Intelligent urological surgery ultrasound image processing system

Abstract

This invention discloses an intelligent urological ultrasound image processing system, belonging to the field of image enhancement technology. The system includes: an ultrasound image acquisition module, a urological ultrasound image processing model construction module, and a urological ultrasound image processing module. Specifically, this invention constructs a dynamic gain compensation coefficient to obtain an ultrasound image after depth attenuation correction. A deep lesion enhancement feature map is generated through multi-scale local feature enhancement blocks. A multi-scale attention map is calculated, and combined with an echo saliency weight map to obtain an echo-weighted multi-scale feature map, improving the completeness of image feature expression. A feature gradient amplitude map is calculated to generate an anatomical edge prior feature map and a lesion multi-scale saliency weight map. Anatomical gating branches and lesion gating branches are constructed to obtain a gated fusion feature map. The enhanced urological ultrasound image is obtained through two sub-pixel convolutions, significantly improving the accuracy of lesion boundary localization and the reliability of image analysis.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement technology, specifically referring to an intelligent urological ultrasound image processing system. Background Technology

[0002] Urological ultrasound image processing systems utilize deep learning technology to enhance urological ultrasound images, improving image quality and lesion identification capabilities. This provides urologists with high-definition, high-fidelity enhanced ultrasound images, helping to accurately locate lesion areas, clarify anatomical boundaries, and develop personalized treatment plans, thereby reducing the risk of missed diagnoses and misdiagnoses. However, existing urological ultrasound image processing systems suffer from the superposition of speckle noise and echo attenuation effects, and the diverse lesion morphologies and scale differences lead to low signal-to-noise ratios, insufficient lesion identification, and difficulty in distinguishing normal tissue from lesion areas, thus affecting subsequent image analysis. Furthermore, existing urological ultrasound image processing systems suffer from blurred organ contours and lesion boundaries, and the difficulty in simultaneously achieving lesion enhancement and structural fidelity. This results in anatomical distortion or weakened lesion details in enhanced images, affecting the accuracy of lesion boundary localization and the reliability of image analysis. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an intelligent urological ultrasound image processing system. Addressing the problems in existing urological ultrasound image processing systems, such as the superposition of speckle noise and echo attenuation effects, and the diverse lesion morphologies and large scale differences leading to low image signal-to-noise ratio, insufficient lesion identification, and difficulty in distinguishing normal tissue from lesion areas, thus affecting subsequent image analysis, this solution combines global image variance estimation of speckle noise intensity, generates denoised ultrasound images through adaptive denoising, constructs dynamic gain compensation coefficients to obtain depth attenuation-corrected ultrasound images, extracts superficial anatomical structure features, generates deep lesion enhancement feature maps through multi-scale local feature enhancement blocks, calculates multi-scale attention maps, and obtains echo-weighted multi-scale feature maps by combining echo saliency weight maps. Finally, echo enhancement feature maps are obtained through channel enhancement and depth convolution, strengthening the... This approach preserves and highlights lesion details and textures, enhancing the integrity of image feature representation and significantly improving the accuracy of subsequent image analysis. Addressing the issues of blurred organ contours and lesion boundaries, and the difficulty in balancing lesion enhancement with structural fidelity in existing urological ultrasound image processing systems, which often result in anatomical distortion or weakened lesion details in enhanced images, thus affecting lesion boundary localization accuracy and image analysis reliability, this solution calculates feature gradient amplitude maps to generate prior anatomical edge feature maps. Based on pooling, it generates multi-scale saliency weight maps for lesions, constructs anatomical and lesion-gated branches to obtain gated fusion feature maps, and generates refined calibration feature maps through multi-scale spatial feature extraction. Finally, it obtains enhanced urological ultrasound images through two sub-pixel convolutions, achieving step-by-step image scale enhancement and significantly improving lesion boundary localization accuracy and image analysis reliability.

[0004] The intelligent urological ultrasound image processing system provided by the present invention includes an ultrasound image acquisition module, a urological ultrasound image processing model construction module, and a urological ultrasound image processing module.

[0005] The ultrasound image acquisition module acquires historical urological ultrasound images and performs preprocessing.

[0006] The urological ultrasound image processing model construction module combines global image variance to estimate speckle noise intensity, generates denoised ultrasound images through adaptive denoising, constructs dynamic gain compensation coefficients to obtain depth attenuation corrected ultrasound images, extracts superficial anatomical structure features, generates deep lesion enhancement feature maps through multi-scale local feature enhancement blocks, calculates multi-scale attention maps, combines echo saliency weight maps to obtain echo-weighted multi-scale feature maps, obtains echo enhancement feature maps through channel enhancement and depth convolution, calculates feature gradient magnitude maps, generates anatomical edge prior feature maps, generates lesion multi-scale saliency weight maps based on pooling, constructs anatomical gated branches and lesion gated branches to obtain gated fusion feature maps, generates refined calibration feature maps through multi-scale spatial feature extraction, and obtains enhanced urological ultrasound images through two sub-pixel convolutions.

[0007] The urological ultrasound image processing module processes real-time urological ultrasound images to obtain enhanced real-time urological ultrasound images.

[0008] Furthermore, the ultrasound image acquisition module acquires historical urological ultrasound images and performs preprocessing, which includes grayscale conversion, size unification, and grayscale normalization.

[0009] Furthermore, the urological ultrasound image processing model construction module is based on deep learning to construct a urological ultrasound image processing model, performs image enhancement processing, and obtains enhanced urological ultrasound images; specifically, it includes the following:

[0010] Local statistical adaptive denoising unit: For the preprocessed urological ultrasound image, statistical analysis is performed on each pixel using a local window, the mean gray value and local variance within the window are calculated, the speckle noise intensity is estimated by combining the global image variance, and adaptive denoising is achieved based on the maximum a posteriori probability weighting to obtain a denoised ultrasound image.

[0011] Depth gain correction unit: For the denoised ultrasound image, a normalized depth coordinate mapping is established along the depth direction of ultrasound propagation, and a dynamic gain compensation coefficient related to depth is constructed to obtain the ultrasound image after depth attenuation correction.

[0012] Deep lesion enhancement unit: For the ultrasound image after depth attenuation correction, convolution is used to perform shallow feature mapping. The feature distribution is stabilized by group normalization to obtain the shallow anatomical structure feature map. The shallow anatomical structure feature map is input into N sequentially stacked multi-scale local feature enhancement blocks to generate the deep lesion enhancement feature map.

[0013] Echo-weighted multi-scale attention unit: Layer normalization is performed on the enhancement feature map of deep lesions, query, key, and value feature maps are generated at three local receptive field scales, multi-scale attention map is calculated using matrix multiplication, echo saliency weight map is calculated based on the global mean and local feature differences of the enhancement feature map of deep lesions, multi-scale attention map and echo saliency weight map are weighted and fused, and learnable relative position bias is added. After normalization, echo-weighted multi-scale feature map is obtained.

[0014] Echo feature enhancement unit: The echo weighted multi-scale feature map is sequentially processed by Ghost Head channel enhancement and deep convolutional feedforward network, and finally the echo enhanced feature map is obtained through residual connection;

[0015] Gradient edge prior unit: Calculate the partial derivatives of the echo enhancement feature map in the horizontal and vertical directions respectively to obtain gradient feature maps in two directions. Based on the gradient feature maps in two directions, further calculate the feature gradient magnitude map. Concatenate the feature gradient magnitude map and the echo enhancement feature map in the channel dimension. Perform feature fusion and channel compression through convolution and normalize to the [0, 1] interval to generate the anatomical edge prior feature map.

[0016] A dual-gated dynamic fusion unit is used to perform global average pooling and global max pooling operations on the echo enhancement feature map. The two pooling results are concatenated along the channel dimension. A multi-scale saliency weight map of the lesion is obtained through convolution and activation. Anatomical gating branches and lesion gating branches are constructed using the prior feature map of the anatomical edge and the multi-scale saliency weight map of the lesion, respectively. Dynamic weighted gating fusion is performed on the echo enhancement feature map and the superficial anatomical structure feature map to obtain the gated fusion feature map.

[0017] Multi-scale fine calibration unit: The gated fusion feature map is extracted using pointwise convolution, standard convolution, and depthwise convolution. The three multi-scale spatial features are then summed element-wise with weights and then activated to obtain the fine calibration feature map.

[0018] The dual-level subpixel reconstruction unit performs subpixel convolution on the refined calibration feature map to obtain a primary high-resolution feature map. It then performs spatial weighting using the anatomical edge prior feature map, and further enhances local details through convolution. It obtains an intermediate high-resolution feature map through residual fusion and performs subpixel convolution. Finally, it completes channel mapping and grayscale reconstruction through pointwise convolution, and outputs an enhanced urological ultrasound image.

[0019] The three-constraint joint optimization unit constructs composite constraints from three dimensions: global grayscale self-consistency, local smoothness of lesion region, and anatomical structure gradient fidelity. After weighted summation of each loss term to obtain the total loss of the three constraints, gradient backpropagation is performed through the Adam optimizer to save the optimal parameters.

[0020] Furthermore, the urological ultrasound image processing module acquires real-time urological ultrasound images, preprocesses them, and then inputs them into a urological ultrasound image processing model constructed based on optimal parameters for further processing to obtain enhanced real-time urological ultrasound images.

[0021] The beneficial effects achieved by the present invention using the above solution are as follows:

[0022] (1) To address the problems in existing urological ultrasound image processing systems, such as the superposition of speckle noise and echo attenuation effects, and the diverse lesion morphology and large scale differences, which lead to low image signal-to-noise ratio, insufficient lesion identification, and difficulty in distinguishing normal tissue from lesion areas, thus affecting subsequent image analysis, this solution combines global image variance estimation of speckle noise intensity and generates denoised ultrasound images through adaptive denoising. While removing noise, it preserves the lesion and organ edge structures, thereby improving the image signal-to-noise ratio. A dynamic gain compensation coefficient is constructed to obtain ultrasound images after depth attenuation correction. The system balances the echo intensity across the entire field, restoring details of deep lesions; it extracts superficial anatomical features and generates enhanced feature maps of deep lesions through multi-scale local feature enhancement blocks, significantly improving lesion identification; it calculates multi-scale attention maps and combines them with echo saliency weight maps to obtain echo-weighted multi-scale feature maps, accurately locating lesions and improving the distinction between lesions and normal tissue; and it obtains echo enhancement feature maps through channel enhancement and depth convolution, strengthening lesion edges and fine anatomical textures, preserving and highlighting lesion details, improving the integrity of image feature expression, and significantly enhancing the accuracy of subsequent image analysis.

[0023] (2) To address the problems of blurred organ contours and lesion boundaries, and difficulty in balancing lesion enhancement and structural fidelity in existing urological ultrasound image processing systems, which lead to anatomical distortion or weakened lesion details in enhanced images, thus affecting the accuracy of lesion boundary localization and the reliability of image analysis, this solution calculates feature gradient amplitude maps to generate prior anatomical edge feature maps, highlighting organ contours and lesion boundary location information, and improving the accuracy of lesion boundary localization; it generates multi-scale saliency weight maps of lesions based on pooling, strengthens the feature differences of lesion regions, and accurately identifies saliency regions of lesions; it constructs anatomical gated branches and lesion gated branches to obtain gated fusion feature maps, dynamically balancing lesion enhancement and anatomical structure fidelity, and avoiding image distortion; it generates refined calibration feature maps through multi-scale spatial feature extraction, adaptively balancing edge preservation and artifact suppression, eliminating image distortion, and improving the refinement of feature maps; it obtains enhanced urological ultrasound images through two sub-pixel convolutions, realizing image scale enhancement step by step, and significantly improving the accuracy of lesion boundary localization and the reliability of image analysis. Attached Figure Description

[0024] Figure 1A schematic diagram of the intelligent urological ultrasound image processing system provided by the present invention;

[0025] Figure 2 A schematic diagram of the module for constructing a urological ultrasound image processing model.

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0029] Example 1, see Figure 1 The intelligent urological ultrasound image processing system provided by the present invention includes an ultrasound image acquisition module, a urological ultrasound image processing model construction module, and a urological ultrasound image processing module.

[0030] The ultrasound image acquisition module acquires historical urological ultrasound images, performs preprocessing, and sends the data to the urological ultrasound image processing model construction module.

[0031] The urological ultrasound image processing model construction module combines global image variance to estimate speckle noise intensity, generates denoised ultrasound images through adaptive denoising, constructs dynamic gain compensation coefficients, obtains ultrasound images after depth attenuation correction, extracts superficial anatomical structure features, generates deep lesion enhancement feature maps through multi-scale local feature enhancement blocks, calculates multi-scale attention maps, obtains echo-weighted multi-scale feature maps by combining echo saliency weight maps, obtains echo enhancement feature maps through channel enhancement and depth convolution, calculates feature gradient magnitude maps, generates anatomical edge prior feature maps, generates lesion multi-scale saliency weight maps based on pooling, constructs anatomical gated branches and lesion gated branches, obtains gated fusion feature maps, generates refined calibration feature maps through multi-scale spatial feature extraction, obtains enhanced urological ultrasound images through two sub-pixel convolutions, and sends the data to the urological ultrasound image processing module.

[0032] The urological ultrasound image processing module processes real-time urological ultrasound images to obtain enhanced real-time urological ultrasound images.

[0033] Example 2, see Figure 1 This embodiment is based on the above embodiment, wherein the ultrasound image acquisition module acquires historical urological ultrasound images and performs preprocessing.

[0034] The preprocessing includes grayscale conversion, size unification, and grayscale normalization;

[0035] The grayscale conversion refers to converting the acquired image into a grayscale image;

[0036] The size unification refers to adjusting the image size to 512×512 using bilinear interpolation.

[0037] The grayscale normalization is to linearly scale the grayscale values ​​of the grayscale image to the range of [0, 1].

[0038] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. The urological ultrasound image processing model construction module is based on deep learning to construct a urological ultrasound image processing model, performs image enhancement processing, and obtains an enhanced urological ultrasound image; specifically, it includes the following:

[0039] Local statistical adaptive denoising unit: The inherent speckle noise in urological ultrasound images can obscure minute lesions and capsule details of organs such as the prostate, bladder, and kidneys. Conventional denoising easily blurs edges and lesion areas. For preprocessed urological ultrasound images, a fixed-size local window is used for statistical analysis of each pixel. The mean gray level and local variance within the window are calculated, and the speckle noise intensity is estimated by combining it with the global image variance. The denoising intensity is increased in flat areas with small gray level changes and decreased in edge and lesion areas. Adaptive denoising is achieved based on maximum a posteriori probability weighting, resulting in a denoised ultrasound image that balances noise removal and structural fidelity, improving the image signal-to-noise ratio. The formulas used are as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] In the formula, It is based on pixels Let i be the average grayscale value within the local window centered on the center, and let i and j be the row and column coordinates, respectively. Let M and N be the height and width of the local window, respectively. It is based on pixels A 7x7 local sliding statistical window centered on the data. These are the coordinates of all pixels within the window. and These are the preprocessed images in and The grayscale value of the location, It is the variance of speckle noise within a local window. It is the global speckle noise variance of the entire ultrasound image. It is a denoised ultrasound image in The pixel value of the location;

[0044] Depth gain correction unit: Ultrasound attenuates at depth during propagation in biological tissues, leading to weak echoes, blurred lesions, and loss of detail in deep regions of urological images, while near-field echoes are excessively strong. For denoising ultrasound images, a normalized depth coordinate mapping is established along the depth direction of ultrasound propagation. Based on the physical law of exponential attenuation of ultrasound propagation in biological tissues, a depth-related dynamic gain compensation coefficient is constructed. Higher compensation weights are assigned to deep regions of the image, while moderate gain suppression is applied to the near-field region, making the echo intensity of the entire tissue more uniform. This eliminates the problems of blurred lesions and loss of detail caused by echo attenuation in deep tissues, thus obtaining a depth-attenuation-corrected ultrasound image. The formula used is as follows:

[0045] ;

[0046] ;

[0047] ;

[0048] In the formula, and They are pixels The corresponding normalized depth coordinates and dynamic gain compensation coefficients, where W is the total horizontal width of the ultrasound image after unification, W=512. It is the initial gain compensation coefficient. λ is the exponential decay adjustment factor. , The ultrasound image after depth attenuation correction is at the pixel level. The grayscale value at that location;

[0049] Deep lesion enhancement unit: Urological lesions are characterized by weak and varying scales, making it difficult for conventional convolution to capture multi-scale fine lesions and anatomical features. For ultrasound images corrected for depth attenuation, a 3×3 convolution is used for shallow feature mapping. Group normalization stabilizes the feature distribution, resulting in a shallow anatomical feature map. This shallow anatomical feature map is then input into N stacked multi-scale local feature enhancement blocks. Each block undergoes multi-scale neighborhood self-attention calculation at different receptive field scales, progressively enhancing lesion and anatomical features to generate a deep lesion enhancement feature map. Adaptive feature extraction across multiple receptive fields focuses on lesions of different sizes, significantly improving the identification of key features such as tumor boundaries, strong echogenicity of stones, and anechoic cysts. The formulas used are as follows:

[0050] ;

[0051] ;

[0052] In the formula, F S It is a diagram of superficial anatomical structures, I ac This is a depth-attenuation corrected ultrasound image. It is a standard 3×3 convolution operation. It is a group normalization, F a and F a-1 These are the output and input features of the a-th multi-scale local feature enhancement block, respectively, where a is the block index. When a=1, the input is F. S When a=N, the output is the enhancement feature map F of deep lesions. D ; ; It is layer normalization. It is the multi-scale self-attention of the a-th block. X is a general intermediate feature placeholder, Q, K, and V are the query, key, and value feature matrices, respectively, T is the transpose, B is the learnable relative position bias matrix, and C is the number of channels for the input features. It is a normalized exponential function. It is the Ghost Head channel enhancement of the a-th block. W1, W2, and W3 are 1×1 convolutional learnable weight matrices. It is Hadamaji. It is a depthwise convolutional feedforward network for the a-th block. W4 is a 1×1 convolutional learnable weight matrix. It is a linear rectified activation function. It is a 5×5 depthwise convolution operation;

[0053] Echo-weighted multi-scale attention unit: In urological ultrasound, hyperechoic and hypoechoic regions are highly correlated with lesions. Traditional attention methods do not incorporate echo characteristics, easily overlooking salient lesion regions. This method performs layer normalization on the enhancement feature map of deep lesions to stabilize feature distribution. Query, key, and value feature maps are generated at three local receptive field scales: 3×3, 5×5, and 7×7. A multi-scale attention map is calculated using matrix multiplication. An echo saliency weight map is calculated based on the global mean and local feature differences of the deep lesion enhancement feature map, assigning higher weights to hyperechoic and hypoechoic regions. The multi-scale attention map and the echo saliency weight map are weighted and fused, with a learnable relative position bias added. After normalization, an echo-weighted multi-scale feature map is obtained, accurately locating lesions, reducing interference from normal tissue, and improving the effectiveness of lesion feature expression. The formulas used are as follows:

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] In the formula, X norm Enhancement feature map F of deep lesions D The standardized feature map obtained after normalization at the execution layer, where s is the multi-scale receptive field index. , and These are the methods used to measure X at the s-th scale. norm The projection is a 1×1 convolutional learnable weight matrix of query, key, and value features, Q. s K s and V s These are the values ​​of X at the s-th scale. norm The query, key, and value feature matrix obtained by linear projection, where T is the transpose. A is the learnable relative position bias matrix at the s-th scale. s This is the neighborhood attention response map at the s-th scale, A multi It is a multi-scale attention map. It is X norm exist The pixel value of the location, μ X and σ X They are X norm The global mean and global standard deviation, It is a smoothing term. , Is The echo significance weight of a location, E, is calculated from all locations. The constructed two-dimensional echo significance weighted map, A weight It is an echo-weighted attention map, C D It is F D Number of feature channels, V multi Value features V at three scales s The multi-scale aggregated value feature obtained after fusion, F attn It is an echo-weighted multi-scale feature map;

[0062] Echo feature enhancement unit: Due to weak features in lesion edges, organ capsules, and fine textures, low channel feature utilization, and easy loss of details during network transmission, the echo-weighted multi-scale feature map is sequentially processed by Ghost Head channel enhancement and deep convolutional feedforward network to enhance lesion edges, capsules, and fine textures. Finally, the echo-enhanced feature map is obtained through residual connections. The formula used is as follows:

[0063] ;

[0064] ;

[0065] ;

[0066] In the formula, F gh This is the feature map after Ghost Head channel enhancement, F ffn It is the feature map after processing by a deep convolutional feedforward network, F df It is an echo enhancement feature map.

[0067] By performing the above operations, this solution addresses the problems in existing urological ultrasound image processing systems where speckle noise and echo attenuation effects overlap, and lesions exhibit diverse morphologies and large scale differences, leading to low image signal-to-noise ratio, insufficient lesion identification, and difficulty in distinguishing normal tissue from lesion areas, thus affecting subsequent image analysis. This solution combines global image variance estimation of speckle noise intensity and generates denoised ultrasound images through adaptive denoising, preserving lesion and organ edge structures while removing noise, thereby improving the image signal-to-noise ratio. A dynamic gain compensation coefficient is constructed to obtain ultrasound images after depth attenuation correction. The image is processed by equalizing the echo intensity of the entire tissue field to restore the details of deep lesions. Superficial anatomical features are extracted and multi-scale local feature enhancement blocks are used to generate deep lesion enhancement feature maps, significantly improving lesion identification. Multi-scale attention maps are calculated and combined with echo saliency weight maps to obtain echo-weighted multi-scale feature maps, accurately locating lesions and improving the distinction between lesions and normal tissues. Echo enhancement feature maps are obtained through channel enhancement and depth convolution, which strengthen the edges and fine anatomical textures of lesions, preserve and highlight lesion details, improve the integrity of image feature expression, and significantly improve the accuracy of subsequent image analysis.

[0068] Example 4, see Figure 1 and Figure 2 This embodiment is based on the above embodiment, and the urological ultrasound image processing model construction module further includes the following:

[0069] Gradient edge prior unit: Due to blurred organ contours and lesion boundaries, model reconstruction is prone to structural misalignment and false edges. Partial derivatives in the horizontal and vertical directions are calculated for the echo enhancement feature map to obtain gradient feature maps in two directions. Based on these two gradient feature maps, feature gradient magnitude maps are further calculated to highlight the positional information of organ contours, capsules, and lesion boundaries. The feature gradient magnitude maps and echo enhancement feature maps are concatenated along the channel dimension, and feature fusion and channel compression are performed through 3×3 convolution. Then, the maps are normalized to the [0, 1] interval using the Sigmoid activation function to generate continuous, smooth anatomical edge prior feature maps with anatomical structural constraints. This accurately locates organ contours and lesion boundaries, providing anatomical structural guidance for subsequent fusion and reconstruction, suppressing unnatural artifacts, and ensuring structural rationality. The formulas used are as follows:

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] In the formula, and They are F df Gradient feature maps in the horizontal x and vertical y directions, and They are respectively for F df Perform spatial gradient calculations in the horizontal and vertical directions. It is F df The corresponding feature gradient magnitude map, It's a splicing operation, F cat It is a fusion of feature maps. It is the Sigmoid non-linear activation function, M ana It is a priori feature map of the anatomical edge;

[0075] A dual-gated dynamic fusion unit is used. Simple lesion enhancement can easily damage anatomical structures, while simple structural fidelity weakens lesion details, making it difficult to balance lesion enhancement and anatomical integrity. Global average pooling and global max pooling operations are performed on the echo enhancement feature map, and the results are concatenated along the channel dimension to enhance the feature differences in the lesion region. A multi-scale saliency weight map of the lesion is obtained through 1×1 convolution and sigmoid activation. Anatomical gating branches and lesion gating branches are constructed using the prior feature map of the anatomical edge and the multi-scale saliency weight map of the lesion, respectively. Dynamic weighted gating fusion is performed on the echo enhancement feature map and the superficial anatomical structure feature map. In the lesion region, lesion gating dominates to enhance high-resolution detail representation; in the anatomical edge region, anatomical gating dominates to maintain structural contour integrity; in the normal region, balanced weighting is used to obtain a gated fusion feature map that balances lesion enhancement and anatomical structure fidelity. The formulas used are as follows:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] In the formula, It is a global average pooling operation. It is a global max pooling operation, F pool It is the result of pooling feature concatenation. It is a 1×1 pointwise convolution operation, M L It is a multi-scale significance weighted map of lesions, g L It is the dynamic gating weight of the lesion, g A It is the anatomical dynamic gating weight, F fus It is a gated fusion feature map;

[0081] A multi-scale fine-tuning calibration unit is used. The gated fusion feature map is extracted using 1×1 pointwise convolution, 3×3 standard convolution, and 5×5 depthwise convolution. Anatomical edge prior feature maps are used to weight the output of large receptive fields by structural regions, and the output of small receptive fields by unstructured regions, achieving an adaptive balance between edge preservation and artifact suppression. The three multi-scale spatial features are element-wise weighted and summed, then nonlinearly smoothed using the GELU activation function to suppress block artifacts, checkerboard artifacts, and unnatural texture distortions that easily occur in Transformer reconstruction, resulting in a fine-tuned calibration feature map that conforms to the physical imaging principles of ultrasound and is free of false structures. The formulas used are as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] In the formula, F conv1 It is the channel calibration feature map obtained by pointwise convolution, F conv3 It is a locally smoothed feature map obtained by standard convolution, F conv5 It is the global structural feature map obtained by depthwise convolution. It is the activation function of the Gaussian error linear unit, F ref It is a refined calibration feature map;

[0087] A dual-level subpixel reconstruction unit addresses the issue of significant detail loss in traditional upsampling methods, which result in blurred reconstructions of lesion edges and fine structures, failing to meet the demands of high-resolution clinical observation. This unit performs subpixel convolution on the refined calibration feature map to obtain a primary high-resolution feature map, achieving coarse scale magnification. Spatial weighting is then applied using prior anatomical edge feature maps to highlight organ capsules and lesion edge structures. Local detail enhancement is further achieved through 3×3 convolution. Residual fusion yields a mid-level high-resolution feature map, which is then subjected to subpixel convolution for fine scale magnification. Finally, 1×1 pointwise convolution completes channel mapping and grayscale reconstruction, outputting an enhanced urological ultrasound image. The coarse and fine magnification steps achieve scale enhancement, while edge weighting strengthens capsule and lesion boundaries, resulting in a high-definition ultrasound image that meets the detailed requirements of precise urological diagnosis. The formulas used are as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, F up1 It is a primary high-resolution feature map. It is a subpixel convolution operation, F edge It is an anatomical edge enhancement feature map, F mid It is a medium-to-high resolution feature map, F up2 It is the final high-resolution feature map, I HR This is an enhanced urological ultrasound image;

[0094] A three-constraint joint optimization unit is used because a single loss function can easily lead to image distortion, uneven lesion regions, and loss of anatomical gradients, resulting in unstable model convergence. A composite constraint is constructed from three dimensions: global gray-level self-consistency, local smoothness of lesion regions, and fidelity of anatomical structure gradients. After weighted summation of each loss term to obtain the total loss of the three constraints, gradient backpropagation is performed through the Adam optimizer to iteratively update all learnable parameters in the model. The optimization unit is activated when the number of model training iterations reaches the preset maximum of 100 or the absolute value of the change in the total loss of the three constraints is less than [a certain value] for three consecutive iterations. When the iteration stops and the current optimal parameters are saved, global grayscale consistency is guaranteed, local smoothing is guaranteed in lesion areas, and gradient fidelity is guaranteed in anatomical areas. Weighted loss enables the model to converge quickly and stably, enhancing the results to better match the characteristics of real clinical ultrasound images. The formulas used are as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] In the formula, , , and These are the global grayscale self-consistency loss, the lesion region local smoothing loss, the anatomical structure gradient fidelity loss, and the total loss under three constraints. pre These are pre-processed urological ultrasound images. It is an L1 norm operation. It is an enhanced image I HR The second-order Laplace gradient plot, and They are I HR and I pre Spatial gradient magnitude map, and These are the sets of learnable parameters of the model after the (t+1)th and tth gradient backpropagation updates, respectively. It's the learning rate. m t and v t These are the first-order moment estimate and the second-order moment estimate of the gradient, respectively.

[0101] By performing the above operations, this solution addresses the problems of existing urological ultrasound image processing systems, such as blurred organ contours and lesion boundaries, and the difficulty in balancing lesion enhancement and structural fidelity. These issues lead to anatomical distortion or weakened lesion details in enhanced images, thus affecting the accuracy of lesion boundary localization and the reliability of image analysis. This solution calculates feature gradient amplitude maps to generate prior anatomical edge feature maps, highlighting organ contours and lesion boundary location information, thereby improving the accuracy of lesion boundary localization. Based on pooling, a multi-scale saliency weight map of the lesion is generated to enhance the feature differences in lesion regions and accurately identify saliency areas. Anatomical and lesion-gated branches are constructed to obtain a gated fusion feature map, dynamically balancing lesion enhancement and anatomical structure fidelity to avoid image distortion. A refined calibration feature map is generated through multi-scale spatial feature extraction, adaptively balancing edge preservation and artifact suppression to eliminate image distortion and improve the refinement of the feature map. Finally, the enhanced urological ultrasound image is obtained through two sub-pixel convolutions, achieving image scale enhancement step by step, significantly improving the accuracy of lesion boundary localization and the reliability of image analysis.

[0102] Example 5, see Figure 1 This embodiment is based on the above embodiment. The urological ultrasound image processing module acquires real-time urological ultrasound images, preprocesses them, and then inputs them into a urological ultrasound image processing model constructed based on optimal parameters for processing to obtain enhanced real-time urological ultrasound images.

[0103] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0105] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent urological ultrasound image processing system, characterized in that: It includes an ultrasound image acquisition module, a urological ultrasound image processing model construction module, and a urological ultrasound image processing module; The ultrasound image acquisition module acquires historical urological ultrasound images and performs preprocessing. The urological ultrasound image processing model construction module combines global image variance to estimate speckle noise intensity, generates denoised ultrasound images through adaptive denoising, constructs dynamic gain compensation coefficients to obtain depth attenuation corrected ultrasound images, extracts superficial anatomical structure features, generates deep lesion enhancement feature maps through multi-scale local feature enhancement blocks, calculates multi-scale attention maps, combines echo saliency weight maps to obtain echo-weighted multi-scale feature maps, obtains echo enhancement feature maps through channel enhancement and depth convolution, calculates feature gradient magnitude maps, generates anatomical edge prior feature maps, generates lesion multi-scale saliency weight maps based on pooling, constructs anatomical gated branches and lesion gated branches to obtain gated fusion feature maps, generates refined calibration feature maps through multi-scale spatial feature extraction, and obtains enhanced urological ultrasound images through two sub-pixel convolutions. The urological ultrasound image processing module processes real-time urological ultrasound images to obtain enhanced real-time urological ultrasound images.

2. The intelligent urological ultrasound image processing system according to claim 1, characterized in that: The urological ultrasound image processing model construction module is based on deep learning to build a urological ultrasound image processing model, performs image enhancement processing, and obtains enhanced urological ultrasound images; specifically, it includes the following: Local statistical adaptive denoising unit: For the preprocessed urological ultrasound image, statistical analysis is performed on each pixel using a local window, the mean gray value and local variance within the window are calculated, the speckle noise intensity is estimated by combining the global image variance, and adaptive denoising is achieved based on the maximum a posteriori probability weighting to obtain a denoised ultrasound image. Depth gain correction unit; Deep lesion enhancement unit; Echo-weighted multiscale attention unit; Echo feature enhancement unit: The echo weighted multi-scale feature map is sequentially processed by Ghost Head channel enhancement and deep convolutional feedforward network, and finally the echo enhanced feature map is obtained through residual connection; Gradient edge prior unit; Dual-gated dynamic fusion unit; Multi-scale fine calibration unit: The gated fusion feature map is extracted using pointwise convolution, standard convolution, and depthwise convolution. The three multi-scale spatial features are then summed element-wise with weights and then activated to obtain the fine calibration feature map. Two-order subpixel reconstruction unit; The three-constraint joint optimization unit constructs composite constraints from three dimensions: global grayscale self-consistency, local smoothness of lesion region, and anatomical structure gradient fidelity. After weighted summation of each loss term to obtain the total loss of the three constraints, gradient backpropagation is performed through the Adam optimizer to save the optimal parameters.

3. The intelligent urological ultrasound image processing system according to claim 2, characterized in that: The depth gain correction unit establishes a normalized depth coordinate mapping along the depth direction of ultrasound propagation in the denoised ultrasound image, and constructs a dynamic gain compensation coefficient related to depth, thereby obtaining an ultrasound image after depth attenuation correction.

4. The intelligent urological ultrasound image processing system according to claim 3, characterized in that: The deep lesion enhancement unit is a shallow feature map obtained by using convolution to map shallow features in the ultrasound image after depth attenuation correction and stabilizing the feature distribution through group normalization. The shallow anatomical structure feature map is then input into a multi-scale local feature enhancement block that is stacked sequentially to generate a deep lesion enhancement feature map.

5. The intelligent urological ultrasound image processing system according to claim 4, characterized in that: The echo-weighted multi-scale attention unit performs layer normalization on the deep lesion enhancement feature map, generates query, key, and value feature maps at three local receptive field scales, calculates the multi-scale attention map using matrix multiplication, calculates the echo saliency weight map based on the global mean and local feature differences of the deep lesion enhancement feature map, and weights and fuses the multi-scale attention map and the echo saliency weight map, adding a learnable relative position bias, and obtains the echo-weighted multi-scale feature map after normalization.

6. The intelligent urological ultrasound image processing system according to claim 5, characterized in that: The gradient edge prior unit calculates the partial derivatives of the echo enhancement feature map in the horizontal and vertical directions to obtain gradient feature maps in two directions. Based on the gradient feature maps in two directions, the feature gradient magnitude map is further calculated. The feature gradient magnitude map and the echo enhancement feature map are concatenated in the channel dimension. Feature fusion and channel compression are performed through convolution, and the data is normalized to the [0, 1] interval to generate the anatomical edge prior feature map.

7. The intelligent urological ultrasound image processing system according to claim 6, characterized in that: The dual-gated dynamic fusion unit performs global average pooling and global max pooling operations on the echo enhancement feature map, concatenates the two pooling results in the channel dimension, and obtains a lesion multi-scale saliency weight map through convolution and activation. It then constructs anatomical gating branches and lesion gating branches using the anatomical edge prior feature map and the lesion multi-scale saliency weight map, respectively, and performs dynamic weighted gating fusion on the echo enhancement feature map and the superficial anatomical structure feature map to obtain a gated fusion feature map.

8. The intelligent urological ultrasound image processing system according to claim 7, characterized in that: The dual-level subpixel reconstruction unit performs subpixel convolution on the refined calibration feature map to obtain a primary high-resolution feature map. It then performs spatial weighting using the anatomical edge prior feature map, followed by local detail enhancement through convolution. Finally, it obtains a mid-level high-resolution feature map through residual fusion and performs subpixel convolution. Through pointwise convolution, it completes channel mapping and grayscale reconstruction, and outputs an enhanced urological ultrasound image.

9. The intelligent urological ultrasound image processing system according to claim 8, characterized in that: The ultrasound image acquisition module acquires historical urological ultrasound images and performs preprocessing, which includes grayscale conversion, size unification, and grayscale normalization.

10. The intelligent urological ultrasound image processing system according to claim 9, characterized in that: The urological ultrasound image processing module acquires real-time urological ultrasound images, preprocesses them, and then inputs them into a urological ultrasound image processing model constructed based on optimal parameters for further processing, resulting in enhanced real-time urological ultrasound images.

Images