An ovarian mass segmentation method based on ultrasound images

Abstract

The application discloses an ovarian mass segmentation method based on an ultrasonic image, and relates to the field of medical image processing, and comprises the following steps: uploading an ultrasonic image, calculating a phase compensation factor based on acoustic propagation parameters of the ultrasonic image, and performing adaptive acoustic phase calibration on the ovarian ultrasonic original image through the phase compensation factor; based on acoustic impedance differences of pixel points, three types of core information, including pixel spatial coordinates, gray gradient changes and adjacent pixel acoustic attenuation correlations, are fused to construct a three-dimensional feature tensor for representing an ovarian mass; the application corrects image deviation through adaptive acoustic phase calibration, and strengthens ovarian mass feature representation by fusing multi-dimensional core information, so that effective areas are accurately screened, and discrete interference pixels are removed; a smooth closed contour is constructed based on feature extreme points and anatomical shape constraints, and multi-objective optimization is used to realize accurate boundary separation of the mass and normal tissues.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method for segmenting ovarian masses based on ultrasound images. Background Technology

[0002] Ovarian masses are growths formed by abnormal proliferation of ovarian tissue. They are a common gynecological problem, and can be benign or malignant. Some are asymptomatic, while others may be accompanied by abdominal pain, bloating, etc. The nature of the mass needs to be determined through ultrasound, tumor marker testing, etc.

[0003] The invention patent application with application number 202410802050.5 discloses an ultrasound-based auxiliary diagnostic method for ovarian adnexal masses. This application aims to solve the problem that "current intelligent diagnostic models still require manual cropping of ultrasound images, which is inefficient and lacks an integrated intelligent ultrasound diagnostic system for ovarian masses that combines detection, segmentation, classification diagnosis, and interpretability analysis, and thus lacks clinical practical value."

[0004] However, most existing ovarian ultrasound image segmentation processing technologies directly output the segmented images, making it difficult to control the segmentation accuracy of the output ultrasound images within the target range, which affects the subsequent use of the segmentation results.

[0005] Therefore, we propose a method for ovarian mass segmentation based on ultrasound images. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a method for segmenting ovarian masses based on ultrasound images, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions;

[0008] This invention discloses a method for segmenting ovarian masses based on ultrasound images, comprising:

[0009] Uploaded ultrasound images are processed, and a phase compensation factor is calculated based on the acoustic propagation parameters of the ultrasound images. Adaptive acoustic phase calibration is then performed on the original ovarian ultrasound images using this phase compensation factor. Based on the acoustic impedance differences between pixels, three core information types—pixel spatial coordinates, grayscale gradient changes, and acoustic attenuation correlation between adjacent pixels—are fused to construct a three-dimensional feature tensor characterizing the ovarian mass. The matching degree between a preset acoustic impedance threshold range and the three-dimensional feature tensor is calculated. The consistency of acoustic attenuation correlation among pixel clusters in the three-dimensional feature tensor is analyzed, and discrete pixel clusters are eliminated by setting a threshold coefficient. Simultaneously, a mutation threshold is set based on the abrupt change characteristics of the acoustic impedance boundary between normal ovarian tissue and the mass to retain continuous feature regions, ultimately determining candidate regions for the ovarian mass. The maximum gradient value of the three-dimensional feature tensor of the candidate region is used as the final determination. Points are used as contour vertices, and an adaptive closed contour curve is constructed based on the distribution of extreme points of the feature tensor. With the goal of maximizing feature tensor similarity and minimizing contour curve curvature energy, the parameters of the closed contour curve are adjusted to separate the boundary between the ovarian mass region and normal ovarian tissue. The acoustic impedance matching degree and boundary contour overlap of the boundary separation result with the three-dimensional feature tensor are calculated. If both indicators meet the preset accuracy threshold, the accuracy assessment result is bound to the boundary separation result to form a complete segmentation data package. If the threshold is not met, the deviation region is located based on the accuracy assessment result, and the deviation region is further optimized and corrected using the three-dimensional feature tensor. The accuracy assessment is re-executed until the threshold is met, and the binding is completed. Finally, the ovarian mass segmentation data package with the bound accuracy assessment result is output.

[0010] Furthermore, the formula for calculating the phase compensation factor is as follows:

[0011] ;

[0012] In the formula: This is the phase compensation factor; The center frequency of the ultrasound signal; This refers to the distance the ultrasound probe travels from the ovarian tissue. This represents the actual velocity of ultrasound in the medium containing the ovary. The ultrasonic attenuation coefficient of the medium; This represents the initial phase of the ultrasound signal; This is the reference sound velocity for ultrasound in standard soft tissue;

[0013] Based on the phase compensation factor The phase of the ultrasound echo signal corresponding to each pixel in the original ovarian ultrasound image is reversed and cancelled to complete the adaptive acoustic phase calibration.

[0014] Furthermore, in the construction of the three-dimensional feature tensor, the fusion of pixel spatial coordinates, gray-level gradient changes, and acoustic attenuation correlations between adjacent pixels adopts a weighted feature fusion model:

[0015] ;

[0016] In the formula: Let be the eigenvalues ​​of the three-dimensional feature tensor at spatial coordinates (x, y, z); To integrate weights, All are positive numbers and their sum is 1; Normalized spatial coordinate features of pixels; The grayscale gradient features of the pixel; The acoustic attenuation correlation features of the pixel.

[0017] Furthermore, the preset acoustic impedance threshold range is set according to the following:

[0018] Calculate the global mean and global standard deviation of the acoustic impedance values ​​of all pixels in the ultrasound image, and determine the threshold interval based on the global mean and global standard deviation. ,in This represents the global mean and global standard deviation. This indicates the preset scaling factor.

[0019] Furthermore, the acoustic attenuation correlation consistency analysis logic for each pixel cluster in the three-dimensional feature tensor is as follows:

[0020] ;

[0021] In the formula: The acoustic attenuation correlation consistency coefficient; This represents the total number of pixels within the pixel cluster. Let i be the set of neighboring pixels of pixel i; , Let be the acoustic attenuation coefficient between pixel i and its neighboring pixel j; The average acoustic attenuation coefficient of normal ovarian tissue;

[0022] The matching degree calculation between the preset acoustic impedance threshold interval and the three-dimensional feature tensor follows the following procedure: first, extract the mean acoustic impedance of each pixel cluster in the three-dimensional feature tensor; then, calculate the proportion of pixels whose mean acoustic impedance of each pixel cluster falls within the preset acoustic impedance threshold interval to the total number of pixels in that cluster; use this proportion as the local matching degree between a single pixel cluster and the preset acoustic impedance threshold interval; finally, use the proportion of pixels in each pixel cluster as the weight to perform a weighted average of the local matching degrees of all single pixel clusters, and record the result as the global matching degree between the preset acoustic impedance threshold interval and the three-dimensional feature tensor.

[0023] During the elimination operation, the absolute value of the acoustic impedance difference between adjacent pixels in the three-dimensional feature tensor is calculated. Based on the statistical data of acoustic impedance difference between healthy ovarian tissue and clinically diagnosed ovarian masses, a mutation threshold is set. Continuous feature regions are screened for each pixel cluster. A subset of continuous pixels in the cluster whose absolute values ​​of acoustic impedance difference between adjacent pixels are all lower than the mutation threshold are retained, and discrete pixels whose absolute values ​​of difference exceed the mutation threshold are eliminated.

[0024] The coefficient threshold , This represents the mean acoustic attenuation correlation consistency coefficient of pixel clusters in a healthy ovarian tissue sample. Represents the confidence coefficient; Indicates the coefficient in a healthy ovarian tissue sample variance Indicates the coefficient in ovarian mass tissue samples variance This represents the total number of pixel clusters in a healthy ovarian tissue sample. This represents the total number of pixel clusters in the ovarian mass tissue sample.

[0025] When pixel cluster If the number of pixels is less than the coefficient threshold, or the local matching degree of a single pixel cluster is lower than the preset local matching threshold, or the number of pixels remaining after continuous feature region filtering is lower than the preset pixel number threshold, the pixel cluster is determined to be invalid and removed. The remaining pixel clusters form the candidate region for ovarian masses.

[0026] Furthermore, when constructing the adaptive closed contour curve based on the distribution of extreme points in the feature tensor, extreme points with gradient magnitudes greater than a preset gradient threshold in the three-dimensional feature tensor are selected. These extreme points are sorted according to the polar angle of their spatial coordinates. Then, a cubic B-spline curve is used to fit the sorted extreme points. The arc length between adjacent extreme points is used as the parameter node of the spline curve. The control points of the spline curve are solved by the least squares method. Finally, the position of the control points is adjusted in combination with the elliptical or circular anatomical morphology constraints of ovarian tissue, so that the fitted curve meets the criteria for closure and smoothness.

[0027] Furthermore, when adjusting the parameters of the closed contour curve, a composite optimization function is constructed with the objectives of maximizing the feature tensor similarity and minimizing the contour curve curvature energy:

[0028] ;

[0029] In the formula: This is the set of parameters for a closed contour curve. These are the weighting coefficients; The similarity between the feature tensor of the region enclosed by the contour curve and the feature tensor template of the ovarian mass; The curvature energy of the contour curve;

[0030] in, Including control point coordinates, spline order, Based on cosine similarity calculation It is obtained by integrating the sum of squared curvatures at each point on the curve.

[0031] Furthermore, the formula for calculating the acoustic impedance matching degree and the boundary profile coincidence degree is as follows:

[0032] ;

[0033] In the formula: Acoustic impedance matching; The total number of pixels in the boundary region; For the set of pixels in the boundary region; The acoustic impedance value of the boundary pixel p; Reference values ​​for acoustic impedance of ovarian masses; The standard deviation of the acoustic impedance of the ovarian mass; The degree of overlap of boundary contours; For the pixel set corresponding to the current boundary contour, and for the pixel set corresponding to the acoustic impedance abrupt boundary in the 3D feature tensor; This indicates the number of elements in the set.

[0034] Furthermore, when locating the deviation area based on the accuracy assessment results, the pixel areas with acoustic impedance matching degree lower than a preset threshold and the pixel areas with boundary contour overlap degree lower than a preset threshold are calculated, and the union of the two types of areas is taken as the deviation area.

[0035] During the secondary contour optimization and correction, the three-dimensional feature tensor within the deviation region is locally enhanced, the extreme points of the region are re-extracted, and the local parameters of the closed contour curve are updated. The optimization and correction are iteratively performed until the deviation region is corrected. and All meet the preset accuracy threshold;

[0036] The complete segmentation data package includes: original ultrasound image data, image data after adaptive acoustic phase calibration, three-dimensional feature tensor data, coordinate set of candidate regions for ovarian masses, final closed contour curve parameters, acoustic impedance matching degree value, boundary contour overlap degree value, and deviation region correction record.

[0037] Furthermore, the local enhancement processing of the three-dimensional feature tensor within the deviation region follows the principle of:

[0038] ;

[0039] ;

[0040] In the formula: The eigenvalues ​​of the locally enhanced 3D feature tensor at coordinates (x, y, z); This is half the size of the Gaussian kernel; The standard deviation of the adaptive Gaussian kernel; It is a three-dimensional adaptive Gaussian kernel function; This represents the original 3D feature tensor that is not enhanced within the deviation region; This is the contrast enhancement factor; The variance of the preset reference feature; The original feature variance of the 3×3×3 neighborhood around coordinates (x,y,z); It is the minimum constant; The neighborhood offset coordinates of the Gaussian kernel.

[0041] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects:

[0042] This invention provides a method for ovarian mass segmentation based on ultrasound images. During execution, the method corrects image deviations through adaptive acoustic phase calibration, enhances the feature representation of ovarian masses by fusing multi-dimensional core information, accurately selects effective regions to eliminate discrete interference pixels, constructs smooth closed contours based on feature extrema points and anatomical morphology constraints, achieves precise boundary separation between the mass and normal tissue through multi-objective optimization, and further improves segmentation accuracy by combining secondary contour correction. Two key evaluation indicators ensure the reliability of the results. Finally, the output data package provides comprehensive support for clinical diagnosis and disease analysis. Overall, it reduces ultrasound image segmentation errors, improves segmentation efficiency and robustness, and is applicable to different types of ovarian mass scenarios. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0044] Figure 1 This is a flowchart illustrating a method for segmenting ovarian masses based on ultrasound images. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0046] The present invention will be further described below with reference to embodiments.

[0047] Example:

[0048] This embodiment presents a method for segmenting ovarian masses based on ultrasound images, such as... Figure 1 As shown, it includes:

[0049] Upload ultrasound images, calculate phase compensation factors based on acoustic propagation parameters of ultrasound images, and perform adaptive acoustic phase calibration on the original ovarian ultrasound images using phase compensation factors;

[0050] The formula for calculating the phase compensation factor is:

[0051] ;

[0052] In the formula: This is the phase compensation factor; The center frequency of the ultrasound signal; This refers to the distance the ultrasound probe travels from the ovarian tissue. This represents the actual velocity of ultrasound in the medium containing the ovary. The ultrasonic attenuation coefficient of the medium; This represents the initial phase of the ultrasonic signal; This is the reference sound velocity for ultrasound in standard soft tissue;

[0053] The above formula comprehensively considers the key influencing factors in the propagation process of ultrasound signals, including the center frequency of the ultrasound signal, propagation distance, actual sound velocity, medium ultrasound attenuation coefficient, initial phase, and standard soft tissue reference sound velocity. By constructing mathematical relationships, the phase compensation effect is accurately quantified. At the same time, the attenuation coefficient value is dynamically adjusted according to the frequency of the ultrasound signal and the density of the medium, so as to achieve reverse cancellation correction of the phase of the ultrasound echo signal corresponding to each pixel in the original ovarian ultrasound image, effectively canceling the phase distortion generated during propagation and ensuring the accuracy of subsequent image processing.

[0054] Obtained by inversion of the propagation time difference of multiple marker points in ultrasound images. It is calculated from the scanning parameters of the ultrasound probe and the image pixel resolution;

[0055] Based on phase compensation factor Inverse cancellation correction is performed on the phase of the ultrasound echo signal corresponding to each pixel in the original ovarian ultrasound image to complete adaptive acoustic phase calibration.

[0056] in, The preset value range is 0.5-3.0dB / (MHz·cm). The higher the ultrasound signal frequency and the greater the medium density, the larger the value. The lower the ultrasound signal frequency and the closer the medium is to normal ovarian soft tissue, the smaller the value.

[0057] Based on the acoustic impedance difference of pixels, a three-dimensional feature tensor for characterizing ovarian masses is constructed by integrating three core types of information: pixel spatial coordinates, gray-level gradient changes, and acoustic attenuation correlation between adjacent pixels.

[0058] In the construction of the 3D feature tensor, a weighted feature fusion model is used to fuse pixel spatial coordinates, gray-level gradient changes, and acoustic attenuation correlations between adjacent pixels.

[0059] ;

[0060] In the formula: Let be the eigenvalues ​​of the three-dimensional feature tensor at spatial coordinates (x, y, z); To integrate weights, All are positive numbers and their sum is 1; Normalized spatial coordinate features of pixels; The grayscale gradient features of the pixel; Acoustic attenuation correlation features of pixels;

[0061] The above formula combines three core features: pixel spatial coordinates, gray-level gradient changes, and acoustic attenuation correlation between adjacent pixels. It constructs a three-dimensional feature tensor using a weighted fusion method. The fusion weight is determined by calculating the variance of each feature in the global range of the ultrasound image, so that the weight allocation is related to the discreteness of the feature itself. At the same time, the spatial coordinates are normalized, the gray-level gradient features are calculated and normalized using the Sobel operator, and the acoustic attenuation correlation features are represented by the average acoustic attenuation coefficient between the pixel and its 8 neighboring pixels. This achieves the integration of the three types of features to represent the three-dimensional feature information of the ovarian mass.

[0062] in, It is obtained by dividing x by the width of the ultrasound image, y by the height of the image, and z by the depth of the image. The gradient is obtained by normalizing the three-dimensional directional gradient calculated using the Sobel operator. The acoustic attenuation coefficients of this pixel and its 8 neighboring pixels are all obtained from the values ​​of the acoustic attenuation coefficients.

[0063] , These represent the variances of the three types of features across the entire ultrasound image;

[0064] The matching degree between the preset acoustic impedance threshold range and the three-dimensional feature tensor is calculated. The acoustic attenuation correlation consistency of each pixel cluster in the three-dimensional feature tensor is analyzed to remove discrete pixel clusters by setting a coefficient threshold. At the same time, a mutation threshold is set based on the acoustic impedance boundary mutation characteristics of normal ovarian tissue and mass to retain continuous feature regions. Finally, the candidate region of ovarian mass is determined.

[0065] The preset acoustic impedance threshold range follows the following rules when set:

[0066] Calculate the global mean and global standard deviation of the acoustic impedance values ​​of all pixels in the ultrasound image, and determine the threshold interval based on the global mean and global standard deviation. ,in This represents the global mean and global standard deviation. This represents a preset scaling factor. The acoustic impedance value is obtained by mapping the grayscale value of the ultrasound image to the acoustic impedance calibration curve. The calibration curve is obtained by calibration using a standard phantom model with known acoustic impedance.

[0067] in, This is the lower limit proportional coefficient. ∈[0.8, 2.5], when the acoustic impedance difference between a low-impedance ovarian mass and normal ovarian tissue in an ultrasound image is small. The larger the value, the greater the difference. The smaller the value; This is the upper limit proportional coefficient. ∈[1,3], when the acoustic impedance difference between a high-impedance ovarian mass and normal ovarian tissue in an ultrasound image is small. The larger the value, the greater the difference. The smaller the value;

[0068] The logic for analyzing the consistency of acoustic attenuation correlation among pixel clusters in the three-dimensional feature tensor is as follows:

[0069] ;

[0070] In the formula: The acoustic attenuation correlation consistency coefficient; This represents the total number of pixels within the pixel cluster. Let i be the set of neighboring pixels of pixel i, either 4-neighborhood or 8-neighborhood pixels; , Let be the acoustic attenuation coefficient between pixel i and its neighboring pixel j; The average acoustic attenuation coefficient of normal ovarian tissue;

[0071] The above formula objectively reflects the consistency of the acoustic characteristics of pixels within a cluster by statistically analyzing the differences in acoustic attenuation coefficients between all pixels and their adjacent pixels, and normalizing the results by combining the average acoustic attenuation coefficient of normal ovarian tissue. Furthermore, the global matching degree calculation uses the size of the pixel cluster as the weight to weight the local matching degree, taking into account the influence of clusters of different sizes. The discrete pixel elimination is achieved by setting a mutation threshold to screen continuous feature regions. Combined with the consistency coefficient threshold, local matching degree threshold, and pixel number threshold, invalid pixel clusters are determined from multiple dimensions, accurately retaining continuous pixel regions that match the preset acoustic impedance threshold range and have consistent characteristics, thus providing support for candidate region determination.

[0072] The acoustic attenuation coefficient is based on the exponential attenuation law of ultrasound signals when they propagate a specified distance in the medium where the ovary is located. It is expressed by dividing the logarithm of the received amplitude to the incident amplitude by the propagation distance by 10 times.

[0073] The matching degree calculation between the preset acoustic impedance threshold range and the three-dimensional feature tensor follows the following procedure: First, extract the mean acoustic impedance of each pixel cluster in the three-dimensional feature tensor. Then, calculate the proportion of pixels whose mean acoustic impedance of each pixel cluster falls within the preset acoustic impedance threshold range to the total number of pixels in that cluster. Use this proportion as the local matching degree between a single pixel cluster and the preset acoustic impedance threshold range. Finally, use the proportion of pixels in each pixel cluster as the weight to calculate the weighted average of the local matching degrees of all single pixel clusters. The result is recorded as the global matching degree between the preset acoustic impedance threshold range and the three-dimensional feature tensor.

[0074] During the removal operation, the absolute value of the acoustic impedance difference between adjacent pixels in the three-dimensional feature tensor is calculated. Based on the statistical data of acoustic impedance difference between healthy ovarian tissue and clinically diagnosed ovarian masses, a mutation threshold is set. Continuous feature regions are screened for each pixel cluster. A subset of continuous pixels in the cluster whose absolute value of the acoustic impedance difference between adjacent pixels is lower than the mutation threshold is retained, and discrete pixels whose absolute value of the difference exceeds the mutation threshold are removed.

[0075] coefficient threshold , This represents the mean acoustic attenuation correlation consistency coefficient of pixel clusters in a healthy ovarian tissue sample. Indicates the confidence coefficient; Indicates the coefficient in a healthy ovarian tissue sample variance Indicates the coefficient in ovarian mass tissue samples variance This represents the total number of pixel clusters in a healthy ovarian tissue sample. This represents the total number of pixel clusters in the ovarian mass tissue sample.

[0076] The above formula is based on the essential difference in acoustic attenuation correlation characteristics between normal ovarian tissue and tumor tissue. It calculates the mean, variance, and sample size of the acoustic attenuation correlation consistency coefficient of pixel clusters in the two types of tissue samples. At the same time, it introduces a confidence coefficient to balance the segmentation sensitivity and the false judgment rate. It uses the mean consistency coefficient of healthy tissue samples as a basic threshold reference, and quantifies the dispersion of data distribution by weighted combination of variance and sample size of the two types of tissue samples. This allows the threshold to adapt to the differences in acoustic characteristics of different samples. The confidence coefficient is used to flexibly match the accuracy requirements for the removal of discrete pixel clusters in clinical diagnosis, so that the threshold setting is deeply coupled with the physiological acoustic characteristics of tissue, sample data distribution, and clinical needs.

[0077] When pixel cluster If the number of pixels is less than the coefficient threshold, or the local matching degree of a single pixel cluster is lower than the preset local matching threshold, or the number of pixels remaining after continuous feature region filtering is lower than the preset number of pixels threshold, it is determined to be an invalid pixel cluster and is removed. The remaining pixel clusters form the candidate region for ovarian mass.

[0078] The preset pixel count threshold is based on the pixel count corresponding to the minimum anatomical size of the ovarian mass.

[0079] Using the maximum gradient point of the 3D feature tensor of the candidate region as the contour vertex, an adaptive closed contour curve is constructed based on the distribution of the extreme points of the feature tensor.

[0080] When constructing an adaptive closed contour curve based on the distribution of extreme points in the feature tensor, extreme points with gradient magnitudes greater than a preset gradient threshold in the three-dimensional feature tensor are selected. The extreme points are sorted according to the polar angle of the spatial coordinates. Then, a cubic B-spline curve is used to fit the sorted extreme points. The arc length between adjacent extreme points is used as the parameter node of the spline curve. The control points of the spline curve are solved by the least squares method. Finally, the position of the control points is adjusted in combination with the elliptical or circular anatomical morphology constraints of ovarian tissue so that the fitted curve meets the criteria for closure and smoothness.

[0081] Among them, the preset gradient threshold is determined based on a preset multiple of the statistical mean of the gradient magnitude of the three-dimensional feature tensor. The closure criterion is characterized by the coordinate deviation of the first and last endpoints of the curve not exceeding the preset deviation threshold. The smoothness criterion is characterized by the first derivative between each continuous point of the standard curve being continuous and the curvature change amplitude not exceeding the preset curvature threshold.

[0082] With the goal of maximizing feature tensor similarity and minimizing contour curve curvature energy, the parameters of the closed contour curve are adjusted to separate the boundary between the ovarian mass region and normal ovarian tissue.

[0083] When adjusting the parameters of the closed contour curve, a composite optimization function is constructed with the objectives of maximizing the feature tensor similarity and minimizing the contour curve curvature energy:

[0084] ;

[0085] In the formula: This is the set of parameters for a closed contour curve. These are the weighting coefficients; The similarity between the feature tensor of the region enclosed by the contour curve and the feature tensor template of the ovarian mass; The curvature energy of the contour curve;

[0086] The above formula takes feature tensor similarity and contour curve curvature energy as the core optimization objectives. It balances the importance of the two through weight coefficients. The similarity is calculated based on cosine similarity to ensure that the region enclosed by the contour is highly consistent with the feature template of the ovarian mass. At the same time, the weight coefficients can be flexibly adjusted according to the feature tensor discriminability and actual needs. When the discriminability is high, the focus is on feature matching accuracy. When the discriminability is low or the smoothness requirement is high, the focus is on contour smoothness. Combined with the anatomical constraints of the ovarian tissue's elliptical or circular shape, the contour curve is accurately optimized, effectively separating the boundary between the ovarian mass and normal tissue.

[0087] in, Including control point coordinates, spline order, Based on cosine similarity calculation It is obtained by integrating the sum of squared curvatures at each point on the curve;

[0088] and The values ​​of are all in the range of (0,1). When the discriminative power of the feature tensor is higher, and the matching accuracy requirement between the segmented region and the ovarian mass feature tensor template is higher, The larger the value, the lower the discriminative power of the feature tensor and the higher the priority of contour curve smoothness over feature matching accuracy. The smaller the value, the higher the requirement for smoothness of the contour curve and the lower the discriminative power of the feature tensor. The larger the value, the higher the priority of feature matching accuracy over contour smoothness and feature tensor discriminative power. The smaller the value;

[0089] The acoustic impedance matching degree and boundary contour coincidence degree of the boundary separation result and the three-dimensional feature tensor are calculated. If both indicators meet the preset accuracy threshold, the accuracy evaluation result is bound to the boundary separation result to form a complete segmentation data package. If the threshold is not met, the deviation area is located based on the accuracy evaluation result, and the deviation area is optimized and corrected by combining the three-dimensional feature tensor. The accuracy evaluation is re-executed until the threshold is met and the binding is completed. Finally, the ovarian mass segmentation data package with the bound accuracy evaluation result is output.

[0090] The formulas for calculating acoustic impedance matching degree and boundary profile coincidence degree are as follows:

[0091] ;

[0092] In the formula: Acoustic impedance matching degree; The total number of pixels in the boundary region; For the set of pixels in the boundary region; The acoustic impedance value of the boundary pixel p; Reference values ​​for acoustic impedance of ovarian masses; The standard deviation of the acoustic impedance of the ovarian mass; The degree of overlap of boundary contours; For the pixel set corresponding to the current boundary contour, and for the pixel set corresponding to the acoustic impedance abrupt boundary in the 3D feature tensor; Indicates the number of elements in the set;

[0093] The above formula exponentially attenuates the difference between the acoustic impedance values ​​of pixels in the boundary region and the reference values, highlighting the contribution of pixels with impedance values ​​close to the reference values, and objectively reflecting the degree of fit between the boundary region and the impedance characteristics of the ovarian mass.

[0094] The boundary contour overlap formula uses the ratio of the number of pixels in the intersection to the average number of pixels in the two sets to quantify the degree of fit between the current contour and the impedance change boundary in the feature tensor. The two formulas evaluate the segmentation accuracy from two key dimensions: impedance characteristics and spatial location, respectively, providing a clear and definite quantitative basis for subsequent deviation area localization and secondary optimization.

[0095] Local enhancement processing of the three-dimensional feature tensor within the deviation region follows the following rules:

[0096] ;

[0097] ;

[0098] In the formula: The eigenvalues ​​of the locally enhanced 3D feature tensor at coordinates (x, y, z); This is half the size of the Gaussian kernel; The standard deviation of the adaptive Gaussian kernel is obtained by inverting the characteristic variance of the 3×3×3 neighborhood around the coordinates (x,y,z) in the deviation region; It is a three-dimensional adaptive Gaussian kernel function; This represents the original 3D feature tensor that is not enhanced within the deviation region; This is the contrast enhancement factor; The variance of the preset reference feature; The original feature variance of the 3×3×3 neighborhood around coordinates (x,y,z); It is the minimum constant; Neighborhood offset coordinates of the Gaussian kernel;

[0099] The above formula uses a three-dimensional adaptive Gaussian kernel function, whose standard deviation is obtained by inverting the variance of the 3×3×3 neighborhood features of the pixels in the deviation area, so that the kernel function can adapt to the local feature distribution. At the same time, a contrast enhancement coefficient is introduced, which dynamically adjusts the enhancement intensity based on the ratio of the local feature variance to the reference variance. The smaller the variance, the greater the enhancement, effectively strengthening the feature details in the deviation area, suppressing noise interference, and improving the discriminativeness of the three-dimensional feature tensor in the deviation area through the combination of Gaussian smoothing and contrast enhancement, providing more accurate feature support for secondary contour optimization and correction.

[0100] When locating the deviation area based on the accuracy assessment results, the pixel areas with acoustic impedance matching degree lower than the preset threshold and the pixel areas with boundary contour overlap degree lower than the preset threshold are calculated, and the union of the two types of areas is taken as the deviation area.

[0101] During the secondary contour optimization and correction, the 3D feature tensor within the deviation region is locally enhanced, the extreme points of this region are re-extracted, and the local parameters of the closed contour curve are updated. The optimization and correction are iteratively performed until the deviation region is corrected. and All meet the preset accuracy threshold;

[0102] The complete segmentation data package includes: raw ultrasound image data, image data after adaptive acoustic phase calibration, three-dimensional feature tensor data, coordinate set of candidate regions for ovarian masses, final closed contour curve parameters, acoustic impedance matching degree values, boundary contour overlap degree values, and deviation region correction records.

[0103] The method described in the above embodiments can accurately distinguish between ovarian masses and normal tissue, improve segmentation accuracy and boundary clarity, adapt to masses with different acoustic characteristics, reduce the influence of interference factors, and ensure reliable results through optimization and correction. It provides accurate data support for clinical diagnosis and treatment, and improves the efficiency and accuracy of diagnosis.

[0104] Based on the method in the above embodiments, an application example of this method is shown below:

[0105] To clarify the extent of a space-occupying lesion in a patient's ovary, the Ultrasound Department of XX Hospital used this method to process the patient's original ovarian ultrasound images. First, the patient's ovarian ultrasound images were uploaded. The actual sound velocity in the medium containing the ovary was obtained by inverting the propagation time difference of multiple marker points in the ultrasound image. Combined with the ultrasound probe's scanning parameters and image pixel resolution, the propagation distance from the probe to the ovarian tissue was calculated. Simultaneously, referring to parameters such as the center frequency of the ultrasound signal, the medium's ultrasound attenuation coefficient, the initial phase, and the standard soft tissue reference sound velocity, a phase compensation factor of 0.82 was calculated. Based on this factor, the phase of the ultrasound echo signal corresponding to each pixel in the original ultrasound image was reverse-cancelled to complete adaptive acoustic phase calibration.

[0106] Subsequently, a three-dimensional feature tensor representing the ovarian mass was constructed: the normalized spatial coordinate features of the pixel, the gray-level gradient features calculated and normalized by the Sobel operator, and the acoustic attenuation correlation features obtained by the average acoustic attenuation coefficients of the pixel and its 8 neighboring pixels were fused through a weighted feature fusion model. The fusion weights of the three types of features were determined to be 0.31, 0.38, and 0.31, respectively, and finally the feature values ​​of the three-dimensional feature tensor at each spatial coordinate were obtained.

[0107] Calculate the global mean and global standard deviation of the acoustic impedance values ​​of all pixels in the ultrasound image, and determine the preset acoustic impedance threshold range as [1.1×10]. 6 Pa·s / m, 2.9×10 6 The proportion of pixels whose acoustic impedance mean falls within the range of Pa·s / m is used as the local matching degree, and the global matching degree is calculated by weighting the proportion of the number of pixel clusters. At the same time, the acoustic attenuation correlation consistency coefficient of each pixel cluster is calculated, and the coefficient threshold is set to 0.75. Discrete pixel clusters with consistency coefficients lower than the threshold, insufficient local matching degree, and insufficient number of consecutive pixels are eliminated. Combined with the acoustic impedance boundary change threshold, continuous feature regions are retained to determine the candidate region of ovarian mass.

[0108] Using the maximum gradient point of the candidate region's 3D feature tensor as the contour vertex, extreme points with gradient magnitudes greater than a preset threshold are selected and sorted by their spatial coordinate polar angle. A cubic B-spline curve is used to fit the sorted extreme points. Control points are adjusted in conjunction with constraints related to the elliptical anatomical morphology of the ovary to obtain an adaptive closed contour curve that meets the requirements of closure and smoothness. With the goal of maximizing feature tensor similarity and minimizing contour curve curvature energy, the coordinates of the control points and the spline order of the contour curve are adjusted, with weighting coefficients... Take 0.62, Using a value of 0.38, precise separation of the ovarian mass area from the boundary of normal tissue was achieved.

[0109] Finally, the acoustic impedance matching degree of the boundary separation results was calculated to be 0.93, and the boundary contour overlap degree was 0.91, both of which met the preset accuracy thresholds. The original ultrasound image data, calibrated image data, three-dimensional feature tensor data, candidate region coordinate set, final contour curve parameters, and the values ​​of the two accuracy indicators were bound together to form a complete ovarian mass segmentation data package, which was then output to provide a precise reference for the mass extent in clinical diagnosis.

[0110] In summary, the method described in the above embodiments corrects image deviations through adaptive acoustic phase calibration during execution, enhances the characterization of ovarian masses by fusing multi-dimensional core information, accurately selects effective regions to eliminate discrete interference pixels, constructs smooth closed contours based on feature extreme points and anatomical morphology constraints, achieves precise boundary separation between masses and normal tissues through multi-objective optimization, and further improves segmentation accuracy by combining secondary contour correction. Two key evaluation indicators ensure the reliability of the results, and finally, the output data package provides comprehensive support for clinical diagnosis and disease analysis. Overall, it reduces ultrasound image segmentation errors, improves segmentation efficiency and robustness, and is applicable to different types of ovarian mass scenarios.

[0111] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for segmenting ovarian masses based on ultrasound images, characterized in that, include: Upload ultrasound images, calculate phase compensation factors based on acoustic propagation parameters of ultrasound images, and perform adaptive acoustic phase calibration on the original ovarian ultrasound images using phase compensation factors; Based on the acoustic impedance difference of pixels, a three-dimensional feature tensor for characterizing ovarian masses is constructed by integrating three core types of information: pixel spatial coordinates, gray-level gradient changes, and acoustic attenuation correlation between adjacent pixels. The matching degree between the preset acoustic impedance threshold range and the three-dimensional feature tensor is calculated. The acoustic attenuation correlation consistency of each pixel cluster in the three-dimensional feature tensor is analyzed to remove discrete pixel clusters by setting a coefficient threshold. At the same time, a mutation threshold is set based on the acoustic impedance boundary mutation characteristics of normal ovarian tissue and mass to retain continuous feature regions. Finally, the candidate region of ovarian mass is determined. Using the maximum gradient point of the 3D feature tensor of the candidate region as the contour vertex, an adaptive closed contour curve is constructed based on the distribution of the extreme points of the feature tensor. With the goal of maximizing feature tensor similarity and minimizing contour curve curvature energy, the parameters of the closed contour curve are adjusted to separate the boundary between the ovarian mass region and normal ovarian tissue. The acoustic impedance matching degree and boundary contour coincidence degree of the boundary separation result with the three-dimensional feature tensor are calculated. If both indicators meet the preset accuracy threshold, the accuracy evaluation result is bound to the boundary separation result to form a complete segmentation data packet. If the threshold is not met, the deviation area is located based on the accuracy evaluation result, and the deviation area is optimized and corrected by combining the three-dimensional feature tensor. The accuracy evaluation is re-executed until the threshold is met and the binding is completed. Finally, the ovarian mass segmentation data packet with the bound accuracy evaluation result is output.

2. The method for segmenting ovarian masses based on ultrasound images according to claim 1, characterized in that, The formula for calculating the phase compensation factor is as follows: ； In the formula: This is the phase compensation factor; The center frequency of the ultrasound signal; This refers to the distance the ultrasound probe travels from the ovarian tissue. This represents the actual velocity of ultrasound in the medium containing the ovary. The ultrasonic attenuation coefficient of the medium; This represents the initial phase of the ultrasonic signal; This is the reference sound velocity for ultrasound in standard soft tissue; Based on the phase compensation factor The phase of the ultrasound echo signal corresponding to each pixel in the original ovarian ultrasound image is reversed and cancelled to complete the adaptive acoustic phase calibration.

3. The method for segmenting ovarian masses based on ultrasound images according to claim 1, characterized in that, In the construction of the three-dimensional feature tensor, the fusion of pixel spatial coordinates, gray-level gradient changes, and acoustic attenuation correlations between adjacent pixels adopts a weighted feature fusion model: ； In the formula: Let be the eigenvalues ​​of the three-dimensional feature tensor at spatial coordinates (x, y, z); To integrate weights, All are positive numbers and their sum is 1; Normalized spatial coordinate features of pixels; The grayscale gradient features of the pixel; The acoustic attenuation correlation features of the pixel.

4. The method for segmenting ovarian masses based on ultrasound images according to claim 1, characterized in that, The preset acoustic impedance threshold range is set according to the following: Calculate the global mean and global standard deviation of the acoustic impedance values ​​of all pixels in the ultrasound image, and determine the threshold interval based on the global mean and global standard deviation. ,in This represents the global mean and global standard deviation. This indicates the preset scaling factor.

5. The method for segmenting ovarian masses based on ultrasound images according to claim 1, characterized in that, The acoustic attenuation correlation consistency analysis logic for each pixel cluster in the three-dimensional feature tensor is as follows: ； In the formula: The acoustic attenuation correlation consistency coefficient; This represents the total number of pixels within the pixel cluster. Let i be the set of neighboring pixels of pixel i; , Let be the acoustic attenuation coefficient between pixel i and its neighboring pixel j; The average acoustic attenuation coefficient of normal ovarian tissue; The matching degree calculation between the preset acoustic impedance threshold interval and the three-dimensional feature tensor follows the following procedure: first, extract the mean acoustic impedance of each pixel cluster in the three-dimensional feature tensor; then, calculate the proportion of pixels whose mean acoustic impedance of each pixel cluster falls within the preset acoustic impedance threshold interval to the total number of pixels in that cluster; use this proportion as the local matching degree between a single pixel cluster and the preset acoustic impedance threshold interval; finally, use the proportion of pixels in each pixel cluster as the weight to perform a weighted average of the local matching degrees of all single pixel clusters, and record the result as the global matching degree between the preset acoustic impedance threshold interval and the three-dimensional feature tensor. During the elimination operation, the absolute value of the acoustic impedance difference between adjacent pixels in the three-dimensional feature tensor is calculated. Based on the statistical data of acoustic impedance difference between healthy ovarian tissue and clinically diagnosed ovarian masses, a mutation threshold is set. Continuous feature regions are screened for each pixel cluster. A subset of continuous pixels in the cluster whose absolute values ​​of acoustic impedance difference between adjacent pixels are all lower than the mutation threshold are retained, and discrete pixels whose absolute values ​​of difference exceed the mutation threshold are eliminated. The coefficient threshold , This represents the mean acoustic attenuation correlation consistency coefficient of pixel clusters in a healthy ovarian tissue sample. Indicates the confidence coefficient; Indicates the coefficient in a healthy ovarian tissue sample variance Indicates the coefficient in ovarian mass tissue samples variance This represents the total number of pixel clusters in a healthy ovarian tissue sample. This represents the total number of pixel clusters in the ovarian mass tissue sample. When pixel cluster If the number of pixels is less than the coefficient threshold, or the local matching degree of a single pixel cluster is lower than the preset local matching threshold, or the number of pixels remaining after continuous feature region filtering is lower than the preset pixel number threshold, the pixel cluster is determined to be invalid and removed. The remaining pixel clusters form the candidate region for ovarian masses.

6. The method for segmenting ovarian masses based on ultrasound images according to claim 1, characterized in that, When constructing an adaptive closed contour curve based on the distribution of extreme points in the feature tensor, extreme points with gradient magnitudes greater than a preset gradient threshold in the three-dimensional feature tensor are selected. These extreme points are sorted according to the polar angle of their spatial coordinates. Then, a cubic B-spline curve is used to fit the sorted extreme points. The arc length between adjacent extreme points is used as the parameter node of the spline curve. The control points of the spline curve are solved by the least squares method. Finally, the position of the control points is adjusted in combination with the elliptical or circular anatomical morphology constraints of ovarian tissue, so that the fitted curve meets the criteria for closure and smoothness.

7. The method for segmenting ovarian masses based on ultrasound images according to claim 1, characterized in that, When adjusting the parameters of the closed contour curve, a composite optimization function is constructed with the objectives of maximizing the feature tensor similarity and minimizing the contour curve curvature energy. ； In the formula: This is the set of parameters for a closed contour curve. These are the weighting coefficients; The similarity between the feature tensor of the region enclosed by the contour curve and the feature tensor template of the ovarian mass; The curvature energy of the contour curve; in, Including control point coordinates, spline order, Based on cosine similarity calculation It is obtained by integrating the sum of squared curvatures at each point on the curve.

8. The method for segmenting ovarian masses based on ultrasound images according to claim 1, characterized in that, The formula for calculating the acoustic impedance matching degree and the boundary profile coincidence degree is as follows: ； In the formula: Acoustic impedance matching degree; The total number of pixels in the boundary region; For the set of pixels in the boundary region; The acoustic impedance value of the boundary pixel p; Reference values ​​for acoustic impedance of ovarian masses; The standard deviation of the acoustic impedance of the ovarian mass; The degree of overlap of boundary contours; For the pixel set corresponding to the current boundary contour, and for the pixel set corresponding to the acoustic impedance abrupt boundary in the 3D feature tensor; This indicates the number of elements in the set.

9. The method for segmenting ovarian masses based on ultrasound images according to claim 1, characterized in that, When locating the deviation area based on the accuracy assessment results, the pixel areas with acoustic impedance matching degree lower than the preset threshold and the pixel areas with boundary contour overlap degree lower than the preset threshold are calculated, and the union of the two types of areas is taken as the deviation area. During the secondary contour optimization and correction, the three-dimensional feature tensor within the deviation region is locally enhanced, the extreme points of the region are re-extracted, and the local parameters of the closed contour curve are updated. The optimization and correction are iteratively performed until the deviation region is corrected. and All meet the preset accuracy threshold; The complete segmentation data package includes: original ultrasound image data, image data after adaptive acoustic phase calibration, three-dimensional feature tensor data, coordinate set of candidate regions for ovarian masses, final closed contour curve parameters, acoustic impedance matching degree value, boundary contour overlap degree value, and deviation region correction record.

10. The method for segmenting ovarian masses based on ultrasound images according to claim 9, characterized in that, The local enhancement processing of the three-dimensional feature tensor within the deviation region follows the following: ； ； In the formula: The eigenvalues ​​of the locally enhanced 3D feature tensor at coordinates (x, y, z); This is half the size of the Gaussian kernel; The standard deviation of the adaptive Gaussian kernel; It is a three-dimensional adaptive Gaussian kernel function; This represents the original 3D feature tensor that is not enhanced within the deviation region; This is the contrast enhancement factor; The variance of the preset reference feature; The original feature variance of the 3×3×3 neighborhood around coordinates (x,y,z); It is the minimum constant; The neighborhood offset coordinates of the Gaussian kernel.

Images