deformable mesh, driving the deformation of the model by calculating the grayscale difference between the model and the preoperative image, gradually aligning it with the corresponding structure in the preoperative image. Specifically, first, the 3D target model is converted into volume data with the same size and voxel spacing as the preoperative image, and a displacement field is initialized to represent the displacement vector of each voxel in the model. Then, the grayscale difference between the model volume data and the preoperative image is calculated, using indices such as the difference of squares. A force field is calculated based on the grayscale difference, driving the model volume data to deform towards the direction of the preoperative image. In each iteration, the displacement field is updated, and the model volume data is deformed according to the new displacement field. This process is repeated until the grayscale difference meets a preset threshold condition or the maximum number of iterations is reached. Finally, an initial deformation field is obtained, which describes the displacement of each voxel in the 3D target model during the registration process, enabling the model to achieve preliminary alignment with the preoperative images in terms of anatomical structure.

[0052] This step provides target 3D information that matches the preoperative images for subsequent puncture path planning, which helps to combine the rich anatomical information from the preoperative images for more precise path design.

[0053] S6. Apply a shadow region smoothing constraint to the initial deformation field to generate a modified deformation field, and generate the final puncture path based on the modified deformation field.

[0054] Specifically, applying a shadow region smoothing constraint to the initial deformation field is to further improve registration accuracy and reduce registration errors caused by acoustic shadows. The principle of the shadow region smoothing constraint is based on the abnormal deformation characteristics that acoustic shadow regions may cause in the deformation field. In the initial deformation field, due to the presence of acoustic shadow regions, localized and unreasonable deformation peaks or valleys may appear, and these abnormal deformations can interfere with the accuracy of the puncture path.

[0055] Therefore, this step analyzes the deformation of the acoustic shadow's resonant region in the initial deformation field and designs a specific smoothing constraint function. For example, using a binary mask of the acoustic shadow region as a reference, constraints are applied to the shadow region and voxels within a certain range around it in the initial deformation field. The constraint function can employ methods such as mean filtering or Gaussian filtering within a local window to smooth the deformation values ​​within the shadow region, while simultaneously considering the deformation trend of the non-shadow region to ensure the continuity and consistency of the entire deformation field. Alternatively, an elastic body constraint model can be introduced, assuming that the tissue deformation in the shadow region conforms to certain elastic mechanical properties. By solving the elastic equation, the deformation field in the shadow region can be adjusted to coordinate with the deformation of the surrounding normal tissue. After smoothing constraint processing, a corrected deformation field is generated. The corrected deformation field more accurately reflects the spatial mapping relationship between intraoperative ultrasound images and preoperative images, effectively reducing deformation deviations caused by acoustic shadows.

[0056] Finally, based on the modified deformation field and the position of the 3D target model in the preoperative image, a specific puncture path is generated according to a specific puncture path planning algorithm (such as straight path planning, optimized path planning to avoid blood vessels and nerves). This puncture path accurately indicates the optimal path for the puncture instrument from the entry point on the body surface to the target point in 3D space, providing precise navigation guidance for interventional treatment and helping to improve the success rate and safety of the treatment.

[0057] The aforementioned method for precise target localization in interventional therapy based on ultrasound imaging extracts acoustic shadow regions from ultrasound images to generate a binary mask. Based on this mask, contrast enhancement parameters for different regions are dynamically adjusted to suppress noise in shadow regions while enhancing target edge details. A shadow mask attention mechanism guides a neural network to accurately segment the target probability map. The segmentation boundary is optimized through level set evolution incorporating shape constraints from shadow regions. Simultaneously, a deformation field correction strategy based on shadow perception is used to locally smooth and compensate for registration deviations between the 3D reconstruction model and preoperative images, ultimately generating a puncture path that avoids dangerous structures. This method effectively solves the problems of blurred target boundaries, accumulated segmentation errors, and inaccurate deformation compensation in traditional methods under acoustic shadow interference, significantly improving the targeting of image enhancement, the accuracy of target localization, and the safety of the interventional path.

[0058] In one optional embodiment, S1 includes the following steps:

[0059] S11. Perform high-pass filtering on the input ultrasound image to obtain a high-frequency component image.

[0060] Specifically, high-pass filtering aims to extract high-frequency information from ultrasound images, namely, areas with abrupt changes in grayscale, such as edges and textures. Acoustic shadow regions are often closely related to this high-frequency information. High-pass filtering methods can include Fourier transform-based high-pass filtering, gradient filtering, and Laplace filtering. Taking gradient-based high-pass filtering as an example, it highlights high-frequency components by calculating the rate of change of grayscale values ​​of each pixel in its neighborhood. In practice, operators such as the Sobel operator, Prewitt operator, or Roberts operator can be used to approximate the gradient magnitude of the image. For example, the Sobel operator, consisting of a pair of 3×3 convolution kernels, is used to calculate the gradient components in the horizontal and vertical directions of the image, respectively. The gradient components in these two directions are vector-synthesized to obtain the gradient magnitude value of each pixel, thereby generating a high-frequency component map. This high-frequency component map preserves detailed information such as edge contours in the ultrasound image, and the edges of acoustic shadow regions also exhibit high gradient magnitude values, laying the foundation for accurate identification of acoustic shadow regions.

[0061] S12. The high-frequency component image is segmented based on the first preset threshold to obtain an initial binary image; wherein, the first preset threshold is the preset quantile value of the pixel intensity in the high-frequency component image.

[0062] Specifically, the high-frequency component image is segmented based on a first preset threshold to initially separate the acoustic shadow region from the background and other tissue regions. A preset quantile value of the pixel intensity in the high-frequency component image can be used as the first preset threshold. The selection of the quantile value is based on statistical analysis of the high-frequency characteristics of the acoustic shadow region and verification with a large amount of experimental data. For example, statistical analysis of the high-frequency components of the acoustic shadow region in a large number of ultrasound images revealed that its pixel intensity usually falls within a specific quantile interval, such as the 70th or 80th quantile. Therefore, selecting the pixel intensity corresponding to this quantile value as the threshold can effectively distinguish the acoustic shadow region from other regions.

[0063] In practice, each pixel in the high-frequency component image is traversed. If the pixel intensity is greater than or equal to a first preset threshold, the pixel is marked as 1 in the initial binary image (indicating that it may belong to the acoustic shadow region), and otherwise marked as 0 (indicating that it does not belong to the acoustic shadow region). Although the initial binary image obtained in this way initially outlines the contour of the acoustic shadow region, due to the influence of noise, artifacts and other factors, its shape may not be complete, the boundaries may not be clear, and there may be problems such as breaks or burrs.

[0064] S13. Perform a closing operation on the initial binary image to generate a binary mask for the acoustic shadow region.

[0065] Specifically, closing is a fundamental operation in morphology used to address shape defects in target regions of binary images. The principle of closing is to first dilate the image, followed by erosion. Dilation involves sliding a structuring element (usually a small square or circular template) across the image. For each pixel, if at least one pixel in the neighborhood covered by the structuring element has a value of 1, then that pixel's value in the output image is set to 1. This causes the target region's boundary to expand outward, filling small gaps and breaks in the boundary. The subsequent erosion operation uses the same structuring element to slide across the dilated image. Only when all pixels in the neighborhood covered by the structuring element have a value of 1 is the corresponding pixel in the output image set to 1. This causes the target region's boundary to shrink inward.

[0066] By performing a closing operation that first expands and then erodes, the broken acoustic shadow region boundaries in the initial binary image can be effectively connected, and small holes in the region caused by noise and other factors can be filled in. This makes the shape of the acoustic shadow region more complete and compact, and the boundaries clearer and smoother. Finally, an accurate acoustic shadow region binary mask is generated, providing reliable acoustic shadow region positioning information for subsequent contrast enhancement and other processing steps.

[0067] refer to Figure 2 In one optional embodiment, S2 includes the following steps:

[0068] S21. Divide the ultrasound image into multiple local image blocks of predefined sizes.

[0069] Specifically, the ultrasound image is divided into multiple predefined local image blocks. The purpose is to perform regional processing of the image to more accurately analyze the impact of acoustic shadowing on each local area. First, the size of the local image blocks is determined based on the resolution of the ultrasound image and the requirements of the actual application scenario. For example, the image can be divided into 8×8, 16×16, or 32×32 pixel square blocks. During division, the image can be gridded sequentially from left to right and from top to bottom to ensure that the position of each local image block in the image is uniquely determined, i.e., the local image block B in the i-th row and j-th column. ij This corresponds to specific regions in the image. This division method allows subsequent processing of each local image patch to more specifically consider the acoustic shadow characteristics within that region, and also facilitates the integration of processing results from different regions.

[0070] S22. Based on the binary mask of the acoustic shadow region, calculate the shadow coverage ratio of each local image patch; the formula for calculating the shadow coverage ratio is:

[0071]

[0072] Among them, B ij N represents a local image patch in the i-th row and j-th column of an ultrasound image. ij M represents the total number of pixels in the local image patch at row i and column j in the ultrasound image. s (x,y) represents the binary mask for the acoustic shadow region at pixel coordinates (x,y) in the ultrasound image. This represents the shadow coverage ratio of a local image patch in the i-th row and j-th column of an ultrasound image.

[0073] Specifically, calculating the shadow coverage ratio of each local image patch based on the binary mask of the acoustic shadow region assesses the degree of shadow influence on the region by quantifying the proportion of acoustic shadow in each local image patch. The specific calculation process is as follows: For local image patch B in the i-th row and j-th column of the ultrasound image... ij First, determine the local image patch within the acoustic shadow region using the binary mask M. s The corresponding region in (x, y). Then, count the number of pixels with a value of 1 (representing an acoustic shadow region) within this region, i.e., calculate... Simultaneously, acquire local image patch B ij Total number of pixels N ij Finally, according to the formula The local image patch B is calculated. ij Shadow coverage ratio

[0074] This ratio A value between 0 and 1 reflects the degree of acoustic shadow coverage within that local image patch. For example, if... A value close to 1 indicates that the local image patch is almost completely covered by acoustic shadows; if... A value close to 0 indicates that the local image patch is almost unaffected by acoustic shadowing. This quantification provides a crucial basis for subsequent dynamic adjustment of contrast enhancement parameters, enabling differentiated operation of contrast enhancement processing for different regions based on their actual degree of shadow influence.

[0075] S23. Dynamically adjust the contrast enhancement parameters according to the shadow coverage ratio, and perform regional enhancement on the ultrasound image according to the contrast enhancement parameters to obtain a regionally enhanced image.

[0076] Specifically, the purpose of this step is to improve the overall contrast of the image while avoiding excessive noise enhancement in acoustic shadow areas, thereby improving image quality.

[0077] First, a set of basic contrast enhancement parameters is determined for the entire ultrasound image. These parameters may include the grayscale mapping range of histogram equalization, contrast adjustment factors, etc. Then, based on each local image patch B... ij Shadow coverage ratio The basic contrast enhancement parameters are dynamically adjusted. For example, for local image patches with a high shadow coverage ratio, the intensity of contrast enhancement is appropriately reduced to avoid noise amplification and artifacts caused by over-enhancement; while for local image patches with a low shadow coverage ratio, the intensity of contrast enhancement can be appropriately increased to better highlight the details of the tissue.

[0078] The adjusted contrast enhancement parameters are used to enhance the local image patch using methods such as histogram equalization and linear grayscale stretching. After enhancing each local image patch individually, the enhanced patches are reassembled according to their positions in the original ultrasound image to obtain a region-enhanced ultrasound image. This region-enhanced image achieves a reasonable improvement in contrast across different regions, effectively suppresses the impact of acoustic shadows on image quality, and provides a clearer and more accurate image basis for subsequent target localization and segmentation operations.

[0079] refer to Figure 3 In an optional embodiment, S23 includes the following steps:

[0080] S231. Calculate the contrast limiting parameter using a dynamic gain function based on the shadow coverage ratio; the expression for the dynamic gain function is:

[0081]

[0082] Among them, C base Here, α is the baseline contrast limit, β is a preset coefficient used to control the gain decay rate in the shadow area, and β is a preset coefficient used to adjust the nonlinear effect of the shadow coverage ratio on the dynamic gain function. This represents the contrast limiting parameter corresponding to the local image block in the i-th row and j-th column of an ultrasound image.

[0083] Specifically, the dynamic gain function is used to calculate contrast limiting parameters based on the shadow coverage ratio, aiming to balance image contrast enhancement and noise control. The baseline contrast limiting value C... base This is the foundation of the entire dynamic gain function, determining the intensity of contrast enhancement under ideal conditions without acoustic shadowing. The preset coefficient α controls the rate of gain decay in shadowed areas, i.e., how quickly the contrast enhancement parameter decreases as the shadow coverage increases. A larger α value causes the contrast enhancement parameter to decrease rapidly in shadowed areas, thus more effectively suppressing noise; while a smaller α value retains relatively more contrast enhancement, suitable for situations with lighter shadows or less noise. The preset coefficient β adjusts the non-linear effect of shadow coverage on the dynamic gain function, determining the curvature of the dynamic gain function curve. A larger β value causes the function to exhibit more significant non-linear changes as the shadow coverage changes, allowing for more refined contrast adjustments in areas with different shadow levels. By adjusting the shadow coverage... Substitute into the dynamic gain function A unique contrast limit parameter can be calculated for each local image patch. This ensures that the contrast enhancement process can be dynamically adjusted according to the actual shadow conditions in each area, thereby enhancing image details while suppressing noise.

[0084] S232. Based on the contrast limiting parameter, perform standard CLAHE enhancement on local image blocks where the shadow coverage ratio is lower than the second preset threshold to obtain a standard enhanced image.

[0085] Specifically, for local image patches with shadow coverage below a second preset threshold, the standard CLAHE (Adaptive Histogram Equalization) enhancement method is used. CLAHE is a commonly used local contrast enhancement technique that improves image contrast by dividing the image into multiple small patches and performing histogram equalization on each patch. In the standard CLAHE enhancement process, the gray-level histogram of the local image patch is first equalized to make the gray-level distribution within the region more uniform, thereby improving contrast. This process effectively enhances tissue details, making key structures such as target points more clearly visible. Because these local image patches have low shadow coverage and relatively little noise impact, standard CLAHE enhancement is sufficient to achieve good enhancement results, yielding a standard enhanced image.

[0086] S233. Based on the contrast limiting parameter, for local image patches where the shadow coverage ratio is higher than or equal to the second preset threshold, noise-constrained CLAHE enhancement is performed to obtain a noise-constrained enhanced image; the expression for noise-constrained CLAHE enhancement is:

[0087]

[0088] Among them, I enh (x,y) represents the noise-reduced constrained enhanced image, and I(x,y) represents the ultrasound image. The result is the ultrasound image enhanced by applying the CLAHE algorithm with contrast limiting parameters. BilateralFilter(I(x,y)) represents the result of applying bilateral filtering to the ultrasound image, and λ is the preset noise reduction weight.

[0089] Specifically, for local image patches with a shadow coverage ratio higher than or equal to a second preset threshold, a noise-constrained CLAHE enhancement method is used. These local image patches, due to their high shadow coverage ratio, are usually accompanied by significant noise interference. Therefore, noise suppression is necessary while enhancing contrast. The expression for noise-constrained CLAHE enhancement combines CLAHE enhancement and bilateral filtering techniques. Among them, Indicates the contrast limiting parameter The following is the result of CLAHE enhancement on ultrasound image I(x,y) to improve image contrast; BilateralFilter(I(x,y)) represents the result of bilateral filtering on ultrasound image I(x,y). Bilateral filtering can effectively suppress noise while preserving edge information. A preset noise reduction weight λ is used to balance the effects of CLAHE enhancement and bilateral filtering. A larger λ value allows more bilateral filtering components to be incorporated into the final result, thus achieving a stronger noise reduction effect, but may slightly weaken the contrast enhancement effect; while a smaller λ value focuses more on contrast enhancement. Through this combination, noise-constrained CLAHE enhancement effectively controls noise amplification while improving contrast, resulting in a noise-constrained enhanced image.

[0090] S234. Based on the standard enhanced image and the noise-reducing constrained enhanced image, obtain the region-specific enhanced image.

[0091] Specifically, after enhancing all local image patches, the standard enhanced image and the noise-constrained enhanced image are integrated to obtain the final region-enhanced image. In practice, the enhanced local image patches are sequentially stitched together according to their position in the original ultrasound image. During the stitching process, smooth transitions between different local image patches are ensured to avoid obvious stitching marks. This can be achieved by weighted fusion of overlapping areas of adjacent local image patches, such as using linear interpolation, to maintain visual consistency in the enhanced image. The resulting region-enhanced image achieves a reasonable increase in contrast across the entire image range while effectively suppressing noise interference in acoustic shadow areas, providing a higher quality and clearer image foundation for subsequent target localization and segmentation operations.

[0092] In an optional embodiment, S3 includes the following steps:

[0093] S31. Perform channel stitching between the region-enhanced image and the acoustic shadow region binary mask to generate dual-channel input data.

[0094] Specifically, the region-enhanced image is a two-dimensional grayscale image where each pixel contains an intensity value representing the image information at that location. The acoustic shadow region binary mask is also a two-dimensional binary image of the same size as the region-enhanced image, where the value of each pixel indicates whether the location belongs to an acoustic shadow region. To integrate these two types of information, the region-enhanced image is used as one channel, and the acoustic shadow region binary mask as another channel, and they are concatenated along the channel dimensions. For example, if the region-enhanced image is a two-dimensional array of shape H×W (height × width), then the concatenated dual-channel input data will be a three-dimensional array of shape H×W×2, where the first layer of the third dimension stores the pixel values ​​of the region-enhanced image, and the second layer stores the pixel values ​​of the acoustic shadow region binary mask. This channel concatenation method allows the model to utilize both the image enhancement information and the acoustic shadow region mask information simultaneously during training and inference, thereby more accurately identifying and locating target points.

[0095] S32. Input the dual-channel input data into the encoder of the SMA-U-Net model, and extract multi-scale encoder feature maps through convolution and downsampling operations.

[0096] Specifically, the encoder is a key component of the SMA-U-Net model. Its role is to progressively extract feature information from the input data and gradually reduce the spatial resolution of the feature maps through downsampling operations, while increasing the number of channels in the feature maps to capture higher-level semantic information. In convolutional operations, different sizes and numbers of convolutional kernels are used to process the input data. Each kernel slides within a local region of the input data, calculating a weighted sum within that region to extract local features. For example, 3×3 or 5×5 convolutional kernels can be used, and by stacking multiple convolutional layers, low-level features such as edges and textures, as well as more abstract high-level features, can be extracted from the input data. Downsampling operations can be implemented through max pooling, average pooling, or stride convolution, aiming to reduce the spatial size of the feature maps while retaining the most important feature information. For example, using a 2×2 max pooling operation can halve the size of the feature map while highlighting salient features. Through multiple convolution and downsampling operations, the encoder can generate a series of multi-scale feature maps containing feature information at different levels from shallow to deep, providing a foundation for subsequent feature fusion and decoding.

[0097] S33. In the encoder skip connection, the encoder feature map of each layer in the SMA-U-Net model is fused with the binary mask of the acoustic shadow region to generate an attention-weighted feature map; the expression for feature fusion is:

[0098]

[0099] in, This represents the encoder feature map output by the k-th layer encoder of the SMA-U-Net model. UpSample(M) represents the attention-weighted feature map corresponding to the encoder feature map output by the k-th layer encoder. s (x,y)) represents the output result after upsampling the binary mask of the acoustic shadow region to the same spatial resolution as the encoder feature map. Indicates to and UpSample(M) s (x,y)) performs channel concatenation operation; ⊙ represents element-wise multiplication; Conv 1×1 (·) represents a 1×1 convolution operation, and σ(·) is the Sigmoid function.

[0100] Specifically, skip connections are a key feature of U-Net-type models, allowing information to be passed from shallow features of the encoder to the decoder, thus preserving image details to some extent and aiding in accurate localization. During feature fusion, the acoustic shadow region binary mask is first upsampled to the same spatial resolution as the encoder feature map of the current layer. Upsampling can be achieved through bilinear interpolation, nearest neighbor interpolation, or deconvolution to ensure the acoustic shadow region binary mask corresponds to the encoder feature map in the same spatial location. Then, the upsampled acoustic shadow region binary mask and the corresponding encoder feature map are concatenated along the channel dimension to form a new feature map. Next, a 1×1 convolution operation is used to process this concatenated feature map. The 1×1 convolution fuses information from different channels, adjusts the number of channels in the feature map, and calculates the attention weight at each location. The Sigmoid activation function normalizes these weights to between 0 and 1, ensuring that the attention weight at each location represents the importance of that location in the current layer's feature map. Finally, these attention weights are multiplied element-wise with the original encoder feature map to obtain an attention-weighted feature map. Attention-weighted feature maps emphasize important features related to acoustic shadow regions and suppress irrelevant features, thereby improving the model's ability to identify target regions.

[0101] S34. Input the attention-weighted feature maps output by each encoder layer into the decoder of the SMA-U-Net model, fuse multi-scale features layer by layer through deconvolution and skip connections, and output the target probability map through the Sigmoid activation function; where the target probability map represents the probability that each pixel in the ultrasound image belongs to the target.

[0102] Specifically, the decoder aims to progressively restore the multi-scale features extracted by the encoder to the same spatial resolution as the input image, while fusing detailed information from skip connections in the encoder. Deconvolution (also known as transposed convolution) increases the spatial size of the feature map, restoring image detail. For example, using a deconvolution with a stride of 2 can double the size of the feature map. Skip connections fuse the feature map in the decoder with the attention-weighted feature map of the corresponding layer in the encoder. This fusion allows the decoder to combine high-level semantic information with low-level detail information, thus more accurately reconstructing the target region in the image. During layer-by-layer upsampling and feature fusion, the semantic information of the feature map is gradually refined, ultimately generating a feature map of the same size as the input image. Finally, the Sigmoid activation function maps the value of each pixel to between 0 and 1, resulting in a target probability map. The value of each pixel in the probability map represents the probability that the pixel belongs to the target region; a higher probability value indicates that the pixel is more likely to belong to the target region. This probability map provides crucial information for subsequent target segmentation, enabling the model to accurately locate and segment targets based on probabilistic information.

[0103] In an optional embodiment, S4 includes the following steps:

[0104] S41. Based on a preset probability threshold, the target probability map is binarized to generate an initial target region binary map.

[0105] Specifically, the preset probability threshold is a value between 0 and 1, which can be determined based on statistical analysis of a large amount of experimental data and the requirements for target detection accuracy. For example, if the probability threshold is set to 0.5, then if the probability value of each pixel in the target probability map is greater than or equal to 0.5, the pixel is considered to belong to the target region and is marked as 1 in the initial target region binary map; conversely, if the probability value is less than 0.5, the pixel is considered not to belong to the target region and is marked as 0. This binarization method based on a fixed threshold is simple and direct, and can quickly convert the probability map into a preliminary target region binary map.

[0106] S42. Perform Euclidean distance transformation on the binary map of the initial target region to generate a symbolic distance function, and set the symbolic distance function as the initial value of the level set.

[0107] Specifically, the Euclidean distance transform is a mathematical operation that converts each pixel in a binary image into its Euclidean distance to the nearest background pixel (i.e., a pixel with a value of 0). The signed distance function is an extension of the Euclidean distance transform; it not only represents the distance but also includes sign information: for pixels belonging to the target region (value 1), the signed distance is positive, representing its distance to the nearest background pixel; for pixels in the background region (value 0), the signed distance is negative, representing its distance to the nearest pixel in the target region. In practical calculations, efficient algorithms such as the fast marching algorithm can be used to compute the signed distance function. Setting the signed distance function as the initial value of the level set aims to ensure that, during the evolution of the level set, the function gradually converges to the true boundary of the target point through an energy minimization process, based on this initial value.

[0108] S43. Construct the energy function of the level set based on the initial value of the level set.

[0109] Specifically, the level set energy function can be composed of internal and external energy terms. The internal energy term ensures the smoothness of the level set function, avoiding overly convoluted boundaries, and can be achieved by penalizing the gradient of the level set function. The external energy term guides the level set function to evolve towards the true boundary of the target point, and can be based on image grayscale information, gradient information, or prior shape information. For example, the grayscale contrast between the target region and the background region can be used to construct the external energy term, resulting in lower energy values ​​for the level set function near regions with significant grayscale changes (i.e., the target boundary). Furthermore, a shape prior term can be added, using the prior shape model of the target point (such as a circle or ellipse) to constrain the external energy term, making the evolution process more consistent with the actual shape characteristics of the target point. By reasonably defining and combining these energy terms, an energy function that accurately guides the level set function to converge to the target boundary can be constructed.

[0110] S44. Update the level set energy function using gradient descent until convergence, and obtain the target segmentation result.

[0111] Specifically, gradient descent is a numerical computation method based on optimization. Its basic idea is to iteratively update the level set function along the negative direction of the energy function gradient, gradually decreasing the energy function value until it reaches its minimum. In each iteration, the gradient of the level set energy function with respect to the level set function is first calculated. This involves differentiating the internal and external energy terms separately and combining them. Then, the level set function is updated based on the gradient value. The choice of the update step size (i.e., learning rate) balances convergence speed and stability; an excessively large step size may lead to oscillations and divergence, while an excessively small step size will prolong the convergence time. During the iteration process, reasonable stopping conditions are set, such as the change in the energy function value being less than a certain threshold, the change in the level set function being less than a certain threshold, or reaching the maximum number of iterations. When the stopping conditions are met, the level set function is considered to have converged to a stable boundary. At this point, the zero level set corresponding to the level set function is extracted, yielding the target segmentation result. This segmentation result is a binary image that identifies the position and contour of the target point in the ultrasound image.

[0112] In an optional embodiment, S6 includes the following steps:

[0113] S61. Based on the acoustic shadow region binary mask, extract the set of three-dimensional spatial coordinates of the shadow region.

[0114] Specifically, the acoustic shadow region binary mask is a set of two-dimensional images corresponding to an ultrasound image sequence, with the location of the acoustic shadow region marked in each two-dimensional image. To locate these shadow regions in three-dimensional space, the two-dimensional coordinates of the acoustic shadow pixels in each two-dimensional image can be converted into three-dimensional spatial coordinates by combining the displacement information of the ultrasound probe and the geometric parameters during image acquisition (such as the scanning step size and angle). For example, assuming that the ultrasound images are acquired through linear scanning, and each two-dimensional image corresponds to a specific lateral position, then the three-dimensional coordinates of each acoustic shadow pixel can be calculated using its row and column coordinates in the two-dimensional image and its corresponding lateral position. Finally, the three-dimensional coordinates of all acoustic shadow pixels are collected to form a set of three-dimensional spatial coordinates.

[0115] S62. Based on the set of three-dimensional spatial coordinates, apply anisotropic smoothing constraints to the initial deformation field within the shadow region to obtain the modified deformation field.

[0116] Specifically, the core of anisotropic smoothing constraints is to apply varying degrees of smoothing in different spatial directions to protect important structural information in the image. In acoustic shadow regions, the initial deformation field may exhibit abnormal deformation vectors due to significant interference with the grayscale distribution and structural information of ultrasound images. Therefore, by utilizing the three-dimensional spatial coordinate set of the shadow region, stronger smoothing constraints can be applied to the deformation vectors within these regions. For example, a smaller smoothing weight can be applied in the normal direction of the shadow region to maintain the sensitivity of the deformation field to changes in that direction, while a larger smoothing weight can be applied in the tangential direction to suppress noise and artifacts in the deformation field. Through this anisotropic smoothing process, the interference of the shadow region on the deformation field can be effectively reduced, making the deformation field more consistent with the actual tissue deformation, resulting in a corrected deformation field.

[0117] S63. Apply the correction deformation field to the coordinate system of the preoperative image to generate a three-dimensional space for deformation correction.

[0118] Specifically, preoperative images (such as CT or MRI images) have their own three-dimensional coordinate system, while the acquisition process of ultrasound images is usually accompanied by a certain degree of deformation. The corrected deformation field describes the corresponding position of each point in the ultrasound image within the preoperative image coordinate system. By applying the corrected deformation field to the preoperative image coordinate system, the preoperative image can be non-rigidly deformed in three-dimensional space to match the anatomical structures in the ultrasound image. This step may involve interpolating the preoperative image to estimate new pixel values ​​at the deformed coordinate positions. For example, methods such as trilinear interpolation or cubic spline interpolation can be used to calculate the pixel values ​​at the deformed positions based on the values ​​of surrounding pixels. The resulting deformed 3D space is a 3D representation of the preoperative image registered with the ultrasound image.

[0119] S64. Mark blood vessels and nerve structures as obstacle regions in the three-dimensional space of deformation correction, and construct an obstacle distance field based on the obstacle regions.

[0120] Specifically, the locations of blood vessels and neural structures are first identified and marked in the deformation-corrected 3D space. This can be achieved by extracting 3D models of blood vessels and neural structures from preoperative images using segmentation algorithms, and then converting these models into the deformation-corrected 3D space.

[0121] After marking the obstacle region, an obstacle distance field is constructed. This obstacle distance field is a three-dimensional scalar field of the same size as the deformation-corrected 3D space, where the value of each voxel represents the distance from that voxel to the nearest obstacle (blood vessel or neural structure). Methods for constructing the obstacle distance field can include fast scan-conversion algorithms or priority queue-based algorithms, which can calculate the shortest distance from each voxel to the obstacle. The obstacle distance field provides crucial reference information for puncture path planning; the puncture path should be as far away from obstacles as possible to avoid damage to blood vessels and neural structures.

[0122] S65. Based on the deformation correction three-dimensional space and obstacle distance field, the final puncture path is generated by the RRT* algorithm.

[0123] Specifically, the RRT* (Rapidly-exploring Random Tree Star) algorithm is a sampling-based algorithm for path planning in high-dimensional space. Its goal is to find the optimal path from the starting point to the target point under certain constraints. In this case, the starting point can be the puncture inlet point on the body surface, and the target point is the center position of the three-dimensional target model. During path planning, the RRT* algorithm explores possible paths by randomly sampling in the deformation-corrected three-dimensional space and constructing a rapidly expanding random tree. Each time the tree is expanded, the algorithm considers distance information from the obstacle distance field to ensure that the newly generated path segment does not collide with obstacles and is as far away from obstacles as possible. In addition, the RRT* algorithm also has the ability to optimize paths, progressively improving the generated path to make it as short and smooth as possible while satisfying the constraints. The final puncture path is a sequence of three-dimensional coordinate points. These points are connected to form a safe and feasible puncture path from the body surface to the target point, providing precise navigation guidance for interventional treatment.

[0124] The aforementioned method for precise target localization in interventional therapy based on ultrasound imaging generates a dynamic mask by extracting the acoustic shadow region from the ultrasound image. Based on the mask, local image blocks are divided to calculate the shadow coverage ratio. Contrast enhancement parameters for different regions are adaptively adjusted according to the shadow coverage ratio, suppressing shadow noise while preserving target edge details. A shadow mask attention mechanism guides a neural network to fuse image features and mask information, generating a high-precision target probability map. The segmentation boundary is optimized by incorporating level set evolution constrained by the shape of the shadow region. Combining the non-rigid registration of preoperative images with the intraoperative 3D reconstruction model, local smoothing correction of the shadow region is applied to the deformation field. Finally, a safe puncture path avoiding blood vessels and nerves is planned in the corrected anatomical space. This method comprehensively solves the problems of target blurring, accumulated segmentation errors, and inaccurate deformation compensation caused by acoustic shadowing, significantly improving the targeting of image enhancement, the accuracy of target localization, and the safety of the interventional path.

[0125] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0126] Based on the same inventive concept, this application also provides a system for implementing the above-mentioned method for precise positioning of interventional treatment targets based on ultrasound imaging. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the system for precise positioning of interventional treatment targets based on ultrasound imaging provided below can be found in the limitations of the method for precise positioning of interventional treatment targets based on ultrasound imaging described above, and will not be repeated here.

[0127] In one exemplary embodiment, such as Figure 4 As shown, an interventional therapy target precision localization system 40 based on ultrasound imaging is provided, comprising:

[0128] The image preprocessing module 41 is used to extract high-frequency features and perform morphological processing on the input ultrasound image to generate a binary mask for the acoustic shadow region.

[0129] The enhancement processing module 42 is used to dynamically adjust the contrast enhancement parameters according to the binary mask of the acoustic shadow region, and perform regional enhancement on the ultrasound image according to the contrast enhancement parameters to obtain a regionally enhanced image.

[0130] The probability map generation module 43 is used to input the enhanced image and the binary mask of the acoustic shadow region into the pre-trained neural network model to generate a target probability map.

[0131] The segmentation calculation module 44 is used to perform shape-constrained level set evolution on the target probability map and output the target segmentation results.

[0132] The 3D reconstruction and registration module 45 is used to reconstruct the 3D target model based on the target segmentation results using the moving cube algorithm; the Demons algorithm is used to register the 3D target model with the preoperative images to obtain the initial deformation field; the preoperative images include CT images and MRI images.

[0133] The path correction module 46 is used to apply a shadow region smoothing constraint to the initial deformation field, generate a corrected deformation field, and generate the final puncture path based on the corrected deformation field.

[0134] Optional, the image preprocessing module includes:

[0135] The high-pass filter unit is used to perform high-pass filtering on the input ultrasound image to obtain a high-frequency component map.

[0136] The first threshold segmentation unit is used to segment the high-frequency component image based on a first preset threshold to obtain an initial binary image; wherein, the first preset threshold is a preset quantile value of the pixel intensity in the high-frequency component image.

[0137] The closing operation unit is used to perform closing operations on the initial binary image to generate a binary mask for the acoustic shadow region.

[0138] Optional, enhanced processing modules include:

[0139] Image block division unit is used to divide ultrasound images into multiple local image blocks of predefined sizes.

[0140] The shadow coverage ratio calculation unit is used to calculate the shadow coverage ratio of each local image patch based on the acoustic shadow region binary mask; the formula for calculating the shadow coverage ratio is:

[0141]

[0142] Among them, B ij N represents a local image patch in the i-th row and j-th column of an ultrasound image. ij M represents the total number of pixels in the local image patch at row i and column j in the ultrasound image. s (x,y) represents the binary mask for the acoustic shadow region at pixel coordinates (x,y) in the ultrasound image. This represents the shadow coverage ratio of a local image patch in the i-th row and j-th column of an ultrasound image.

[0143] The contrast enhancement parameter adjustment and enhancement unit is used to dynamically adjust the contrast enhancement parameters according to the shadow coverage ratio, and to enhance the ultrasound image by region according to the contrast enhancement parameters to obtain a region-enhanced image.

[0144] Optional, the contrast enhancement parameter adjustment and enhancement unit includes:

[0145] The contrast limiting parameter calculation subunit is used to calculate the contrast limiting parameter based on the shadow coverage ratio using a dynamic gain function; the expression for the dynamic gain function is:

[0146]

[0147] Among them, Cbase Here, α is the baseline contrast limit, β is a preset coefficient used to control the gain decay rate in the shadow area, and β is a preset coefficient used to adjust the nonlinear effect of the shadow coverage ratio on the dynamic gain function. This represents the contrast limiting parameter corresponding to the local image block in the i-th row and j-th column of an ultrasound image.

[0148] The standard enhancement subunit is used to perform standard CLAHE enhancement on local image blocks with a shadow coverage ratio lower than a second preset threshold based on the contrast limit parameter, so as to obtain a standard enhanced image.

[0149] The noise reduction constraint enhancement subunit is used to perform noise reduction constraint CLAHE enhancement on local image patches whose shadow coverage ratio is higher than or equal to a second preset threshold, based on a contrast constraint parameter, to obtain a noise reduction constraint enhanced image; the expression for noise reduction constraint CLAHE enhancement is:

[0150]

[0151] Among them, I enh (x,y) represents the noise-reduced constrained enhanced image, and I(x,y) represents the ultrasound image. The result is the ultrasound image enhanced by applying the CLAHE algorithm with contrast limiting parameters. BilateralFilter(I(x,y)) represents the result of applying bilateral filtering to the ultrasound image, and λ is the preset noise reduction weight.

[0152] The region-based image enhancement generation subunit is used to obtain a region-based enhanced image based on the standard enhanced image and the noise-reducing constrained enhanced image.

[0153] Optionally, the probability map generation module includes:

[0154] The channel stitching unit is used to stitch the region-enhanced image with the acoustic shadow region binary mask to generate dual-channel input data.

[0155] The encoder processing unit is used to input dual-channel input data into the encoder of the SMA-U-Net model and extract multi-scale encoder feature maps through convolution and downsampling operations.

[0156] The attention feature fusion unit is used in the encoder skip connections to fuse the encoder feature maps of each layer in the SMA-U-Net model with the binary mask of the acoustic shadow region to generate an attention-weighted feature map; the expression for feature fusion is:

[0157]

[0158] in, This represents the encoder feature map output by the k-th layer encoder of the SMA-U-Net model. UpSample(M) represents the attention-weighted feature map corresponding to the encoder feature map output by the k-th layer encoder. s (x,y)) represents the output result after upsampling the binary mask of the acoustic shadow region to the same spatial resolution as the encoder feature map. Indicates to and UpSample(M) s (x,y)) performs channel concatenation operation; ⊙ represents element-wise multiplication; Conv 1×1 (·) represents a 1×1 convolution operation, and σ(·) is the Sigmoid function.

[0159] The decoder processing unit is used to input the attention-weighted feature maps output by each encoder layer into the decoder of the SMA-U-Net model. It fuses multi-scale features layer by layer through deconvolution and skip connections and outputs a target probability map through the Sigmoid activation function. The target probability map represents the probability that each pixel in the ultrasound image belongs to the target.

[0160] Optionally, the segmentation calculation module includes:

[0161] The binarization processing unit is used to perform binarization processing on the target probability map based on a preset probability threshold to generate an initial target region binary map.

[0162] The distance transformation and initial value setting unit is used to perform Euclidean distance transformation on the binary map of the initial target area, generate a symbolic distance function, and set the symbolic distance function as the initial value of the level set.

[0163] Energy function construction unit, used to construct the energy function of the level set based on the initial value of the level set.

[0164] The energy function update unit is used to update the energy function of the level set using the gradient descent method until convergence, thus obtaining the target segmentation result.

[0165] Optional, the path correction module includes:

[0166] The 3D coordinate extraction unit is used to extract the set of 3D spatial coordinates of the shadow region based on the acoustic shadow region binary mask.

[0167] The smoothing constraint element is used to apply anisotropic smoothing constraints to the initial deformation field within the shaded region based on the set of three-dimensional spatial coordinates, thereby obtaining the modified deformation field.

[0168] The three-dimensional spatial correction unit is used to apply the correction deformation field to the coordinate system of the preoperative image to generate a deformation correction three-dimensional space.

[0169] The obstacle distance field construction unit is used to mark blood vessels and neural structures as obstacle regions in the deformation correction three-dimensional space, and construct the obstacle distance field based on the obstacle regions.

[0170] The path planning unit is used to generate the final puncture path based on the deformation-corrected 3D space and obstacle distance field using the RRT* algorithm.

[0171] Embodiments of this application also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned method embodiments.

[0172] Embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0173] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0174] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for precise target localization in interventional therapy based on ultrasound imaging, characterized in that, The method includes: S1. Perform high-frequency feature extraction and morphological processing on the input ultrasound image to generate a binary mask for the acoustic shadow region. S2. Based on the binary mask of the acoustic shadow region, dynamically adjust the contrast enhancement parameters, and perform regional enhancement on the ultrasound image according to the contrast enhancement parameters to obtain a regionally enhanced image. S3. Input the enhanced image and the acoustic shadow region binary mask into a pre-trained neural network model to generate a target probability map; S4. Perform shape-constrained level set evolution on the target probability map and output the target segmentation result; S5. Based on the target segmentation results, a three-dimensional target model is reconstructed using the moving cube algorithm; the three-dimensional target model is registered with the preoperative images using the Demons algorithm to obtain the initial deformation field; wherein, the preoperative images include CT images and MRI images; S6. Apply a shadow region smoothing constraint to the initial deformation field to generate a modified deformation field, and generate the final puncture path based on the modified deformation field.

2. The method according to claim 1, characterized in that, S1 includes: S11. Perform high-pass filtering on the input ultrasound image to obtain a high-frequency component image; S12. The high-frequency component image is segmented based on a first preset threshold to obtain an initial binary image; wherein, the first preset threshold is a preset quantile value of the pixel intensity in the high-frequency component image; S13. Perform a closing operation on the initial binary image to generate the acoustic shadow region binary mask.

3. The method according to claim 2, characterized in that, S2 includes: S21. Divide the ultrasound image into multiple local image blocks of predefined sizes; S22. Based on the acoustic shadow region binary mask, calculate the shadow coverage ratio of each local image block; the formula for calculating the shadow coverage ratio is: Among them, B ij N represents the local image patch in the i-th row and j-th column of the ultrasound image. ij M is the total number of pixels in the local image patch in the i-th row and j-th column of the ultrasound image. s (x,y) represents the binary mask of the acoustic shadow region at pixel coordinates (x,y) in the ultrasound image. This represents the shadow coverage ratio of the local image patch in the i-th row and j-th column of the ultrasound image; S23. Dynamically adjust the contrast enhancement parameters according to the shadow coverage ratio, and perform regional enhancement on the ultrasound image according to the contrast enhancement parameters to obtain the regional enhanced image.

4. The method according to claim 3, characterized in that, S23 includes: S231. Based on the shadow coverage ratio, calculate the contrast limiting parameter using a dynamic gain function; the expression for the dynamic gain function is: Among them, C base Here, α is the baseline contrast limit value, β is a preset coefficient used to control the gain attenuation rate in the shadow area, and β is a preset coefficient used to adjust the nonlinear effect of the shadow coverage ratio on the dynamic gain function. The contrast limiting parameter represents the local image block in the i-th row and j-th column of the ultrasound image. S232. Based on the contrast limiting parameter, perform standard CLAHE enhancement on the local image block whose shadow coverage ratio is lower than the second preset threshold to obtain a standard enhanced image; S233. Based on the contrast limiting parameter, for the local image patch where the shadow coverage ratio is higher than or equal to the second preset threshold, noise reduction constraint CLAHE enhancement is performed to obtain a noise reduction constraint enhanced image; the expression for the noise reduction constraint CLAHE enhancement is: Among them, I enh (x,y) represents the noise-reduced constrained enhanced image, and I(x,y) represents the ultrasound image. The result of applying the CLAHE algorithm to the ultrasound image and enhancing it with the contrast limiting parameters is represented by BilateralFilter(I(x,y)), which represents the result of applying bilateral filtering to the ultrasound image, and λ is a preset noise reduction weight. S234. Based on the standard enhanced image and the noise-reducing constrained enhanced image, the region-specific enhanced image is obtained.

5. The method according to claim 4, characterized in that, S3 includes: S31. The region-enhanced image and the acoustic shadow region binary mask are spliced ​​together to generate dual-channel input data. S32. Input the dual-channel input data into the encoder of the SMA-U-Net model, and extract multi-scale encoder feature maps through convolution and downsampling operations; S33. In the encoder skip connection, the encoder feature map of each layer in the SMA-U-Net model is fused with the acoustic shadow region binary mask to generate an attention-weighted feature map; the expression for the feature fusion is: in, This represents the encoder feature map output by the k-th layer encoder of the SMA-U-Net model. UpSample(M) represents the attention-weighted feature map corresponding to the encoder feature map output by the k-th layer encoder. s (x,y) represents the output result after upsampling the binary mask of the acoustic shadow region to the same spatial resolution as the encoder feature map. Indicates to and UpSample(M) s (x,y)) performs channel concatenation operation; ⊙ represents element-wise multiplication; Conv 1×1 (·) represents a 1×1 convolution operation, and σ(·) is the Sigmoid function; S34. Input the attention-weighted feature maps output by each layer encoder into the decoder of the SMA-U-Net model, fuse multi-scale features layer by layer through deconvolution and skip connections, and output the target probability map through the Sigmoid activation function; wherein, the target probability map represents the probability that each pixel in the ultrasound image belongs to the target.

6. The method according to claim 5, characterized in that, S4 includes: S41. Based on a preset probability threshold, the target probability map is binarized to generate an initial target region binary map. S42. Perform Euclidean distance transformation on the binary map of the initial target region to generate a symbolic distance function, and set the symbolic distance function as the initial value of the level set; S43. Construct the level set energy function based on the initial value of the level set; S44. Update the energy function of the level set using the gradient descent method until convergence, and obtain the target segmentation result.

7. The method according to any one of claims 1 to 6, characterized in that, S6 includes: S61. Based on the acoustic shadow region binary mask, extract the set of three-dimensional spatial coordinates of the shadow region; S62. Based on the set of three-dimensional spatial coordinates, apply anisotropic smoothing constraints to the initial deformation field within the shadow region to obtain the modified deformation field; S63. Apply the modified deformation field to the coordinate system of the preoperative image to generate a three-dimensional space for deformation correction; S64. Mark blood vessels and nerve structures as obstacle regions in the deformation correction three-dimensional space, and construct an obstacle distance field based on the obstacle regions; S65. Based on the deformation correction three-dimensional space and the obstacle distance field, the final puncture path is generated using the RRT* algorithm.

8. A precise target localization system for interventional therapy based on ultrasound imaging, characterized in that, The system includes: The image preprocessing module is used to extract high-frequency features and perform morphological processing on the input ultrasound images to generate a binary mask for the acoustic shadow region. The enhancement processing module is used to dynamically adjust the contrast enhancement parameters according to the acoustic shadow region binary mask, and to perform regional enhancement on the ultrasound image according to the contrast enhancement parameters to obtain a regionally enhanced image. The probability map generation module is used to input the enhanced image and the acoustic shadow region binary mask into a pre-trained neural network model to generate a target probability map; The segmentation calculation module is used to perform shape-constrained level set evolution on the target probability map and output the target segmentation result; The 3D reconstruction and registration module is used to reconstruct a 3D target model based on the target segmentation results using the moving cube algorithm; and to register the 3D target model with the preoperative images using the Demons algorithm to obtain the initial deformation field; wherein, the preoperative images include CT images and MRI images; The path correction module is used to apply a shadow region smoothing constraint to the initial deformation field, generate a corrected deformation field, and generate the final puncture path based on the corrected deformation field.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images