Ultrasonic tumor boundary accurate segmentation method based on sparse representation

Abstract

The invention discloses an ultrasonic tumor boundary accurate segmentation method based on sparse representation. The method comprises the following steps: S1, obtaining a preprocessed ultrasonic tumor image; s2, obtaining a finally optimized multi-scale feature dictionary; s3, performing local block segmentation on the preprocessed ultrasonic tumor image by using the finally optimized multi-scale feature dictionary, and performing sparse coding solution on each local block to obtain a corresponding multi-scale sparse representation coefficient; s4, forming a preliminary ultrasonic tumor boundary diagram based on the multi-scale sparse representation coefficient and the regularization optimization result of the diagram; s5, further correcting and optimizing the tumor boundary by adopting a structure continuity and smoothness optimization method; and S6, fusing the optimized ultrasonic tumor boundary segmentation result with the preprocessed ultrasonic tumor image to form a final ultrasonic tumor boundary segmentation image. According to the method, the capability of extracting the real tumor edge in the low-contrast image is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of tumor imaging technology, and in particular to a method for precise segmentation of ultrasound tumor boundaries based on sparse representation. Background Technology

[0002] With the continuous development of medical imaging technology, ultrasound imaging has been widely used in tumor screening and clinical diagnosis due to its non-invasiveness, safety and low cost. In particular, in the detection of tumors in the liver, breast and thyroid, ultrasound images are an important basis for doctors to make preliminary judgments and further treatment decisions. However, ultrasound images themselves have inherent defects such as low resolution, poor contrast, blurred boundaries and serious speckle noise, which makes the accurate identification of tumor areas, especially the boundaries, extremely challenging.

[0003] Currently, most methods for tumor boundary segmentation in ultrasound images still rely on human experience or traditional image processing techniques, such as thresholding, edge detection, active contour, or region growing methods. While effective under specific conditions, these methods generally suffer from several significant problems: First, traditional edge detection algorithms are prone to producing broken or false edges when faced with strong speckle noise and blurred boundaries in ultrasound images; second, active contour methods are sensitive to the initial contour and struggle to adapt to complex scenarios with varying tumor shapes and irregular boundaries; and third, most traditional methods lack effective modeling of local structures and contextual information in the image, which can easily lead to boundary contour shifts or even missed detections.

[0004] In recent years, deep learning and sparse representation data-driven technologies have been gradually introduced into medical image segmentation. Although some progress has been made in performance, existing methods often ignore the structural continuity and local complexity variations of tumor boundary regions in images, resulting in low segmentation accuracy in areas with blurred boundaries or tissue transitions, which is difficult to meet the actual needs of precision medicine. Therefore, there is an urgent need for a new segmentation method that can fully preserve local structural features and improve the stability and accuracy of boundary detection to overcome the limitations of existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a precise segmentation method for ultrasound tumor boundaries based on sparse representation, which effectively improves the ability to extract the true edges of tumors in low-contrast images.

[0006] A method for precise segmentation of ultrasound tumor boundaries based on sparse representation according to an embodiment of the present invention includes the following steps:

[0007] S1. Perform noise suppression, contrast enhancement, and speckle noise removal on the input raw ultrasound tumor image to obtain a preprocessed ultrasound tumor image;

[0008] S2. Based on the preprocessed ultrasound tumor images and the corresponding labeled ultrasound tumor images, a feature dictionary describing the local boundary features of ultrasound tumors is constructed using a sparse dictionary learning method. Graph regularization constraints are introduced during dictionary training to preserve the local structural relationships between adjacent ultrasound tumor image regions, thus obtaining the final optimized multi-scale feature dictionary.

[0009] S3. The preprocessed ultrasound tumor image is segmented into local blocks using the final optimized multi-scale feature dictionary. Sparse coding is performed on each local block to obtain the corresponding multi-scale sparse representation coefficients.

[0010] S4. Based on the multi-scale sparse representation coefficients and their graph regularization optimization results, edge detection algorithms and morphological processing methods are used to extract preliminary boundary information of ultrasound tumors and form a preliminary ultrasound tumor boundary map.

[0011] S5. Based on the preliminary ultrasound tumor boundary map and the local structural information retained in the graph regularization model, the tumor boundary is further corrected and optimized using structural continuity and smoothness optimization methods;

[0012] S6. The optimized ultrasound tumor boundary segmentation results are fused with the preprocessed ultrasound tumor image to form the final ultrasound tumor boundary segmentation image.

[0013] Optionally, S1 includes the following steps:

[0014] S11. Input ultrasound tumor image set for acquiring ultrasound tumor images ,in, Indicates the first Original ultrasound tumor image, For each original ultrasound tumor image, the total number of ultrasound tumor images is [number]. Gaussian filtering is performed to suppress background random noise using a two-dimensional Gaussian kernel function. Convolution smoothing is performed to obtain a noise-suppressed ultrasound tumor image;

[0015] S12. Contrast enhancement is performed on the noise-suppressed ultrasound tumor image to enhance local contrast and obtain an enhanced ultrasound tumor image. Then, speckle noise removal is performed to smooth local texture information while preserving edge structure, resulting in a preprocessed ultrasound tumor image. This results in the final preprocessed ultrasound tumor image set. .

[0016] Optionally, S2 includes the following steps:

[0017] S21. Based on the preprocessed ultrasound tumor image set and corresponding labeled ultrasound tumor image set Extracting tumor tumor images of size [size] from each ultrasound tumor image using a sliding window method. Pixel-level ultrasound tumor image patches are used to construct ultrasound tumor image patch sets at multiple scales. ,in, This indicates the scale of the ultrasound tumor image patch. Representing scale Next A block of ultrasound tumor images, Representing scale Total number of tumor images on ultrasound;

[0018] S22. For each ultrasound tumor image patch at a given scale, construct an initial sparse feature dictionary. ,in Ultrasound tumor image block Dimensions after flattening The dictionary contains the number of atoms at this scale, used to represent the feature composition of the boundary ultrasound tumor image structure at this scale, and is based on the set of ultrasound tumor image patches. Based on the correspondence between tumor boundaries in labeled ultrasound tumor images, a sparse representation model with graph regularization constraints is constructed:

[0019]

[0020] in, Ultrasound tumor image block In sparse feature dictionary The sparse representation coefficients under the following conditions Representing scale The set of sparse coefficients for all ultrasound tumor image patches. For sparsity control parameters, For graph regularization coefficients, A set of edges that exist between ultrasound tumor image patches, indicating boundary continuity or spatial adjacency. Ultrasound tumor image block With another ultrasound tumor image block Graph regularization structure weights in labeled boundary structures;

[0021] S23. Regularize the weights of the graph structure. A dynamic region sensing mechanism is introduced, which dynamically adjusts the influence of graph constraints based on the structural complexity of each ultrasound tumor image block, and defines dynamic graph regularization structure weights. :

[0022]

[0023] in, Represents ultrasound tumor image blocks The structural complexity factor, when the structural complexity factor The larger the value, the more complex and drastic the tumor boundary in the region, and the stronger the regularization response.

[0024] S24. Jointly optimize the sparse feature dictionary using an alternating minimization strategy. and sparse coefficient set At every scale The process continues until the objective function converges, yielding the final optimized multi-scale feature dictionary. ;

[0025] S25. Weighted fusion of sparse representation results at different scales is performed. By fusing representations at different scales, the ability to characterize tumor boundary contours, structural aberrations, and blurred boundary regions in ultrasound tumor images is optimized, resulting in the final representation of ultrasound tumor image blocks.

[0026]

[0027] in, Indicates the final fusion of the first A block of ultrasound tumor images, For each scale in ultrasound tumor image patches Adaptive fusion weights.

[0028] Optionally, S3 includes the following steps:

[0029] S31. Based on the preprocessed ultrasound tumor image set, each ultrasound tumor image is divided into several local ultrasound tumor image blocks to construct a set of local ultrasound tumor image blocks. ,in, Representing ultrasound tumor images In scale The next A local ultrasound tumor image patch The number of ultrasound tumor image blocks obtained by segmenting the ultrasound tumor image.

[0030] S32. For each local ultrasound tumor image patch Using the final optimized multi-scale feature dictionary Sparse representation encoding is performed while maintaining consistency in feature dimensions:

[0031]

[0032] in, Local ultrasound tumor image patch The final optimized multi-scale feature dictionary The sparse representation coefficients are used to characterize the ultrasound tumor boundary structure information of the local region at the current scale.

[0033] S33. All scales of ultrasound tumor images Sparse representation coefficients under By fusing the coefficients, a multi-scale sparse representation coefficient can be constructed. :

[0034]

[0035] in, Indicates the fused first Multiscale sparse representation coefficients of ultrasound tumor image patches Local ultrasound tumor image patch In scale The fusion weights below.

[0036] Optionally, S4 includes the following steps:

[0037] S41. Based on multi-scale sparse representation coefficients Construction and preprocessing of ultrasound tumor images Uniformly sized sparse coefficient mapping Each multi-scale sparse representation coefficient Mapped to one-dimensional response intensity To characterize the degree of response of ultrasound tumor image patches in terms of boundary structure expression;

[0038] S42. All one-dimensional response intensities Recombined to form a sparse response map This results in each position corresponding to a sparse representation of the intensity response;

[0039] S43. Based on sparse response graph Using gradient operator , Extract the sparsity variation rates in the horizontal and vertical directions and calculate the edge intensity map. :

[0040]

[0041] in, Indicates the location of a tumor in an ultrasound image. The sparse edge response intensity reflects whether the location is in a boundary change region;

[0042] S44. Regularize the structure weights of the dynamic graph. Integrating into the edge detection process as a structure enhancement factor to adjust the boundary discrimination intensity of different regions in the sparse edge response map at each ultrasound tumor image location. Based on the spatial adjacency of the corresponding ultrasound tumor image patch in the sparse coefficient graph, the dynamic graph regularization structure weights of its neighboring ultrasound tumor image patches are statistically analyzed, and the graph regularization structure enhancement factor corresponding to that position is constructed by weighted averaging. The graph regularization structure enhancement factor With edge intensity map The corresponding positions are multiplied pixel by pixel to obtain the structure-enhanced response map. The structural enhancement response map is used to optimize the boundary discrimination capability of the boundary response map within the continuous region of the structure and to suppress the response values ​​of isolated, broken or falsely detected edge regions.

[0043] S45. Response diagram to structural reinforcement Nonmaximum suppression and double thresholding operations were used to extract the boundary regions with the strongest edge response continuity and integrity, resulting in a preliminary ultrasound tumor boundary map. ,in This indicates that the pixel has been preliminarily identified as a tumor boundary region point.

[0044] Optionally, S5 includes the following steps:

[0045] S51. Preliminary ultrasound tumor boundary map Contour extraction is performed at the mid-boundary location, and an initial set of boundary curves is constructed. ,in Representing an image The Middle Boundary curves This represents the initial number of boundary lines in the image;

[0046] S52. Regularize the structure weights based on the dynamic graph. For the set of boundary curves We construct a structural continuity function by weighted modeling of the structural relationships between points. The structural continuity function is used to characterize the connection strength and directional consistency between adjacent points in a boundary curve:

[0047]

[0048] in, Represents boundary curves The set of adjacent pixel pairs. and These represent the coordinates of adjacent boundary points on the curve. Its structural consistency weight;

[0049] S53. Constructing Boundary Smoothness Constraint Functions The boundary smoothness constraint function is used to control the local curvature changes of the boundary curve, suppressing edge jaggedness and local abnormal oscillations.

[0050]

[0051] in, Represents curve Upper A boundary point, Given the total number of points on the curve, the boundary smoothness constraint function calculates the local discrete curvature formed by each group of three consecutive points;

[0052] S54. Constructing the overall objective function for boundary optimization by combining the structural continuity function and the boundary smoothness function. Joint optimization of the initial boundary curves;

[0053] S55. Set the optimized boundary curves. Mapping back to image space to construct an optimized ultrasound tumor boundary map ,in This indicates that the pixel was ultimately identified as a tumor boundary region point.

[0054] Optionally, S6 includes the following steps:

[0055] S61. Identify all positions with a pixel value of 1 in the optimized ultrasound tumor boundary map as tumor boundary points, and extract the set of boundary pixel positions;

[0056] S62. Using the preprocessed ultrasound tumor image as the base ultrasound tumor image, the boundary pixel location set is layered and fused to generate a fused ultrasound tumor image. During the fusion process, for all locations belonging to the boundary pixel set, ultrasound tumor image enhancement operations are performed, including brightening the pixel, coloring the pixel with a highlight color, or drawing an outline around it to highlight the boundary structure. For non-boundary pixel locations, their original ultrasound tumor image grayscale values ​​are kept unchanged.

[0057] S63. The fused ultrasound tumor image and the optimized ultrasound tumor boundary map are output together as the final ultrasound tumor boundary segmentation result. The fused ultrasound tumor image is used to improve the doctor's visual recognition ability of the tumor contour location and boundary integrity, while the ultrasound tumor boundary map is retained for clinical treatment, diagnostic assistance or system evaluation.

[0058] The beneficial effects of this invention are:

[0059] (1) In the process of constructing a sparse dictionary learning model, the present invention introduces graph regularization constraints and further designs dynamic graph regularization structure weights based on the complexity of regional structure. It can dynamically adjust the strength of regularization terms according to the local structural features of different image blocks, thereby enhancing the ability to express structural information in areas with drastic changes in tumor boundaries or complex textures, and effectively suppressing false edges and noise interference.

[0060] (2) This invention constructs a multi-scale sparse feature dictionary, independently trains and optimizes the sparse representation model at different scales, and then adaptively weights and fuses the representation results at each scale to form a more discriminative boundary response feature, which is particularly significant in the region of blurred boundaries, effectively improving the ability to extract the true edge of the tumor in low-contrast images.

[0061] (3) The present invention constructs a sparse response map and introduces a graph structure enhancement factor. By multiplying the factor with the sparse edge intensity map pixel by pixel, the boundary response of the continuous region of the image structure is enhanced. At the same time, the isolated response that is not related to the structure is attenuated, thereby significantly improving the integrity and accuracy of the boundary. In the subsequent boundary curve optimization, the structural continuity function and the boundary smoothness function are further introduced for joint modeling, which effectively eliminates edge jaggedness, breakpoints and pseudo contours. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a flowchart of a method for precise segmentation of ultrasound tumor boundaries based on sparse representation proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figure 1 A method for coordinated cooling control of a box-type transformer based on a dynamic region allocation algorithm includes the following steps:

[0066] S1. Perform noise suppression, contrast enhancement, and speckle noise removal on the input raw ultrasound tumor image to obtain a preprocessed ultrasound tumor image;

[0067] S2. Based on the preprocessed ultrasound tumor images and the corresponding labeled ultrasound tumor images, a feature dictionary describing the local boundary features of ultrasound tumors is constructed using a sparse dictionary learning method. Graph regularization constraints are introduced during dictionary training to preserve the local structural relationships between adjacent ultrasound tumor image regions, thus obtaining the final optimized multi-scale feature dictionary.

[0068] S3. The preprocessed ultrasound tumor image is segmented into local blocks using the final optimized multi-scale feature dictionary. Sparse coding is performed on each local block to obtain the corresponding multi-scale sparse representation coefficients.

[0069] S4. Based on the multi-scale sparse representation coefficients and their graph regularization optimization results, edge detection algorithms and morphological processing methods are used to extract preliminary boundary information of ultrasound tumors and form a preliminary ultrasound tumor boundary map.

[0070] S5. Based on the preliminary ultrasound tumor boundary map and the local structural information retained in the graph regularization model, the tumor boundary is further corrected and optimized using structural continuity and smoothness optimization methods;

[0071] S6. The optimized ultrasound tumor boundary segmentation results are fused with the preprocessed ultrasound tumor image to form the final ultrasound tumor boundary segmentation image.

[0072] In this embodiment, S1 includes the following steps:

[0073] S11. Input ultrasound tumor image set for acquiring ultrasound tumor images ,in, Indicates the first Original ultrasound tumor image, For each original ultrasound tumor image, the total number of ultrasound tumor images is [number]. Gaussian filtering is performed to suppress background random noise using a two-dimensional Gaussian kernel function. Convolution smoothing is performed to obtain a noise-suppressed ultrasound tumor image;

[0074] S12. Contrast enhancement is performed on the noise-suppressed ultrasound tumor image to enhance local contrast and obtain an enhanced ultrasound tumor image. Then, speckle noise removal is performed to smooth local texture information while preserving edge structure, resulting in a preprocessed ultrasound tumor image. This results in the final preprocessed ultrasound tumor image set. .

[0075] In this embodiment, S2 includes the following steps:

[0076] S21. Based on the preprocessed ultrasound tumor image set and corresponding labeled ultrasound tumor image set Extracting tumor tumor images of size [size] from each ultrasound tumor image using a sliding window method. Pixel-level ultrasound tumor image patches are used to construct ultrasound tumor image patch sets at multiple scales. ,in, This indicates the scale of the ultrasound tumor image patch. Representing scale Next A block of ultrasound tumor images, Representing scale Total number of tumor images on ultrasound;

[0077] S22. For each ultrasound tumor image patch at a given scale, construct an initial sparse feature dictionary. ,in Ultrasound tumor image block Dimensions after flattening The dictionary contains the number of atoms at this scale, used to represent the feature composition of the boundary ultrasound tumor image structure at this scale, and is based on the set of ultrasound tumor image patches. Based on the correspondence between tumor boundaries in labeled ultrasound tumor images, a sparse representation model with graph regularization constraints is constructed:

[0078]

[0079] in, Ultrasound tumor image block In sparse feature dictionary The sparse representation coefficients under the following conditions Representing scale The set of sparse coefficients for all ultrasound tumor image patches. For sparsity control parameters, For graph regularization coefficients, A set of edges that exist between ultrasound tumor image patches, indicating boundary continuity or spatial adjacency. Ultrasound tumor image block With another ultrasound tumor image block Graph regularization structure weights in labeled boundary structures;

[0080] S23. Regularize the weights of the graph structure. A dynamic region sensing mechanism is introduced, which dynamically adjusts the influence of graph constraints based on the structural complexity of each ultrasound tumor image block, and defines dynamic graph regularization structure weights. :

[0081]

[0082] in, Represents ultrasound tumor image blocks The structural complexity factor, when the structural complexity factor The larger the value, the more complex and drastic the tumor boundary in the region, and the stronger the regularization response.

[0083] S24. Jointly optimize the sparse feature dictionary using an alternating minimization strategy. and sparse coefficient set At every scale The process continues until the objective function converges, yielding the final optimized multi-scale feature dictionary. ;

[0084] S25. Weighted fusion of sparse representation results at different scales is performed. By fusing representations at different scales, the ability to characterize tumor boundary contours, structural aberrations, and blurred boundary regions in ultrasound tumor images is optimized, resulting in the final representation of ultrasound tumor image blocks.

[0085]

[0086] in, Indicates the final fusion of the first A block of ultrasound tumor images, For each scale in ultrasound tumor image patches Adaptive fusion weights.

[0087] In this embodiment, S3 includes the following steps:

[0088] S31. Based on the preprocessed ultrasound tumor image set, each ultrasound tumor image is divided into several local ultrasound tumor image blocks to construct a set of local ultrasound tumor image blocks. ,in, Representing ultrasound tumor images In scale The next A local ultrasound tumor image patch The number of ultrasound tumor image blocks obtained by segmenting the ultrasound tumor image.

[0089] S32. For each local ultrasound tumor image patch Using the final optimized multi-scale feature dictionary Sparse representation encoding is performed while maintaining consistency in feature dimensions:

[0090]

[0091] in, Local ultrasound tumor image patch The final optimized multi-scale feature dictionary The sparse representation coefficients are used to characterize the ultrasound tumor boundary structure information of the local region at the current scale.

[0092] S33. All scales of ultrasound tumor images Sparse representation coefficients under By fusing the coefficients, a multi-scale sparse representation coefficient can be constructed. :

[0093]

[0094] in, Indicates the fused first Multiscale sparse representation coefficients of ultrasound tumor image patches Local ultrasound tumor image patch In scale The fusion weights below.

[0095] In this embodiment, S4 includes the following steps:

[0096] S41. Based on multi-scale sparse representation coefficients Construction and preprocessing of ultrasound tumor images Uniformly sized sparse coefficient mapping Each multi-scale sparse representation coefficient Mapped to one-dimensional response intensity To characterize the degree of response of ultrasound tumor image patches in terms of boundary structure expression;

[0097] S42. All one-dimensional response intensities Recombined to form a sparse response map This results in each position corresponding to a sparse representation of the intensity response;

[0098] S43. Based on sparse response graph Using gradient operator , Extract the sparsity variation rates in the horizontal and vertical directions and calculate the edge intensity map. :

[0099]

[0100] in, Indicates the location of a tumor in an ultrasound image. The sparse edge response intensity reflects whether the location is in a boundary change region;

[0101] S44. Regularize the structure weights of the dynamic graph. Integrating into the edge detection process as a structure enhancement factor to adjust the boundary discrimination intensity of different regions in the sparse edge response map at each ultrasound tumor image location. Based on the spatial adjacency of the corresponding ultrasound tumor image patch in the sparse coefficient graph, the dynamic graph regularization structure weights of its neighboring ultrasound tumor image patches are statistically analyzed, and the graph regularization structure enhancement factor corresponding to that position is constructed by weighted averaging. The graph regularization structure enhancement factor With edge intensity map The corresponding positions are multiplied pixel by pixel to obtain the structure-enhanced response map. The structural enhancement response map is used to optimize the boundary discrimination capability of the boundary response map within the continuous region of the structure and to suppress the response values ​​of isolated, broken or falsely detected edge regions.

[0102] S45. Response diagram to structural reinforcement Nonmaximum suppression and double thresholding operations were used to extract the boundary regions with the strongest edge response continuity and integrity, resulting in a preliminary ultrasound tumor boundary map. ,in This indicates that the pixel has been preliminarily identified as a tumor boundary region point.

[0103] In this embodiment, S5 includes the following steps:

[0104] S51. Preliminary ultrasound tumor boundary map Contour extraction is performed at the mid-boundary location, and an initial set of boundary curves is constructed. ,in Representing an image The Middle Boundary curves This represents the initial number of boundary lines in the image;

[0105] S52. Regularize the structure weights based on the dynamic graph. For the set of boundary curves We construct a structural continuity function by weighted modeling of the structural relationships between points. The structural continuity function is used to characterize the connection strength and directional consistency between adjacent points in a boundary curve:

[0106]

[0107] in, Represents boundary curves The set of adjacent pixel pairs. and These represent the coordinates of adjacent boundary points on the curve. Its structural consistency weight;

[0108] S53. Constructing Boundary Smoothness Constraint Functions The boundary smoothness constraint function is used to control the local curvature changes of the boundary curve, suppressing edge jaggedness and local abnormal oscillations.

[0109]

[0110] in, Represents curve Upper A boundary point, Given the total number of points on the curve, the boundary smoothness constraint function calculates the local discrete curvature formed by each group of three consecutive points;

[0111] S54. Constructing the overall objective function for boundary optimization by combining the structural continuity function and the boundary smoothness function. Joint optimization of the initial boundary curves;

[0112] S55. Set the optimized boundary curves. Mapping back to image space to construct an optimized ultrasound tumor boundary map ,in This indicates that the pixel was ultimately identified as a tumor boundary region point.

[0113] In this embodiment, S6 includes the following steps:

[0114] S61. Identify all positions with a pixel value of 1 in the optimized ultrasound tumor boundary map as tumor boundary points, and extract the set of boundary pixel positions;

[0115] S62. Using the preprocessed ultrasound tumor image as the base ultrasound tumor image, the boundary pixel location set is layered and fused to generate a fused ultrasound tumor image. During the fusion process, for all locations belonging to the boundary pixel set, ultrasound tumor image enhancement operations are performed, including brightening the pixel, coloring the pixel with a highlight color, or drawing an outline around it to highlight the boundary structure. For non-boundary pixel locations, their original ultrasound tumor image grayscale values ​​are kept unchanged.

[0116] S63. The fused ultrasound tumor image and the optimized ultrasound tumor boundary map are output together as the final ultrasound tumor boundary segmentation result. The fused ultrasound tumor image is used to improve the doctor's visual recognition ability of the tumor contour location and boundary integrity, while the ultrasound tumor boundary map is retained for clinical treatment, diagnostic assistance or system evaluation.

[0117] Example 1:

[0118] On December 8, 2024, at 9:42 a.m., the image processing workstation of the imaging diagnostic center of a hospital in Province A received a request to upload a liver ultrasound image from the internal medicine outpatient department. The patient was a 58-year-old male, Mr. Wang, who was initially suspected of having a hypoechoic nodule in the right liver during a physical examination. The doctor recommended further image enhancement and segmentation to assess the integrity of the boundary and the trend of size changes.

[0119] The uploaded image number is JSRY20241208015, and the resolution is [not specified]. In the process of constructing the sparse dictionary learning model, this invention introduces graph regularization constraints and further designs dynamic graph regularization structural weights based on the complexity of regional structures. The mechanism can dynamically adjust the strength of the regularization terms according to the local structural features of different image blocks, thereby enhancing the ability to express structural information in areas with drastic changes in tumor boundaries or complex textures, and effectively suppressing false edges and noise interference.

[0120] This invention constructs a multi-scale sparse feature dictionary, independently trains and optimizes the sparse representation model at different scales, and then adaptively weights and fuses the representation results at each scale to form a more discriminative boundary response feature, which is particularly significant in areas with blurred boundaries and effectively improves the ability to extract the true edge of a tumor in low-contrast images.

[0121] This invention constructs a sparse response map and introduces a graph structure enhancement factor. By multiplying this factor with the sparse edge intensity map pixel by pixel, the boundary response of the continuous region of the image structure is enhanced, while the isolated response that is not related to structure is attenuated, thereby significantly improving the integrity and accuracy of the boundary. In the subsequent boundary curve optimization, the structural continuity function and the boundary smoothness function are further introduced for joint modeling, which effectively eliminates edge jaggedness, breakpoints and pseudo-contours. The image size is 512×512, the gray level of the image is low, the tumor boundary area is blurred under naked eye observation, and the contrast with the surrounding liver tissue signal is low. There is a lot of uncertainty in the doctor's initial boundary labeling. After opening the image, A, an employee of the imaging center, noticed that there was a hypoechoic mass of about 28mm in the lower left quadrant of the image near the portal vein branch. The edge was rough, but there was a "window area" in the 3 o'clock position of the image. Traditional methods could not close the contour.

[0122] Employee A first used the traditional ACM method for boundary evolution processing, setting the initial contour as an ellipse surrounding the low-echo region. After 100 iterations, the image showed that the boundary deviated significantly from the solid region, and a pseudo-contour expansion appeared in the upper left corner of the image. The algorithm took about 2.21 seconds, and the final segmentation area was 812 pixels. The boundary error was too large, and the doctor rejected the processing result.

[0123] At 9:58 AM, employee A switched to the method of this invention, and after loading the JSRY20241208015 image, the following process was automatically initiated:

[0124] The image was first preprocessed using a two-dimensional Gaussian filter with a standard deviation of 1.2 and a convolution kernel size of 5×5, resulting in a noise reduction of 18.7%.

[0125] Subsequently, local contrast enhancement was performed, which increased the overall brightness index of the image by 23% and significantly improved the texture discernibility.

[0126] The system automatically extracts three types of image patches from the image: 32×32, 48×48, and 64×64, generating a total of 4387 image patches, which are then integrated into a trained multi-scale sparse dictionary.

[0127] During the encoding process, the system recognizes the region complexity factor δ=1.83 in the fourth quadrant of the image, and the weight of the dynamic graph regularization term is increased, making the boundary expression of this region more sensitive.

[0128] In the sparse coefficient mapping diagram, the response value of the empty window region at the 3 o'clock position increases from 0.14 to 0.78, and the previously unclosable boundary gradually forms a closed loop;

[0129] During the edge extraction stage, the maximum boundary strength shown by the gradient map increased from 0.61 in the traditional method to 0.94, which the system judged as a structural reinforcement boundary.

[0130] Finally, smoothness optimization and structural continuity modeling were performed, with a total system time of 1.64 seconds, and the optimized boundary image was output.

[0131] Employee A imported the processing results into the visualization platform, which automatically highlighted the tumor boundary line. The image number JSRY20241208015_fusion showed that the tumor area was completely closed. Based on the image, the doctor confirmed that the tumor's long diameter was 29.4 mm and its wide diameter was 21.8 mm, with clear boundaries, and included it as a high-risk tumor sample.

[0132] At 2 PM that day, at the segmentation effect comparison report meeting in the imaging center, the image was presented as case number D015 to show the comparison results. In the slides, employee A also showed the boundary maps of the JSRY20241208015 image after processing using three methods, including traditional ACM, U-Net neural network, and the method of the present invention. After observation, the doctors unanimously agreed that the boundary of the traditional ACM method was deformed, the edge of the U-Net processing was obviously jagged, while the boundary provided by the method of the present invention best matched the visual interpretation and pathological corresponding area.

[0133] Throughout December, the center processed a total of 380 images suspected of containing tumors, of which 210 images were processed using the method of this invention. The statistical results are as follows:

[0134] Image number Segmentation methods Average Dice coefficient Hausdorff distance (in pixels) Average processing time (seconds) Manual correction? JSRY20241208015 Method of the present invention 0.91 6.3 1.64 no JSRY20241207342 U-Net Method 0.85 9.8 1.91 yes JSRY20241205087 ACM Methodology 0.77 13.2 2.35 yes

[0135] According to the expert panel's evaluation report, the manual correction rate of the method of this invention is less than 8%, while that of the traditional method is as high as 42%, and the image segmentation consistency is significantly better.

[0136] More importantly, at 3 p.m. on December 15, the surgical team determined the integrity of the tumor based on the image JSRY20241208015_fusion when discussing Wang's condition. This avoided the tendency of misdiagnosis caused by the blurriness of traditional image segmentation. The doctor proposed a local resection plan based on the image, and the tumor range was confirmed in the postoperative pathological section to be almost consistent with the automatic segmentation of the system, with an error of less than 2 mm.

[0137] As can be seen from Example 1, the present invention not only effectively improves the recognition accuracy of tumor boundaries in ultrasound images, especially in low-contrast and structurally complex areas, but also significantly reduces the repetitive work and diagnostic burden of doctors, and has high practical value in actual medical work.

[0138] This invention introduces graph regularization constraints in the process of constructing a sparse dictionary learning model, and further designs dynamic graph regularization structural weights based on the complexity of regional structures. This can dynamically adjust the strength of the regularization terms according to the local structural features of different image blocks, thereby enhancing the ability to express structural information in areas with drastic changes in tumor boundaries or complex textures, and effectively suppressing false edges and noise interference.

[0139] This invention constructs a multi-scale sparse feature dictionary, independently trains and optimizes the sparse representation model at different scales, and then adaptively weights and fuses the representation results at each scale to form a more discriminative boundary response feature, which is particularly significant in areas with blurred boundaries and effectively improves the ability to extract the true edge of a tumor in low-contrast images.

[0140] This invention constructs a sparse response map and introduces a graph structure enhancement factor. By multiplying this factor with the sparse edge intensity map pixel by pixel, the boundary response of the continuous region of the image structure is enhanced, while the isolated response that is not related to the structure is attenuated. This significantly improves the integrity and accuracy of the boundary. In the subsequent boundary curve optimization, the structural continuity function and the boundary smoothness function are further introduced for joint modeling, which effectively eliminates edge jaggedness, breakpoints and pseudo-contours.

[0141] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for precise segmentation of ultrasound tumor boundaries based on sparse representation, characterized in that, Includes the following steps: S1. Perform noise suppression, contrast enhancement, and speckle noise removal on the input raw ultrasound tumor image to obtain a preprocessed ultrasound tumor image; S2. Based on the preprocessed ultrasound tumor images and the corresponding labeled ultrasound tumor images, a feature dictionary describing the local boundary features of ultrasound tumors is constructed using a sparse dictionary learning method. Graph regularization constraints are introduced during dictionary training to preserve the local structural relationships between adjacent ultrasound tumor image regions, thus obtaining the final optimized multi-scale feature dictionary. S3. The preprocessed ultrasound tumor image is segmented into local blocks using the final optimized multi-scale feature dictionary. Sparse coding is performed on each local block to obtain the corresponding multi-scale sparse representation coefficients. S4. Based on the multi-scale sparse representation coefficients and their graph regularization optimization results, edge detection algorithms and morphological processing methods are used to extract preliminary boundary information of ultrasound tumors and form a preliminary ultrasound tumor boundary map. S5. Based on the preliminary ultrasound tumor boundary map and the local structural information retained in the graph regularization model, the tumor boundary is further corrected and optimized using structural continuity and smoothness optimization methods; S6. The optimized ultrasound tumor boundary segmentation results are fused with the preprocessed ultrasound tumor image to form the final ultrasound tumor boundary segmentation image.

2. The method for precise segmentation of ultrasound tumor boundaries based on sparse representation according to claim 1, characterized in that, S1 includes the following steps: S11. Input ultrasound tumor image set for acquiring ultrasound tumor images ,in, Indicates the first Original ultrasound tumor image, For each original ultrasound tumor image, the total number of ultrasound tumor images is [number]. Gaussian filtering is performed to suppress background random noise using a two-dimensional Gaussian kernel function. Convolution smoothing is performed to obtain a noise-suppressed ultrasound tumor image; S12. Contrast enhancement is performed on the noise-suppressed ultrasound tumor image to enhance local contrast and obtain an enhanced ultrasound tumor image. Then, speckle noise removal is performed to smooth local texture information while preserving edge structure, resulting in a preprocessed ultrasound tumor image. This results in the final preprocessed ultrasound tumor image set. .

3. The method for precise segmentation of ultrasound tumor boundaries based on sparse representation according to claim 2, characterized in that, S2 includes the following steps: S21. Based on the preprocessed ultrasound tumor image set and corresponding labeled ultrasound tumor image set Extracting tumor tumor images of size [size] from each ultrasound tumor image using a sliding window method. Pixel-level ultrasound tumor image patches are used to construct ultrasound tumor image patch sets at multiple scales. ,in, This indicates the scale of the ultrasound tumor image patch. Representing scale Next A block of ultrasound tumor images, Representing scale Total number of tumor images on ultrasound; S22. For each ultrasound tumor image patch at a given scale, construct an initial sparse feature dictionary. ,in Ultrasound tumor image block Dimensions after flattening The dictionary contains the number of atoms at this scale, used to represent the feature composition of the boundary ultrasound tumor image structure at this scale, and is based on the set of ultrasound tumor image patches. Based on the correspondence between tumor boundaries in labeled ultrasound tumor images, a sparse representation model with graph regularization constraints is constructed: in, Ultrasound tumor image block In sparse feature dictionary The sparse representation coefficients under the following conditions Representing scale The set of sparse coefficients for all ultrasound tumor image patches. For sparsity control parameters, For graph regularization coefficients, A set of edges that exist between ultrasound tumor image patches, indicating boundary continuity or spatial adjacency. Ultrasound tumor image block With another ultrasound tumor image block Graph regularization structure weights in labeled boundary structures; S23. Regularize the weights of the graph structure. A dynamic region sensing mechanism is introduced, which dynamically adjusts the influence of graph constraints based on the structural complexity of each ultrasound tumor image block, and defines dynamic graph regularization structure weights. : in, Represents ultrasound tumor image blocks The structural complexity factor, when the structural complexity factor The larger the value, the more complex and drastic the tumor boundary in the region, and the stronger the regularization response. S24. Jointly optimize the sparse feature dictionary using an alternating minimization strategy. and sparse coefficient set At every scale The process continues until the objective function converges, yielding the final optimized multi-scale feature dictionary. ; S25. Weighted fusion of sparse representation results at different scales is performed. By fusing representations at different scales, the ability to characterize tumor boundary contours, structural aberrations, and blurred boundary regions in ultrasound tumor images is optimized, resulting in the final representation of ultrasound tumor image blocks. in, Indicates the final fusion of the first A block of ultrasound tumor images, For each scale in ultrasound tumor image patches Adaptive fusion weights.

4. The method for precise segmentation of ultrasound tumor boundaries based on sparse representation according to claim 1, characterized in that, S3 includes the following steps: S31. Based on the preprocessed ultrasound tumor image set, each ultrasound tumor image is divided into several local ultrasound tumor image blocks to construct a set of local ultrasound tumor image blocks. ,in, Representing ultrasound tumor images In scale The next A local ultrasound tumor image patch The number of ultrasound tumor image blocks obtained by segmenting the ultrasound tumor image. S32. For each local ultrasound tumor image patch Using the final optimized multi-scale feature dictionary Sparse representation encoding is performed while maintaining consistency in feature dimensions: in, Local ultrasound tumor image patch The final optimized multi-scale feature dictionary The sparse representation coefficients are used to characterize the ultrasound tumor boundary structure information of the local region at the current scale. S33. All scales of ultrasound tumor images Sparse representation coefficients under By fusing the coefficients, a multi-scale sparse representation coefficient can be constructed. : in, Indicates the fused first Multiscale sparse representation coefficients of ultrasound tumor image patches Local ultrasound tumor image patch In scale The fusion weights below.

5. The method for precise segmentation of ultrasound tumor boundaries based on sparse representation according to claim 1, characterized in that, S4 includes the following steps: S41. Based on multi-scale sparse representation coefficients Construction and preprocessing of ultrasound tumor images Uniformly sized sparse coefficient mapping Each multi-scale sparse representation coefficient Mapped to one-dimensional response intensity To characterize the degree of response of ultrasound tumor image patches in terms of boundary structure expression; S42. All one-dimensional response intensities Recombined to form a sparse response map This results in each position corresponding to a sparse representation of the intensity response; S43. Based on sparse response graph Using gradient operator , Extract the sparsity variation rates in the horizontal and vertical directions and calculate the edge intensity map. : in, Indicates the location of a tumor in an ultrasound image. The sparse edge response intensity reflects whether the location is in a boundary change region; S44. Regularize the structure weights of the dynamic graph. Integrating into the edge detection process as a structure enhancement factor to adjust the boundary discrimination intensity of different regions in the sparse edge response map at each ultrasound tumor image location. Based on the spatial adjacency of the corresponding ultrasound tumor image patch in the sparse coefficient graph, the dynamic graph regularization structure weights of its neighboring ultrasound tumor image patches are statistically analyzed, and the graph regularization structure enhancement factor corresponding to that position is constructed by weighted averaging. The graph regularization structure enhancement factor With edge intensity map The corresponding positions are multiplied pixel by pixel to obtain the structure-enhanced response map. The structural enhancement response map is used to optimize the boundary discrimination capability of the boundary response map within the continuous region of the structure and to suppress the response values ​​of isolated, broken or falsely detected edge regions. S45. Response diagram to structural reinforcement Nonmaximum suppression and double thresholding operations were used to extract the boundary regions with the strongest edge response continuity and integrity, resulting in a preliminary ultrasound tumor boundary map. ,in This indicates that the pixel has been preliminarily identified as a tumor boundary region point.

6. The method for precise segmentation of ultrasound tumor boundaries based on sparse representation according to claim 5, characterized in that, S5 includes the following steps: S51. Preliminary ultrasound tumor boundary map Contour extraction is performed at the mid-boundary location, and an initial set of boundary curves is constructed. ,in Representing an image The Middle Boundary curves This represents the initial number of boundary lines in the image; S52. Regularize the structure weights based on the dynamic graph. For the set of boundary curves We construct a structural continuity function by weighted modeling of the structural relationships between points. The structural continuity function is used to characterize the connection strength and directional consistency between adjacent points in a boundary curve: in, Represents boundary curves The set of adjacent pixel pairs. and These represent the coordinates of adjacent boundary points on the curve. Its structural consistency weight; S53. Constructing Boundary Smoothness Constraint Functions The boundary smoothness constraint function is used to control the local curvature changes of the boundary curve, suppressing edge jaggedness and local abnormal oscillations. in, Represents curve Upper A boundary point, Given the total number of points on the curve, the boundary smoothness constraint function calculates the local discrete curvature formed by each group of three consecutive points; S54. Constructing the overall objective function for boundary optimization by combining the structural continuity function and the boundary smoothness function. Joint optimization of the initial boundary curves; S55. Set the optimized boundary curves. Mapping back to image space to construct an optimized ultrasound tumor boundary map ,in This indicates that the pixel was ultimately identified as a tumor boundary region point.

7. The method for precise segmentation of ultrasound tumor boundaries based on sparse representation according to claim 6, characterized in that, S6 includes the following steps: S61. Identify all positions with a pixel value of 1 in the optimized ultrasound tumor boundary map as tumor boundary points, and extract the set of boundary pixel positions; S62. Using the preprocessed ultrasound tumor image as the base ultrasound tumor image, the boundary pixel location set is layered and fused to generate a fused ultrasound tumor image. During the fusion process, for all locations belonging to the boundary pixel set, ultrasound tumor image enhancement operations are performed, including brightening the pixel, coloring the pixel with a highlight color, or drawing an outline around it to highlight the boundary structure. For non-boundary pixel locations, their original ultrasound tumor image grayscale values ​​are kept unchanged. S63. The fused ultrasound tumor image and the optimized ultrasound tumor boundary map are output together as the final ultrasound tumor boundary segmentation result. The fused ultrasound tumor image is used to improve the doctor's visual recognition ability of the tumor contour location and boundary integrity, while the ultrasound tumor boundary map is retained for clinical treatment, diagnostic assistance or system evaluation.

Images