Cooperative segmentation method and system for vaginal canal

Abstract

The invention discloses a collaborative segmentation method and system for a vaginal canal, and relates to the technical field of medical image processing, and the method comprises the steps: outputting the comprehensive features of a plurality of vaginal canal regions based on a plurality of vaginal canal target images; performing focus region segmentation on a target vaginal canal three-dimensional shape model constructed according to each comprehensive feature to generate a plurality of initial focus region images and corresponding initial focus segmentation boundary curves; determining an initial adaptive topology constraint through each initial lesion area image, and determining an optimized adaptive topology constraint based on the initial adaptive topology constraint; determining a priori curve and a plurality of priori segmentation curves according to the optimized adaptive topological constraint and an adaptive grid refinement strategy; determining a plurality of target focus area images based on the priori curve, the priori curvature distribution of each priori segmentation curve and the optimized adaptive topology constraint; and converting each target lesion area image into a plurality of vaginal canal lesion visual images. Based on the scheme, the reliability of vaginal canal segmentation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a collaborative segmentation method and system for vaginal canal. Background Art

[0002] In modern medical examinations, image segmentation technology has shown great potential in segmenting lesions in medical images obtained by computed tomography, magnetic resonance imaging, etc.

[0003] The vaginal canal is a complex three-dimensional tubular structure, and its morphology changes with changes in physiological states, such as vaginal relaxation and vaginal spasm. These changes will cause changes in the local morphology and topological structure of the vaginal canal, which brings challenges to lesion segmentation. Traditional vaginal canal segmentation methods usually rely on fixed topological models, which are difficult to adapt to the dynamic changes of the vaginal canal, resulting in low reliability of vaginal canal segmentation. Summary of the invention

[0004] The present invention provides a collaborative segmentation method and system for the vaginal canal, which solves the technical problem that the traditional vaginal canal segmentation method usually relies on a fixed topological model and is difficult to adapt to the dynamic change characteristics of the vaginal canal, resulting in low reliability of vaginal canal segmentation.

[0005] The present invention provides a collaborative segmentation method for vaginal canal, comprising:

[0006] Acquire multiple vaginal canal medical images according to a preset step length and perform preprocessing to determine multiple vaginal canal target images;

[0007] Based on each of the vaginal canal target images, multi-channel feature extraction and cross-dimensional feature interaction are performed to output comprehensive features of multiple vaginal canal regions;

[0008] Performing parametric modeling of the vaginal canal according to the comprehensive features to generate a target vaginal canal three-dimensional morphological model, performing lesion area segmentation on the target vaginal canal three-dimensional morphological model, and generating a plurality of initial lesion area images and corresponding initial lesion segmentation boundary curves;

[0009] Determining initial adaptive topological constraints based on each of the initial lesion region images, performing multi-directional collaborative analysis on each of the initial lesion region images under the constraints of the initial adaptive topological constraints, and determining optimized adaptive topological constraints;

[0010] Under the constraint of the optimized adaptive topological constraint, construct a priori curve according to each of the initial lesion segmentation boundary curves, and determine a plurality of priori segmentation curves according to the prior curve through an adaptive grid refinement strategy;

[0011] Performing image segmentation on the target vaginal canal three-dimensional morphological model based on the prior curve and the prior curvature distribution of each prior segmentation curve, and performing multi-directional collaborative analysis under the constraint of the optimized adaptive topological constraint to determine multiple target lesion area images;

[0012] The target lesion area images are subjected to image enhancement rendering and converted into a plurality of vaginal canal lesion visual images.

[0013] Optionally, the multi-channel feature extraction and cross-dimensional feature interaction are performed based on each of the vaginal canal target images to output comprehensive features of multiple vaginal canal regions, including:

[0014] Perform multi-channel feature extraction on each of the vaginal canal target images to obtain multiple channel features of each of the vaginal canal target images;

[0015] According to each of the vaginal canal target images, the associated channel features are sequentially input into the cross-latitude interaction model for feature processing to obtain a plurality of vaginal canal change sensitive area images of each of the vaginal canal target images;

[0016] Using a cluster analysis algorithm to classify the images of the sensitive areas of the vaginal canal changes, and determining a plurality of vaginal canal areas;

[0017] Performing edge detection on the vaginal canal change sensitive region image of each of the vaginal canal regions according to an edge detection algorithm to obtain a plurality of initial contours of each of the vaginal canal regions;

[0018] splicing the initial contours of the vaginal canal regions to obtain a sequence of initial contours of the vaginal canal regions;

[0019] The time series features of each initial contour sequence are extracted respectively through a preset first time series model, and the dynamic change features of each vaginal canal area are output;

[0020] Using a preset second time series model to respectively detect data difference values ​​between adjacent initial contours in each initial contour sequence, and associating relatively later initial contours with data difference values ​​greater than or equal to a preset tolerance value to form shape features of each of the vaginal canal regions;

[0021] The shape features and dynamic change features of each vaginal canal region are converted into the same dimension and then spliced ​​to output the comprehensive features of each vaginal canal region.

[0022] Optionally, the parametric modeling of the vaginal canal is performed according to each of the comprehensive features to generate a target vaginal canal three-dimensional morphological model, and the lesion area is segmented on the target vaginal canal three-dimensional morphological model to generate a plurality of initial lesion area images and corresponding initial lesion segmentation boundary curves, including:

[0023] The comprehensive features are fused by a Poisson fusion algorithm to obtain a four-dimensional feature array;

[0024] The elements of the four-dimensional feature array are regarded as nodes in a minimum spanning tree, and a minimum tree generation algorithm is used to perform parameterization processing on the four-dimensional feature array to output the geometric shape and topological structure of the vaginal canal;

[0025] Performing curve fitting on each of the comprehensive features to obtain multiple smooth curves;

[0026] Using the smooth curves, the geometric shapes and the topological structures, an initial vaginal canal three-dimensional morphological model is constructed based on surface reconstruction technology;

[0027] Performing lesion area feature enhancement on the initial vaginal canal three-dimensional morphological model through self-calibration convolution to obtain a target vaginal canal three-dimensional morphological model;

[0028] The U-net network is used to perform lesion region segmentation on the target vaginal canal three-dimensional morphological model, and a plurality of initial lesion region images and corresponding initial lesion region segmentation boundary curves are determined.

[0029] Optionally, determining an initial adaptive topological constraint based on each of the initial lesion region images, performing multi-directional collaborative analysis on each of the initial lesion region images under the constraints of the initial adaptive topological constraint, and determining an optimized adaptive topological constraint, comprises:

[0030] Using a minimum graph cut algorithm to perform image segmentation on each of the initial lesion region images to obtain a plurality of initial segmented lesion region images and a plurality of initial cut sets;

[0031] Performing image feature analysis on each of the initial segmented lesion region images and each of the initial cut sets to determine initial adaptive topological constraints;

[0032] Under the constraint of the initial adaptive topological constraint, feature extraction is performed on each of the initial lesion region images in different directions, and multiple direction analysis results of each of the initial lesion region images are output;

[0033] If the directional analysis results in any same direction satisfy the preset consistency difference range, the initial adaptive topological constraint is used as a regularization term to construct an initial energy functional;

[0034] Determine the extreme value condition of the initial energy functional by using a variational method, and iteratively solve the initial energy functional according to the extreme value condition by using a gradient descent method, and output an optimized cut set;

[0035] Performing image segmentation on the target vaginal canal three-dimensional morphological model using a minimum graph cut algorithm according to the optimized cut set to obtain multiple intermediate lesion area images;

[0036] Calculating an intersection-over-union ratio using each of the initial lesion region images and the corresponding intermediate lesion region images;

[0037] If any of the intersection-and-union ratios is greater than the intersection-and-union ratio threshold, the initial adaptive topology constraint is used as the optimized adaptive topology constraint;

[0038] If the directional analysis results in any of the same directions do not satisfy the preset consistency difference range or any of the intersection-and-union ratios is less than or equal to the intersection-and-union ratio threshold, the corresponding directional analysis results of each of the initial lesion area images are respectively input into the cross-dimensional interaction model for feature processing, and multiple optimized fusion directional analysis results of each of the initial lesion area images are output;

[0039] The optimized fusion direction analysis results are concatenated to generate an optimized multi-directional collaborative result, and the optimized multi-directional collaborative result is input into an attention model to output a directional weighted feature;

[0040] The directional weighted feature is weighted with the initial adaptive topological constraint to obtain a new initial adaptive topological constraint, and the number of updates is counted in real time;

[0041] If the update times are equal to the preset times threshold, the initial adaptive topology constraint is used as the optimized adaptive topology constraint;

[0042] If the update number is less than the preset number threshold, multiple new directional analysis results of each initial lesion area image are determined under the constraint of the initial adaptive topological constraint until any directional analysis results in the same direction meet the preset consistency difference range or the update number is equal to the preset number threshold.

[0043] Optionally, constructing a priori curves according to each of the initial lesion segmentation boundary curves under the constraint of the optimized adaptive topological constraint, and determining a plurality of priori segmentation curves according to the priori curves by an adaptive grid refinement strategy, comprises:

[0044] Extracting a plurality of discrete points from each of the initial lesion segmentation boundary curves respectively, and generating a priori curves based on each of the discrete points using a Bezier curve under the constraint of the optimized adaptive topological constraint;

[0045] Inputting the priori curve into a convolutional neural network and outputting a curve residual;

[0046] Using the curve residual as a regularization term to correct the priori curve to obtain a calibration curve;

[0047] filtering the calibration curve to output scale curves of multiple scales;

[0048] Calculating the curvature of each scale curve to determine the curvature distribution of each scale;

[0049] Extracting high curvature points and inflection points from each of the curvature distributions, and performing adaptive grid refinement on the regions where each of the high curvature points and each of the inflection points are located in the calibration curve to determine a refined curve;

[0050] Using a minimum graph cut algorithm to perform image segmentation on the thinning curve to determine a thin segmentation curve;

[0051] The priori curve is used to calculate errors respectively with each of the refined segmentation curves, and a plurality of the errors are used to form an error distribution;

[0052] When the error distribution does not meet the preset error condition, adaptive grid refinement and control point density adjustment are performed in the refinement curve according to the error distribution, and a new refinement curve is determined until the error distribution meets the preset error condition, and the refined segmentation curve is determined as the priori segmentation curve.

[0053] Optionally, the image segmentation of the target vaginal canal three-dimensional morphological model based on the prior curve and the prior curvature distribution of each prior segmentation curve, and multi-directional collaborative analysis under the constraint of the optimized adaptive topological constraint to determine multiple target lesion area images, including:

[0054] Performing multi-scale curvature analysis on the prior curve and each of the prior segmentation curves to obtain a plurality of prior curvature distributions;

[0055] The priori curve and the priori curvature distribution of each priori segmentation curve are fitted respectively by spline curve fitting technology to obtain a plurality of spline curves;

[0056] The curvature change rate of each spline curve is used for fitting to obtain a curvature change function, and the curvature change function is used as a regularization term to construct a curvature energy functional;

[0057] Performing image segmentation on the target vaginal canal three-dimensional morphological model based on the curvature energy functional by using a minimum graph cut algorithm to obtain a plurality of curvature lesion area images;

[0058] Under the constraint of optimizing the adaptive topological constraint, performing feature extraction in different directions on each of the curvature lesion region images, and outputting a plurality of curvature direction analysis results of each of the curvature lesion region images;

[0059] If the preset consistency difference range is satisfied between the analysis results of curvature direction in any same direction, the optimized adaptive topological constraint is used as a regularization term to construct the target energy functional;

[0060] If the directional analysis results in any same direction do not meet the preset consistency difference range, the curvature directional analysis results of each of the curvature lesion area images are respectively input into the cross-latitude interaction model for feature processing, and multiple curvature fusion directional analysis results of each of the curvature lesion area images are output and spliced ​​to generate a curvature multi-directional collaborative result;

[0061] The curvature direction weighted feature obtained by inputting the curvature multi-directional coordination result into the attention model is weighted with the optimized adaptive topological constraint to obtain a new optimized adaptive topological constraint, and the number of optimization updates is counted in real time;

[0062] If the number of optimization updates is equal to a preset optimization number threshold, the new optimization adaptive topological constraint is used as a regularization term to construct a target energy functional;

[0063] If the number of optimization updates is less than the preset optimization number threshold, multiple new curvature direction analysis results of each curvature lesion area image are determined under the constraint of the optimization adaptive topological constraint, until any curvature direction analysis results in the same direction meet the preset consistency difference range or the number of optimization updates is equal to the preset optimization number threshold;

[0064] Using a minimum graph cut algorithm to perform image segmentation on the target vaginal canal three-dimensional morphological model based on the target energy functional to obtain a plurality of optimized lesion area images;

[0065] If the standard deviation of the curvature distribution of any of the optimized lesion area images does not meet the convergence conditions, a new optimized adaptive topological constraint is determined based on the multi-directional collaborative results of the curvature of each of the curvature lesion area images until the standard deviation of the curvature distribution of any of the optimized lesion area images meets the preset convergence conditions or the number of optimization updates is equal to the preset optimization number threshold, then the optimized lesion area image is determined as the target lesion area image.

[0066] Optionally, the feature processing process of the cross-dimensional interaction model includes:

[0067] Perform full permutation and combination of multiple input features of the associated input to obtain multiple feature combinations;

[0068] Dividing the input features in each of the feature combinations into a plurality of weighted features and fusion features according to the sorting;

[0069] After weighted summing each weighted feature in each feature combination, channel splicing is performed with the corresponding fusion feature to obtain the splicing feature of each feature combination;

[0070] The key area features of each spliced ​​feature are extracted through a cross-dimensional attention mechanism to obtain multiple output features.

[0071] Optionally, performing image enhancement rendering on each of the target lesion area images to convert them into a plurality of vaginal canal lesion visual images includes:

[0072] Performing Gaussian smoothing filtering and morphological opening and closing operations on each of the target lesion area images, and outputting a plurality of corrected lesion area images;

[0073] Calculating the segmentation accuracy using each of the corrected lesion region images and the corresponding standard lesion region images;

[0074] If any of the segmentation accuracy does not meet the expected segmentation standard, each of the corrected lesion area images is used as a new target lesion area image until the segmentation accuracy meets the expected segmentation standard and is determined as the target vaginal canal segmentation lesion image;

[0075] Each of the target vaginal canal segmentation lesion images is converted into a vaginal canal lesion visual image through image rendering technology.

[0076] A second aspect of the present invention provides a collaborative segmentation system for a vaginal canal, comprising:

[0077] An image processing module, used for acquiring a plurality of vaginal canal medical images according to a preset step length and performing preprocessing to determine a plurality of vaginal canal target images;

[0078] A feature extraction module, used for performing multi-channel feature extraction and cross-dimensional feature interaction based on each of the vaginal canal target images, and outputting comprehensive features of multiple vaginal canal regions;

[0079] A preliminary segmentation module is used to perform parametric modeling of the vaginal canal according to each of the comprehensive features to generate a target vaginal canal three-dimensional morphological model, perform lesion area segmentation on the target vaginal canal three-dimensional morphological model, and generate a plurality of initial lesion area images and corresponding initial lesion segmentation boundary curves;

[0080] A constraint optimization module, used for determining an initial adaptive topological constraint based on each of the initial lesion region images, performing a multi-directional collaborative analysis on each of the initial lesion region images under the constraints of the initial adaptive topological constraint, and determining an optimized adaptive topological constraint;

[0081] A curve optimization module, used for constructing a priori curves according to each of the initial lesion segmentation boundary curves under the constraint of the optimization adaptive topological constraint, and determining a plurality of priori segmentation curves according to the prior curves through an adaptive grid refinement strategy;

[0082] A target segmentation module, used for performing image segmentation on the target vaginal canal three-dimensional morphological model based on the prior curve and the prior curvature distribution of each prior segmentation curve, and performing multi-directional collaborative analysis under the constraint of the optimized adaptive topological constraint to determine multiple target lesion area images;

[0083] The visualization module is used to perform image enhancement rendering on each of the target lesion area images and convert them into multiple vaginal canal lesion visual images.

[0084] A third aspect of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of any of the methods described above.

[0085] It can be seen from the above technical solutions that the present invention has the following advantages:

[0086] The above scheme of the present invention provides a collaborative segmentation method for the vaginal canal, including: acquiring multiple vaginal canal medical images of the vaginal canal according to a preset step length and performing preprocessing to determine multiple vaginal canal target images; performing multi-channel feature extraction and cross-dimensional feature interaction based on each vaginal canal target image, and outputting comprehensive features of multiple vaginal canal regions; performing parametric modeling of the vaginal canal according to each comprehensive feature to generate a target vaginal canal three-dimensional morphological model, performing lesion region segmentation on the target vaginal canal three-dimensional morphological model, and generating multiple initial lesion region images and corresponding initial lesion segmentation boundary curves; determining initial adaptive topological constraints based on each initial lesion region image, and generating a plurality of initial lesion region images and corresponding initial lesion segmentation boundary curves in the initial lesion region image. Under the constraint of the initial adaptive topological constraint, a multi-directional collaborative analysis is performed on each initial lesion area image to determine the optimized adaptive topological constraint; under the constraint of the optimized adaptive topological constraint, a priori curve is constructed according to each initial lesion segmentation boundary curve, and multiple priori segmentation curves are determined through an adaptive mesh refinement strategy based on the prior curve; image segmentation is performed on the target vaginal canal three-dimensional morphological model based on the priori curve and the prior curvature distribution of each priori segmentation curve, and a multi-directional collaborative analysis is performed under the constraint of the optimized adaptive topological constraint to determine multiple target lesion area images; image enhancement rendering is performed on each target lesion area image to convert it into multiple vaginal canal lesion visual images. Based on the above scheme, the medical images of the vaginal canal are preprocessed to ensure the clarity and integrity of the data. The three-dimensional morphological model of the target vaginal canal is reconstructed by capturing the target image features of the vaginal canal and parameterizing modeling, which is conducive to adapting to the dynamic change characteristics of the vaginal canal to accurately present the spatial structure and morphological characteristics of the vaginal canal. The collaborative segmentation mechanism of image segmentation, adaptive topological constraints, adaptive grid refinement strategy and curvature analysis is adopted to facilitate the combination of the dynamic change characteristics of the vaginal canal with topological constraints to achieve real-time response and adaptive adjustment to the changes of the vaginal canal, thereby improving the reliability of vaginal canal segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0088] Figure 1 A flowchart of the steps of a collaborative segmentation method for a vaginal canal provided by an embodiment of the present invention;

[0089] Figure 2 A schematic diagram of feature processing of a cross-dimensional interaction model provided by an embodiment of the present invention;

[0090] Figure 3 A structural block diagram of a collaborative segmentation system for the vaginal canal provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0091] The embodiments of the present invention provide a collaborative segmentation method and system for the vaginal canal, which is used to solve the technical problem that traditional vaginal canal segmentation methods usually rely on a fixed topological model and are difficult to adapt to the dynamic changes of the vaginal canal, resulting in low reliability of vaginal canal segmentation.

[0092] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0093] See also Figure 1 , Figure 1 A flowchart of the steps of a collaborative segmentation method for a vaginal canal provided in an embodiment of the present invention.

[0094] The present invention provides a collaborative segmentation method for vaginal canal, comprising:

[0095] Step 101: Acquire multiple vaginal canal medical images according to a preset step length and perform preprocessing to determine multiple vaginal canal target images.

[0096] Step 101 includes the following sub-steps:

[0097] S11 uses an image processing algorithm to perform preliminary denoising on each vaginal canal medical image to obtain a first medical image.

[0098] It should be noted that medical images of the vaginal canal can be obtained through high-resolution medical imaging equipment (such as MRI or CT). Assuming that we use a 3T MRI device and scan at a resolution of 5 mm, we can obtain multiple three-dimensional image data of the vaginal canal, which will show the condition of the vaginal canal from different angles and time points. These data are usually interfered by noise, so it is necessary to use an image processing algorithm for preliminary denoising. For example, a denoising algorithm based on non-local means (Non-LocalMeans) can smooth the image in the spatial domain while retaining edge details.

[0099] S12: If the image resolution of each of the first medical images is insufficient, a super-resolution reconstruction algorithm is used to perform image reconstruction to obtain a second medical image.

[0100] It should be noted that if the image resolution after denoising is still insufficient, a super-resolution reconstruction algorithm can be used, such as the super-resolution reconstruction network (SRCNN) based on the convolutional neural network (CNN), to perform super-resolution reconstruction of the image data to achieve a high resolution of 25 mm.

[0101] S13. Using an image registration algorithm to perform registration and calibration on each second medical image to obtain a third medical image.

[0102] It should be noted that for image data acquired at different time points or angles, feature point-based image registration technology can be used, such as using the SIFT (Scale-Invariant Feature Transform) algorithm to extract and match feature points, and using the RANSAC (Random Sample Consensus) algorithm for registration and calibration to obtain high-quality image data in a unified coordinate system.

[0103] S14. If each of the third medical images has image distortion or deformation, a geometric correction algorithm is used to perform geometric correction to obtain a fourth medical image.

[0104] It should be noted that if there is image distortion or deformation in the image data, the image data can be geometrically corrected using a geometric correction algorithm, such as applying a geometric correction algorithm based on a B-spline deformation model, to ensure that the morphological structure of the image data is accurate.

[0105] S15. Perform enhancement processing on each fourth medical image using an image enhancement algorithm to obtain a fifth medical image.

[0106] It should be noted that, based on the above data, the data is enhanced by an image enhancement algorithm, for example, a contrast enhancement algorithm based on histogram equalization is used to make the tissue structure clearer and the contrast higher.

[0107] S16. Collect blood perfusion images of the vaginal canal through dynamic contrast-enhanced MRI, and fuse each fifth medical image with the corresponding blood perfusion image using a multimodal image fusion algorithm to obtain multiple vaginal canal target images.

[0108] It should be noted that by combining the image data with the blood perfusion information inside the vaginal canal, the dynamic contrast-enhanced MRI (DCE-MRI) technology can be used to obtain blood perfusion images, and multimodal image fusion technology can be used to fuse the structural image and blood perfusion information through a fusion algorithm based on mutual information to obtain a target image of the vaginal canal including the blood perfusion situation. In this way, the above-mentioned high processing accuracy and enhancement technology can be comprehensively utilized to obtain high-quality and detailed medical imaging data for clinical diagnosis and treatment.

[0109] Step 102: Perform multi-channel feature extraction and cross-dimensional feature interaction based on each vaginal canal target image, and output comprehensive features of multiple vaginal canal regions.

[0110] Step 102 includes the following sub-steps:

[0111] S21, performing multi-channel feature extraction on each vaginal canal target image to obtain multiple channel features of each vaginal canal target image.

[0112] It should be noted that multi-channel feature extraction is performed on each vaginal canal target image to extract information of different dimensions, such as channel features of the R channel, the G channel, and the B channel.

[0113] S22. According to each vaginal canal target image, the associated channel features are sequentially input into the cross-dimensional interaction model for feature processing to obtain multiple vaginal canal change sensitive area images of each vaginal canal target image.

[0114] Preferably, the feature processing process of the cross-dimensional interaction model includes:

[0115] Perform full permutation and combination of multiple input features of the associated input to obtain multiple feature combinations;

[0116] Divide the input features in each feature combination into multiple weighted features and fusion features according to the sorting;

[0117] After weighted summing each weighted feature in each feature combination, channel splicing is performed with the corresponding fusion feature to obtain the splicing feature of each feature combination;

[0118] The key area features of each spliced ​​feature are extracted through the cross-dimensional attention mechanism to obtain multiple output features.

[0119] It should be noted that, in this embodiment, Figure 2As shown, taking each vaginal canal target image (data) as a unit, the associated multiple channel features are used as input features (X0, X1, ..., Xn) of the cross-dimensional interaction model for full permutation and combination. Taking the existence of three channel features (RGB) as an example, there are 3*2*1=6 combinations for full permutation, thereby obtaining 6 feature combinations; in each feature combination, multiple input features (X0, X1, ..., Xn-1) are used as weighted features in sequence, and the input feature (Xn) is used as a fusion feature, and the associated multiple weighted features are weighted summed, and the fusion feature, that is, the channel data without combining the hierarchical dimension, is introduced as supplementary data for channel splicing, and the spliced ​​features are output, for example, the R and G channel features are weighted summed, and the B channel features are used as supplementary data; after that, each spliced ​​feature can be evaluated through a cross-dimensional attention mechanism (Triplet Attention) or an attention layer such as channel attention or spatial attention to identify the key area, and obtain the area where the structure changes most sensitively such as the vaginal canal wall thickness and the lumen diameter, so as to output the image of the sensitive area of ​​the vaginal canal change.

[0120] S23, using a cluster analysis algorithm to classify the images of each vaginal canal change-sensitive region to determine multiple vaginal canal regions.

[0121] It should be noted that after the images of the sensitive areas of vaginal canal changes are obtained, they are divided into different categories using a cluster analysis algorithm, thereby determining multiple vaginal canal areas.

[0122] S24, performing edge detection on the vaginal canal change sensitive region image of each vaginal canal region according to an edge detection algorithm to obtain multiple initial contours of each vaginal canal region.

[0123] It should be noted that an edge detection algorithm such as the Canny edge detection algorithm is used to binarize the images of each vaginal canal change-sensitive area by setting thresholds such as 50 and 150, identify significant boundary changes in the area, and extract contours to obtain multiple initial contours of each vaginal canal area.

[0124] S25, splicing the initial contours of the vaginal canal regions to obtain an initial contour sequence of the vaginal canal regions.

[0125] It should be noted that the multiple images contain n time points, and the initial contours of multiple vaginal canal regions are detected in the images at each time point. The initial contours of each vaginal canal region at n time points are spliced ​​to obtain the corresponding initial contour sequence.

[0126] S26. Extracting time series features from each initial contour sequence using a preset first time series model, and outputting dynamic change features of each vaginal canal region.

[0127] It should be noted that the initial contour sequence is input into the first time series model, such as the LSTM model, for time series analysis. By acquiring data within the abnormal time point, the changes in each vaginal canal area are tracked, such as changes in vaginal canal wall thickness, changes in lumen diameter, and overall shape deformation, and other dynamic characteristics, thereby extracting the corresponding dynamic change features.

[0128] S27, using a preset second time series model to respectively detect data difference values ​​between adjacent initial contours in each initial contour sequence, and associating relatively later initial contours with data difference values ​​greater than or equal to a preset tolerance value to form shape features of each vaginal canal region.

[0129] It should be noted that the initial contours at different times in each initial contour sequence are integrated through a second time series model such as an LSTM model, and whether the data difference value between the nth initial contour and the n-1th initial contour is greater than or equal to a preset tolerance value (such as 1e-4) is determined to determine whether there is a significant difference in the data. When there is a significant difference, it is determined that there is difference data between the two initial contours that does not belong to the same area. The difference data part, i.e., the nth initial contour, is saved and finally uniformly constructed into the shape features of the vaginal canal area.

[0130] S28, converting the shape features and dynamic change features of each vaginal canal region into the same dimension and then splicing them to output the comprehensive features of each vaginal canal region.

[0131] It should be noted that, through the integration of shape features and dynamic change features, comprehensive features reflecting the anatomical characteristics of each vaginal canal region can be formed.

[0132] Step 103, performing parametric modeling of the vaginal canal according to the comprehensive features to generate a target vaginal canal three-dimensional morphological model, performing lesion area segmentation on the target vaginal canal three-dimensional morphological model, and generating multiple initial lesion area images and corresponding initial lesion segmentation boundary curves.

[0133] Step 103 includes the following sub-steps:

[0134] S31. The comprehensive features are fused through the Poisson fusion algorithm to obtain a four-dimensional feature array.

[0135] It should be noted that the Poisson fusion algorithm is used to fuse all comprehensive features to fill the gaps in feature representation and complete the data. In a manner similar to MRI slice splicing, a four-dimensional feature array (x, y, z, t) is output, where xyz represents the coordinate system and t represents time.

[0136] S32, regarding the elements of the four-dimensional feature array as nodes in a minimum spanning tree, using a minimum tree generation algorithm to perform parameterization processing on the four-dimensional feature array, and outputting the geometric shape and topological structure of the vaginal canal.

[0137] It should be noted that parameterizing the four-dimensional feature array can be understood as constructing a patch with every three elements in a certain order to obtain the corresponding geometric shape and topological structure of the vaginal canal. This order can be determined by treating the elements as nodes in a minimum spanning tree using a minimum spanning tree algorithm. The parameterization method can be used to describe the geometric shape of the vaginal canal as being similar to a curved cylinder, whose parameters such as length and diameter can accurately represent the size of the vaginal canal. As for the topological structure, it can be described as a closed pipe with two openings.

[0138] S33, performing curve fitting on each comprehensive feature to obtain multiple smooth curves.

[0139] It should be noted that, through curve fitting technology, the data points in the comprehensive features can be connected into a smooth curve.

[0140] S34. Using various smooth curves, geometric shapes and topological structures, an initial vaginal canal three-dimensional morphological model is constructed based on surface reconstruction technology.

[0141] It should be noted that based on various smooth curves, geometric shapes and topological structures, a three-dimensional morphological model similar to the real vaginal canal can be constructed through surface reconstruction technology. Please refer to the prior art for details. In this initial three-dimensional morphological model of the vaginal canal, the subtle wrinkles of the vaginal canal wall and possible morphological variations can be clearly seen.

[0142] S35, enhancing the lesion region features of the initial vaginal canal 3D morphology model through self-calibration convolution to obtain a target vaginal canal 3D morphology model;

[0143] It should be noted that the initial vaginal canal 3D morphological model is input into the self-calibration convolution. When it is detected that the data features of a certain area are different from those of the surrounding area, the self-calibration convolution will focus on this area to capture its local structural changes and detailed features. If the color, texture and other features in a certain area are different from those of normal tissues, it may be a potential lesion area, and the self-calibration convolution will identify and mark the feature points of this area.

[0144] In addition, in order to improve the accuracy and robustness of modeling, the present embodiment can adopt an improved self-calibration convolution, that is, a self-calibration convolution that introduces batch normalization. By performing batch normalization after the convolution operation of the self-calibration convolution, it is ensured that the image features such as brightness and contrast in different regions remain consistent during the feature extraction process, thereby reducing the impact of deviations and improving the convergence speed of the convolution layer when processing complex forms. At the same time, the three-dimensional morphological model of the vaginal canal can also dynamically adjust the learning rate and the parameters of the self-calibration convolution layer according to different data. For example, when the lesion features are not obvious, the learning rate and convolution layer parameters are automatically adjusted to better adapt to local features and more accurately capture the subtle differences in the lesion area.

[0145] S36. Use a U-net network to segment the target vaginal canal three-dimensional morphological model into lesion regions, and determine a plurality of initial lesion region images and corresponding initial lesion region segmentation boundary curves.

[0146] It should be noted that when the target vaginal canal 3D morphological model is input into the U-net network, the U-net network will perform a preliminary segmentation of the lesion area according to the learned features, and obtain multiple initial lesion area images and corresponding initial lesion area segmentation boundary curves. The lesion area segmentation boundary curve can be understood as a curve that separates the lesion area from the normal tissue area and depicts the outline of the lesion.

[0147] Step 104: determining initial adaptive topological constraints based on each initial lesion region image, performing multi-directional collaborative analysis on each initial lesion region image under the constraints of the initial adaptive topological constraints, and determining optimized adaptive topological constraints.

[0148] Step 104 includes the following sub-steps:

[0149] S41, performing image segmentation on each initial lesion region image using a minimum graph cut algorithm to obtain a plurality of initial segmented lesion region images and a plurality of initial cut sets;

[0150] It should be noted that based on the initial lesion area image, the minimum graph cut algorithm is applied to calculate the minimum cut. Assuming that the size of the input image is 100x100 pixels, the minimum graph cut algorithm first constructs a weight matrix by calculating the weights between pixels. These weights can be calculated by the pixel grayscale difference. For example, the weight between a pixel and its neighboring pixels can be set as w(i,j)=exp(|I(i)-I(j)| 2 / σ 2 ), where σ is the adjustment parameter; through the minimum cut operation, the initial segmented lesion area image and the initial cut set are obtained. The cut set can be understood as a set of edges that divide the pixels of the image into the foreground and the background, mapping the potential topological structure in the image.

[0151] S42, performing image feature analysis on each initial segmented lesion region image and each initial cut set to determine initial adaptive topological constraints.

[0152] It should be noted that, based on the initial segmented lesion area image and the initial cut set, the texture and edge feature analysis of the image is performed to determine the initial adaptive topological constraints, such as requiring the connectivity and smoothness of the segmentation curve.

[0153] S43, performing feature extraction in different directions on each initial lesion region image under the constraint of the initial adaptive topological constraint, and outputting multiple direction analysis results of each initial lesion region image.

[0154] It should be noted that, through the multi-head convolutional neural network (MCNN) based on multiple parallel convolutional neural network layers, features of the lesion area image can be extracted from multiple different angles, such as horizontal, vertical, diagonal and other directions, or the direction can be set from 0 to 180 degrees in increments of 10 degrees. Its output is the direction analysis result after the feature extraction of each direction. These direction analysis results reflect information such as gradient changes in various directions;

[0155] When performing directional analysis, adaptive topological constraints can be used as a reference standard to ensure the rationality of the segmentation results. For example, adaptive topological constraints require that the segmentation curve has connectivity and smoothness. When the initial lesion area image is directional analyzed from 0 to 180 degrees in increments of 10 degrees, the adaptive topological constraint is like a "filter" to determine whether the results obtained in each direction meet the overall topological requirements. If the gradient change amplitude obtained in a certain direction looks reasonable locally, but causes the segmentation curve to be disconnected or not smooth, then it does not meet the adaptive topological constraint.

[0156] S44. If the directional analysis results in any same direction satisfy the preset consistency difference range, the initial adaptive topological constraint is used as a regularization term to construct the initial energy functional.

[0157] It should be noted that if the gradient change amplitudes between the analysis results in multiple directions in each direction are consistent within a certain range of difference, it is considered that the multi-directional synergy condition is met, and the initial adaptive topological constraint is integrated into the energy functional as a regularization term to construct the initial energy functional E(u)=Data(u)+λReg(u), where Data(u) represents the data term, Reg(u) represents the regularization term, λ is the weight parameter, and u represents the cut set.

[0158] S45. Determine the extreme value condition of the initial energy functional by using the variational method, and iteratively solve the initial energy functional according to the extreme value condition by using the gradient descent method, and output the optimized cut set.

[0159] It should be noted that the initial energy functional is processed by the variational method, and δE / δu=0 is set to obtain the extreme value condition of the optimal solution. When the segmentation requirements are met, the gradient descent method is used to calculate the gradient ∇E(u) and update the curve parameters to output the optimized cut set.

[0160] S46. Perform image segmentation on the target vaginal canal three-dimensional morphological model using a minimum graph cut algorithm according to the optimized cut set to obtain multiple intermediate lesion area images.

[0161] S47, using each initial lesion region image and the corresponding intermediate lesion region image to calculate the intersection-over-union ratio.

[0162] S48. If any intersection-and-union ratio is greater than the intersection-and-union ratio threshold, the initial adaptive topology constraint is used as the optimized adaptive topology constraint.

[0163] It should be noted that by comparing the initial lesion area image with the intermediate lesion area image, the intersection over union (IoU) and the number of connected areas are used to determine whether the segmentation accuracy is improved. If the IoU increases from 6 to 8, it means that the segmentation accuracy is significantly improved. Therefore, the initial adaptive topological constraint is used as the optimized adaptive topological constraint.

[0164] S49. If the directional analysis results in any same direction do not satisfy the preset consistency difference range or any intersection-and-union ratio is less than or equal to the intersection-and-union ratio threshold, the corresponding directional analysis results of each initial lesion area image are respectively input into the cross-dimensional interaction model for feature processing, and multiple optimized fused directional analysis results of each initial lesion area image are output.

[0165] S410, concatenating the analysis results of each optimized fusion direction to generate an optimized multi-directional collaborative result, and inputting the optimized multi-directional collaborative result into the attention model to output the direction weighted feature.

[0166] S411, weighting the direction weighted feature and the initial adaptive topology constraint to obtain a new initial adaptive topology constraint, and counting the number of updates in real time.

[0167] S412: If the number of updates is equal to a preset number threshold, the initial adaptive topology constraint is used as an optimized adaptive topology constraint.

[0168] It should be noted that, in this embodiment, the threshold of the number of iterative updates of the adaptive topology constraint can be determined according to the number of directions in the directionality analysis, for example, the number of directions is used as the threshold of the number of times.

[0169] S413. If the number of updates is less than the preset number threshold, determine multiple new directional analysis results of each initial lesion area image under the constraint of the initial adaptive topological constraint until any directional analysis results in the same direction meet the preset consistency difference range or the number of updates is equal to the preset number threshold.

[0170] It should be noted that if the preset consistency difference range is not satisfied between the directional analysis results in any direction, it is considered that the multi-directional coordination conditions are not met, and the adaptive topological constraints need to be corrected and updated through an iterative generation process in multiple specific directions; first, the directional analysis results of each initial lesion area image are input into the cross-dimensional interaction model for feature processing, and multiple optimized fusion directional analysis results of each initial lesion area image are output; then, each optimized fusion directional analysis result is spliced ​​to obtain a multi-directional coordination result. The multi-directional coordination result can be understood as the weight of the initial lesion area image extracted by the multi-head convolutional neural network. The optimized multi-directional coordination result The input attention model is input to adjust the weights for refinement, and the obtained directional weighted features are weighted with the initial adaptive topological constraints to update the new initial adaptive topological constraints, and the number of updates is counted in real time; if the number of updates is less than the preset number threshold, then jump to S43 to re-analyze the directionality of the initial lesion area image under the constraint of the new initial adaptive topological constraint, and if the multi-directional coordination condition is met, continue to jump to S44-S48 to output the optimized adaptive topological constraint, and if the multi-directional coordination condition is still not met until the number of updates is equal to the preset number threshold, then the current initial adaptive topological constraint is output as the optimized adaptive topological constraint;

[0171] If any of the intersection-and-union ratios calculated in S47 is greater than the intersection-and-union ratio threshold, the adaptive topology constraint needs to be corrected and updated, and the initial adaptive topology constraint is updated according to steps S49-S13 to output the optimized adaptive topology constraint.

[0172] Step 105: construct a priori curve according to each initial lesion segmentation boundary curve under the constraint of optimizing the adaptive topological constraint, and determine multiple priori segmentation curves according to the priori curve through an adaptive grid refinement strategy.

[0173] Step 105 includes the following sub-steps:

[0174] S51. Extract multiple discrete points from each initial lesion segmentation boundary curve, and generate a priori curves based on each discrete point using a Bezier curve under the constraint of optimizing the adaptive topological constraint.

[0175] It should be noted that for an initial lesion segmentation boundary curve consisting of a set of discrete points, an n-order (e.g., n=3) Bezier curve can be used to form a relatively smooth prior curve under the constraint of optimizing the adaptive topological constraint.

[0176] S52. Input the prior curve into the convolutional neural network and output the curve residual.

[0177] It should be noted that the user convolutional neural network or the self-calibrated convolutional neural network can predict the residual of the prior curve and output the curve residual.

[0178] S53, using the curve residual as a regularization term to correct the priori curve to obtain a calibration curve.

[0179] It should be noted that the curve residual can be used as a regularization term to adjust the shape of the prior curve, thereby obtaining a calibrated calibration curve.

[0180] S54, filtering the calibration curve, and outputting scale curves of multiple scales.

[0181] S55. Calculate the curvature of each scale curve to determine the curvature distribution of each scale.

[0182] It should be noted that the curvature distribution of multiple scales is determined by performing multi-scale curvature analysis on the calibration curve; first, the calibration curve is filtered by using Gaussian smoothing, bilateral filtering, wavelet transform, etc. to obtain scale curves of different scales, and then the curvature is calculated by solving the second-order derivative at each point of the scale curve to determine the curvature distribution.

[0183] S56, extracting high curvature points and inflection points from each curvature distribution, performing adaptive mesh refinement on the regions where each high curvature point and each inflection point are located in the calibration curve, and determining a refined curve.

[0184] It should be noted that the use of a multi-scale approach can strike a balance between details and global features, identify key points, i.e., high curvature points and inflection points, from the curvature distribution of each scale, and near these key points of the calibration curve, increase the degree of refinement of the original grid slices (for example, increase the grid density by 20% in the high curvature area) to apply an adaptive grid refinement strategy to obtain a refined curve. In this way, the density of local control points can be increased, thereby improving the expressiveness of the curve.

[0185] S57, using a minimum graph cut algorithm to perform image segmentation on the refined curve to determine a refined segmentation curve.

[0186] It should be noted that the minimum graph cut algorithm is used to segment the refinement curve into several parts based on the energy minimization principle to obtain a refinement segmentation curve, so as to achieve accurate boundary segmentation.

[0187] S58, using the priori curve and each refined segmentation curve to calculate errors respectively, and using multiple errors to form an error distribution.

[0188] It should be noted that by refining the segmentation curve, the deviation between the refined segmentation curve and the prior curve shape can be determined. The error analysis can use the mean square error (MSE) to evaluate the difference between the segmentation result and the prior shape to obtain the error distribution.

[0189] S59. When the error distribution does not satisfy the preset error condition, adaptive mesh refinement and control point density adjustment are performed in the refinement curve according to the error distribution, and a new refinement curve is determined until the error distribution satisfies the preset error condition, and the refined segmentation curve is determined as the priori segmentation curve.

[0190] It should be noted that the error distribution can display the error conditions of different areas in the refinement curve, helping to determine which areas need to be adjusted and the degree of adjustment. Therefore, the adaptive grid refinement and control point density adjustment of the refinement curve according to the error distribution can be achieved through the error feedback mechanism. For example, additional control points are added or the grid density is adjusted in the area where the error is greater than a specific threshold. After the adjustment, jump to S57-S59 and use the minimum graph cut algorithm again for image segmentation and error analysis. When the error distribution meets the preset error condition, the current refined segmentation curve is output as a priori segmentation curve to determine the refined segmentation curve. Through this processing process, the representation of the curve shape and the segmentation accuracy can be effectively improved.

[0191] Step 106: perform image segmentation on the target vaginal canal three-dimensional morphological model based on the prior curve and the prior curvature distribution of each prior segmentation curve, and perform multi-directional collaborative analysis under the constraint of optimizing the adaptive topological constraint to determine multiple target lesion area images.

[0192] Step 106 includes the following sub-steps:

[0193] S61. Perform multi-scale curvature analysis on the prior curve and each prior segmentation curve, and obtain a plurality of corresponding prior curvature distributions.

[0194] It should be noted that multi-scale curvature analysis refers to identifying and quantifying the curvature changes of a curve at different scales, including filtering the curve and calculating the curvature to construct a curvature distribution. For details, please refer to S54-S55. The multi-scale curvature analysis is used to determine the prior curve and multiple prior curvature distributions of each curve in each prior segmentation curve.

[0195] S62, fitting the priori curve and the priori curvature distribution of each priori segmentation curve respectively by using a spline curve fitting technology to obtain a plurality of spline curves;

[0196] It should be noted that, for the multi-scale curvature analysis results, spline curve fitting technology can be used to obtain the polynomial of the prior curvature distribution in three-dimensional space, and the geometric characteristics of the prior curvature distribution at different scales can be captured through the polynomial, and the prior curve and the spline curve of each curve in each prior segmentation curve can be output.

[0197] S63, fitting the curvature change rate of each spline curve to obtain a curvature change function, and constructing the curvature change function as a regularization term to obtain a curvature energy functional;

[0198] It should be noted that the corresponding curvature change rate can be determined according to the curvature of each spline curve. The curvature change function is fitted to each curvature change rate and integrated into the energy functional as a regularization term to obtain the curvature energy functional E = E_data + λE_curvature, where λ represents the regularization parameter used to balance the influence of the data term and the curvature term, E_curvature represents the curvature change function, and E_data represents the data term, which can optimize the fitting accuracy of the high curvature area.

[0199] S64. Perform image segmentation on the target vaginal canal three-dimensional morphological model based on the curvature energy functional by using a minimum graph cut algorithm to obtain multiple curvature lesion area images.

[0200] It should be noted that by minimizing the curvature energy functional, the minimum graph cut algorithm is used to re-segment the target vaginal canal three-dimensional morphological model to obtain multiple curvature lesion area images.

[0201] S65. Under the constraint of optimizing the adaptive topological constraint, extract features in different directions for each curvature lesion region image, and output multiple curvature direction analysis results for each curvature lesion region image.

[0202] S66. If the curvature direction analysis results in any same direction satisfy the preset consistency difference range, the optimized adaptive topological constraint is used as a regularization term to construct the target energy functional.

[0203] S67. If the directional analysis results in any same direction do not meet the preset consistency difference range, the curvature directional analysis results of each curvature lesion area image are respectively input into the cross-dimensional interaction model for feature processing, and multiple curvature fusion directional analysis results of each curvature lesion area image are output and spliced ​​to generate curvature multi-directional collaborative results.

[0204] It should be noted that if any directional analysis results in the same direction do not meet the preset consistency difference range, then taking the image of each curvature lesion area as a unit, the associated multiple curvature directional analysis results are used as the input features of the cross-dimensional interaction model, and the feature processing process of the cross-dimensional interaction model is used to perform permutation, combination, weighted splicing and feature extraction to output the curvature fusion directional analysis results, and the various curvature fusion directional analysis results are spliced ​​to obtain the curvature multi-directional collaborative results.

[0205] S68. Input the curvature multi-directional collaborative results into the attention model to obtain the curvature direction weighted features, weight them with the optimized adaptive topological constraints, obtain new optimized adaptive topological constraints, and count the number of optimization updates in real time.

[0206] S69. If the number of optimization updates is equal to a preset optimization number threshold, the new optimization adaptive topological constraint is used as a regularization term to construct a target energy functional.

[0207] S610. If the number of optimization updates is less than the preset optimization threshold, multiple new curvature direction analysis results of each curvature lesion area image are determined under the constraints of the optimization adaptive topological constraints until any curvature direction analysis results in the same direction meet the preset consistency difference range or the number of optimization updates is equal to the preset optimization threshold.

[0208] S611. Use a minimum graph cut algorithm to perform image segmentation on the target vaginal canal three-dimensional morphological model based on the target energy functional to obtain multiple optimized lesion area images.

[0209] S612. If the standard deviation of the curvature distribution of any optimized lesion area image does not meet the convergence condition, a new optimized adaptive topological constraint is determined according to the multi-directional collaborative results of the curvature of each curvature lesion area image until the standard deviation of the curvature distribution of any optimized lesion area image meets the preset convergence condition or the number of optimization updates is equal to the preset optimization number threshold, then the optimized lesion area image is determined as the target lesion area image.

[0210] It should be noted that convergence can be determined by calculating whether the standard deviation of the curvature distribution of the optimized lesion area image is less than 1. If the convergence state reaches the preset standard, the iterative adaptive topological constraint is stopped, and the final vaginal canal segmentation result, i.e., the target lesion area image, is obtained; if it has not converged, when the number of optimization updates is less than the preset optimization number threshold, the iterative adaptive topological constraint continues. When the number of optimization updates is equal to the preset optimization number threshold, the latest optimized lesion area image is output as the target lesion area image.

[0211] Step 107: Perform image enhancement rendering on each target lesion region image to convert it into a plurality of vaginal canal lesion visual images.

[0212] Step 107 includes the following sub-steps:

[0213] S71. Perform Gaussian smoothing filtering and morphological opening and closing operations on each target lesion region image, and output a plurality of corrected lesion region images.

[0214] It should be noted that the target lesion area image is subjected to Gaussian smoothing filtering to obtain a preliminary smoothed lesion area image. In specific implementation, a Gaussian function with a standard deviation of 5 can be used for convolution operation. If the Gaussian smoothing filtering result shows that the noise is still large, the Gaussian function parameters are adjusted to improve the noise suppression effect and obtain a new smooth lesion area image. The smooth lesion area image is subjected to morphological opening and closing operations. In specific implementation, a structural element with a radius of 3 can be used for morphological operations to improve the boundary and morphology of the segmentation result and obtain a corrected lesion area image.

[0215] S72. Calculate segmentation accuracy using each corrected lesion region image and the corresponding standard lesion region image.

[0216] It should be noted that the accuracy of segmentation is evaluated by calculating the segmentation accuracy based on the corrected lesion area image by comparing it with the standard lesion area image used as a benchmark. In specific implementation, the Dice coefficient or Jaccard index can be used for evaluation.

[0217] S73. If any segmentation accuracy does not meet the expected segmentation standard, each corrected lesion region image is used as a new target lesion region image until the segmentation accuracy meets the expected segmentation standard and is determined as the target vaginal canal segmentation lesion image.

[0218] It should be noted that if any segmentation accuracy does not meet the expected segmentation standard (such as 9), the corrected lesion area image is used as the new target lesion area image, and the process jumps to S71 to iteratively adjust the parameters of Gaussian filtering and morphological operations to obtain a more accurate corrected lesion area image.

[0219] In addition, by adding Gaussian noise and salt and pepper noise of different intensities to the target lesion area image, its adaptability to different noise and interference conditions can be evaluated to ensure that the collaborative segmentation method still maintains stable performance under the conditions of noise standard deviation of 5 and noise density of 0.2.

[0220] S74, converting each target vaginal canal segmentation lesion image into a vaginal canal lesion visual image by using image rendering technology.

[0221] It should be noted that the use of image rendering techniques such as pseudo-color coding or three-dimensional reconstruction can convert the target vaginal canal segmentation lesion image into an intuitive visual image, which is convenient for subsequent image observation and analysis, provides a visual imaging reference for exploring the disease characteristics and lesions related to the vaginal canal, and provides support for doctors in formulating appropriate treatment plans.

[0222] In an embodiment of the present invention, the medical image of the vaginal canal is preprocessed to ensure the clarity and integrity of the data. The target vaginal canal three-dimensional morphological model is reconstructed by capturing the target image features of the vaginal canal and parameterizing modeling, which is conducive to adapting to the dynamic change characteristics of the vaginal canal to accurately present the spatial structure and morphological characteristics of the vaginal canal. A collaborative segmentation mechanism of image segmentation, adaptive topological constraints, adaptive grid refinement strategy and curvature analysis is adopted to facilitate combining the dynamic change characteristics of the vaginal canal with topological constraints to achieve real-time response and adaptive adjustment to changes in the vaginal canal, improve the reliability of vaginal canal segmentation, and visualize the image of the target lesion area. The entire process is achieved through automated algorithms and image processing technology without the need for human intervention, thereby improving the efficiency of medical analysis and the accuracy of the results.

[0223] See also Figure 3 , Figure 3 A structural block diagram of a collaborative segmentation system for the vaginal canal provided in an embodiment of the present invention.

[0224] The present invention provides a collaborative segmentation system for vaginal canal, comprising:

[0225] An image processing module 301 is used to obtain a plurality of vaginal canal medical images according to a preset step length and perform preprocessing to determine a plurality of vaginal canal target images;

[0226] A feature extraction module 302 is used to perform multi-channel feature extraction and cross-dimensional feature interaction based on each vaginal canal target image, and output comprehensive features of multiple vaginal canal regions;

[0227] The preliminary segmentation module 303 is used to perform parametric modeling of the vaginal canal according to the comprehensive features to generate a target vaginal canal three-dimensional morphological model, perform lesion area segmentation on the target vaginal canal three-dimensional morphological model, and generate multiple initial lesion area images and corresponding initial lesion segmentation boundary curves;

[0228] The constraint optimization module 304 is used to determine the initial adaptive topological constraints based on each initial lesion region image, and perform multi-directional collaborative analysis on each initial lesion region image under the constraint of the initial adaptive topological constraints to determine the optimized adaptive topological constraints;

[0229] The curve optimization module 305 is used to construct a priori curve according to each initial lesion segmentation boundary curve under the constraint of optimizing the adaptive topological constraint, and determine a plurality of priori segmentation curves according to the priori curve through an adaptive grid refinement strategy;

[0230] The target segmentation module 306 is used to perform image segmentation on the target vaginal canal three-dimensional morphological model based on the prior curve and the prior curvature distribution of each prior segmentation curve, and perform multi-directional collaborative analysis under the constraint of optimized adaptive topological constraints to determine multiple target lesion area images;

[0231] The visualization module 307 is used to perform image enhancement rendering on each target lesion area image and convert it into a plurality of vaginal canal lesion visual images.

[0232] Optionally, the feature extraction module 302 is specifically used for:

[0233] Perform multi-channel feature extraction on each vaginal canal target image to obtain multiple channel features of each vaginal canal target image;

[0234] According to each vaginal canal target image, the associated channel features are sequentially input into the cross-latitude interaction model for feature processing, so as to obtain multiple vaginal canal change sensitive area images of each vaginal canal target image;

[0235] Cluster analysis algorithm is used to classify the images of each vaginal canal change sensitive area to determine multiple vaginal canal areas;

[0236] Performing edge detection on the vaginal canal change sensitive region image of each vaginal canal region according to an edge detection algorithm to obtain multiple initial contours of each vaginal canal region;

[0237] splicing the initial contours of each vaginal canal region to obtain an initial contour sequence of each vaginal canal region;

[0238] The time series features of each initial contour sequence are extracted respectively through a preset first time series model, and the dynamic change features of each vaginal canal area are output;

[0239] Using a preset second time series model to respectively detect data difference values ​​between adjacent initial contours in each initial contour sequence, and associating relatively later initial contours with data difference values ​​greater than or equal to a preset tolerance value to form shape features of each vaginal canal region;

[0240] The shape features and dynamic change features of each vaginal canal region are converted into the same dimension and then spliced ​​to output the comprehensive features of each vaginal canal region.

[0241] Optionally, the preliminary segmentation module 303 is specifically used for:

[0242] The comprehensive features are fused through the Poisson fusion algorithm to obtain a four-dimensional feature array;

[0243] The elements of the four-dimensional feature array are regarded as nodes in a minimum spanning tree, and the four-dimensional feature array is parameterized using a minimum tree generation algorithm to output the geometric shape and topological structure of the vaginal canal;

[0244] Perform curve fitting on each comprehensive feature to obtain multiple smooth curves;

[0245] Using various smooth curves, geometric shapes and topological structures, an initial vaginal canal three-dimensional morphological model was constructed based on surface reconstruction technology;

[0246] The lesion area features of the initial vaginal canal 3D morphology model are enhanced by self-calibration convolution to obtain the target vaginal canal 3D morphology model;

[0247] The U-net network is used to segment the lesion area of ​​the target vaginal canal three-dimensional morphological model, and multiple initial lesion area images and corresponding initial lesion area segmentation boundary curves are determined.

[0248] Optionally, the constraint optimization module 304 is specifically used for:

[0249] Using a minimum graph cut algorithm to segment each initial lesion region image, a plurality of initial segmented lesion region images and a plurality of initial cut sets are obtained;

[0250] Perform image feature analysis on each initial segmented lesion region image and each initial cut set to determine the initial adaptive topological constraints;

[0251] Under the constraints of the initial adaptive topological constraints, feature extraction is performed on each initial lesion region image in different directions, and multiple direction analysis results of each initial lesion region image are output;

[0252] If the preset consistency difference range is satisfied between the directional analysis results in any same direction, the initial adaptive topological constraint is used as a regularization term to construct the initial energy functional;

[0253] The variational method is used to determine the extreme value conditions of the initial energy functional, and the initial energy functional is iteratively solved according to the extreme value conditions through the gradient descent method, and the optimized cut set is output;

[0254] According to the optimized cut set, the minimum graph cut algorithm is used to segment the target vaginal canal three-dimensional morphological model to obtain multiple intermediate lesion area images;

[0255] The intersection-over-union ratio is calculated using each initial lesion region image and the corresponding intermediate lesion region image;

[0256] If any intersection-and-union ratio is greater than the intersection-and-union ratio threshold, the initial adaptive topology constraint is used as the optimized adaptive topology constraint;

[0257] If the directional analysis results in any of the same directions do not satisfy the preset consistency difference range or any intersection-over-union ratio is less than or equal to the intersection-over-union ratio threshold, the corresponding directional analysis results of each initial lesion area image are respectively input into the cross-dimensional interaction model for feature processing, and multiple optimized fusion directional analysis results of each initial lesion area image are output;

[0258] The optimized multi-directional collaborative results are generated by splicing the analysis results of each optimized fusion direction, and the optimized multi-directional collaborative results are input into the attention model to output the direction weighted features;

[0259] The directional weighted features are weighted with the initial adaptive topological constraints to obtain new initial adaptive topological constraints, and the number of updates is counted in real time;

[0260] If the number of updates is equal to the preset number threshold, the initial adaptive topology constraint is used as the optimized adaptive topology constraint;

[0261] If the number of updates is less than the preset threshold, multiple new directional analysis results of each initial lesion area image are determined under the constraints of the initial adaptive topological constraints until any directional analysis results in the same direction meet the preset consistency difference range or the number of updates is equal to the preset threshold.

[0262] Optionally, the curve optimization module 305 is specifically used for:

[0263] Extracting multiple discrete points from each initial lesion segmentation boundary curve, and generating a priori curves based on each discrete point using Bezier curves under the constraint of optimizing adaptive topological constraints;

[0264] Input the prior curve into the convolutional neural network and output the curve residual;

[0265] The curve residual is used as a regularization term to correct the prior curve and obtain the calibration curve;

[0266] Filtering the calibration curve to output scale curves of multiple scales;

[0267] Calculate the curvature of each scale curve and determine the curvature distribution of each scale;

[0268] Extract high curvature points and inflection points from each curvature distribution, perform adaptive mesh refinement on the regions where each high curvature point and each inflection point are located in the calibration curve, and determine a refinement curve;

[0269] The image is segmented by using the minimum graph cut algorithm to determine the refined segmentation curve;

[0270] The errors are calculated using the priori curve and each refined segmentation curve, and multiple errors are used to form an error distribution;

[0271] When the error distribution does not meet the preset error condition, adaptive mesh refinement and control point density adjustment are performed in the refinement curve according to the error distribution, and a new refinement curve is determined until the error distribution meets the preset error condition, and the refined segmentation curve is determined as the prior segmentation curve.

[0272] Optionally, the target segmentation module 306 is specifically used for:

[0273] Perform multi-scale curvature analysis on the prior curve and each prior segmentation curve, and obtain multiple prior curvature distributions accordingly;

[0274] The priori curve and the priori curvature distribution of each priori segmentation curve are fitted respectively by spline curve fitting technology to obtain multiple spline curves;

[0275] The curvature variation function is obtained by fitting the curvature variation rate of each spline curve, and the curvature variation function is used as a regular term to construct the curvature energy functional.

[0276] The target vaginal canal 3D morphological model is segmented by using the minimum graph cut algorithm based on the curvature energy functional to obtain multiple curvature lesion area images.

[0277] Under the constraint of optimizing the adaptive topological constraint, the features of each curvature lesion region image in different directions are extracted respectively, and multiple curvature direction analysis results of each curvature lesion region image are output;

[0278] If the preset consistency difference range is satisfied between the analysis results of curvature direction in any same direction, the optimized adaptive topological constraint is used as a regularization term to construct the target energy functional;

[0279] If the directional analysis results in any of the same directions do not meet the preset consistency difference range, the curvature directional analysis results of each curvature lesion area image are respectively input into the cross-latitude interaction model for feature processing, and multiple curvature fusion directional analysis results of each curvature lesion area image are output and spliced ​​to generate curvature multi-directional collaborative results;

[0280] The curvature direction weighted features obtained by inputting the curvature multi-directional coordination results into the attention model are weighted with the optimized adaptive topological constraints to obtain new optimized adaptive topological constraints, and the number of optimization updates is counted in real time;

[0281] If the number of optimization updates is equal to the preset optimization number threshold, the new optimization adaptive topological constraint is used as a regularization term to construct the target energy functional;

[0282] If the number of optimization updates is less than the preset optimization number threshold, multiple new curvature direction analysis results of each curvature lesion area image are determined under the constraint of the optimization adaptive topological constraint until any curvature direction analysis results in the same direction meet the preset consistency difference range or the number of optimization updates is equal to the preset optimization number threshold;

[0283] The minimum graph cut algorithm is used to segment the target vaginal canal 3D morphological model based on the target energy functional to obtain multiple optimized lesion area images.

[0284] If the standard deviation of the curvature distribution of any optimized lesion area image does not meet the convergence condition, a new optimized adaptive topological constraint is determined according to the multi-directional collaborative results of the curvature of each curvature lesion area image until the standard deviation of the curvature distribution of any optimized lesion area image meets the preset convergence condition or the number of optimization updates is equal to the preset optimization number threshold, then the optimized lesion area image is determined as the target lesion area image.

[0285] Optionally, the feature processing process of the cross-dimensional interaction model includes:

[0286] Perform full permutation and combination of multiple input features of the associated input to obtain multiple feature combinations;

[0287] Divide the input features in each feature combination into multiple weighted features and fusion features according to the sorting;

[0288] After weighted summing each weighted feature in each feature combination, channel splicing is performed with the corresponding fusion feature to obtain the splicing feature of each feature combination;

[0289] The key area features of each spliced ​​feature are extracted through the cross-dimensional attention mechanism to obtain multiple output features.

[0290] Optionally, the visualization module 307 is specifically used for:

[0291] Performing Gaussian smoothing filtering and morphological opening and closing operations on each target lesion area image, and outputting multiple corrected lesion area images;

[0292] The segmentation accuracy is calculated using each corrected lesion region image and the corresponding standard lesion region image;

[0293] If any segmentation accuracy does not meet the expected segmentation standard, each corrected lesion area image is used as a new target lesion area image until the segmentation accuracy meets the expected segmentation standard and is determined as the target vaginal canal segmentation lesion image;

[0294] Each target vaginal canal segmentation lesion image is converted into a vaginal canal lesion visual image through image rendering technology.

[0295] An embodiment of the present invention further provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the processor executes the steps of the collaborative segmentation method for the vaginal canal as described in any of the above embodiments.

[0296] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system and module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0297] In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0298] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A collaborative segmentation method for vaginal canal, characterized in that: include: Acquire multiple vaginal canal medical images according to a preset step length and perform preprocessing to determine multiple vaginal canal target images; Based on each of the vaginal canal target images, multi-channel feature extraction and cross-dimensional feature interaction are performed to output comprehensive features of multiple vaginal canal regions; Performing parametric modeling of the vaginal canal according to the comprehensive features to generate a target vaginal canal three-dimensional morphological model, performing lesion area segmentation on the target vaginal canal three-dimensional morphological model, and generating a plurality of initial lesion area images and corresponding initial lesion segmentation boundary curves; Determining initial adaptive topological constraints based on each of the initial lesion region images, performing multi-directional collaborative analysis on each of the initial lesion region images under the constraints of the initial adaptive topological constraints, and determining optimized adaptive topological constraints; Under the constraint of the optimized adaptive topological constraint, construct a priori curve according to each of the initial lesion segmentation boundary curves, and determine a plurality of priori segmentation curves according to the prior curve through an adaptive grid refinement strategy; Performing image segmentation on the target vaginal canal three-dimensional morphological model based on the prior curve and the prior curvature distribution of each prior segmentation curve, and performing multi-directional collaborative analysis under the constraint of the optimized adaptive topological constraint to determine multiple target lesion area images; The target lesion area images are subjected to image enhancement rendering and converted into a plurality of vaginal canal lesion visual images.

2. The collaborative segmentation method for the vaginal canal according to claim 1, characterized in that: The method of performing multi-channel feature extraction and cross-dimensional feature interaction based on each of the vaginal canal target images to output comprehensive features of multiple vaginal canal regions includes: Perform multi-channel feature extraction on each of the vaginal canal target images to obtain multiple channel features of each of the vaginal canal target images; According to each of the vaginal canal target images, the associated channel features are sequentially input into the cross-latitude interaction model for feature processing to obtain a plurality of vaginal canal change sensitive area images of each of the vaginal canal target images; Using a cluster analysis algorithm to classify the images of the sensitive areas of the vaginal canal changes, and determining a plurality of vaginal canal areas; Performing edge detection on the vaginal canal change sensitive region image of each of the vaginal canal regions according to an edge detection algorithm to obtain a plurality of initial contours of each of the vaginal canal regions; splicing the initial contours of the vaginal canal regions to obtain a sequence of initial contours of the vaginal canal regions; The time series features of each initial contour sequence are extracted respectively through a preset first time series model, and the dynamic change features of each vaginal canal area are output; Using a preset second time series model to respectively detect data difference values ​​between adjacent initial contours in each initial contour sequence, and associating relatively later initial contours with data difference values ​​greater than or equal to a preset tolerance value to form shape features of each of the vaginal canal regions; The shape features and dynamic change features of each vaginal canal region are converted into the same dimension and then spliced ​​to output the comprehensive features of each vaginal canal region.

3. The collaborative segmentation method for vaginal canal according to claim 1, characterized in that: The method of performing parametric modeling of the vaginal canal according to the comprehensive features to generate a target vaginal canal three-dimensional morphological model, performing lesion area segmentation on the target vaginal canal three-dimensional morphological model, and generating a plurality of initial lesion area images and corresponding initial lesion segmentation boundary curves includes: The comprehensive features are fused by a Poisson fusion algorithm to obtain a four-dimensional feature array; The elements of the four-dimensional feature array are regarded as nodes in a minimum spanning tree, and a minimum tree generation algorithm is used to perform parameterization processing on the four-dimensional feature array to output the geometric shape and topological structure of the vaginal canal; Performing curve fitting on each of the comprehensive features to obtain multiple smooth curves; Using the smooth curves, the geometric shapes and the topological structures, an initial vaginal canal three-dimensional morphological model is constructed based on surface reconstruction technology; Performing lesion area feature enhancement on the initial vaginal canal three-dimensional morphological model through self-calibration convolution to obtain a target vaginal canal three-dimensional morphological model; The U-net network is used to perform lesion region segmentation on the target vaginal canal three-dimensional morphological model, and a plurality of initial lesion region images and corresponding initial lesion region segmentation boundary curves are determined.

4. The collaborative segmentation method for vaginal canal according to claim 1, characterized in that: The step of determining the initial adaptive topological constraint based on each of the initial lesion region images, performing multi-directional collaborative analysis on each of the initial lesion region images under the constraint of the initial adaptive topological constraint, and determining the optimized adaptive topological constraint comprises: Using a minimum graph cut algorithm to perform image segmentation on each of the initial lesion region images to obtain a plurality of initial segmented lesion region images and a plurality of initial cut sets; Performing image feature analysis on each of the initial segmented lesion region images and each of the initial cut sets to determine initial adaptive topological constraints; Under the constraint of the initial adaptive topological constraint, feature extraction is performed on each of the initial lesion region images in different directions, and multiple direction analysis results of each of the initial lesion region images are output; If the directional analysis results in any same direction satisfy the preset consistency difference range, the initial adaptive topological constraint is used as a regularization term to construct an initial energy functional; Determine the extreme value condition of the initial energy functional by using a variational method, and iteratively solve the initial energy functional according to the extreme value condition by using a gradient descent method, and output an optimized cut set; Performing image segmentation on the target vaginal canal three-dimensional morphological model using a minimum graph cut algorithm according to the optimized cut set to obtain multiple intermediate lesion area images; Calculating an intersection-over-union ratio using each of the initial lesion region images and the corresponding intermediate lesion region images; If any of the intersection-and-union ratios is greater than the intersection-and-union ratio threshold, the initial adaptive topology constraint is used as the optimized adaptive topology constraint; If the directional analysis results in any of the same directions do not satisfy the preset consistency difference range or any of the intersection-and-union ratios is less than or equal to the intersection-and-union ratio threshold, the corresponding directional analysis results of each of the initial lesion area images are respectively input into the cross-dimensional interaction model for feature processing, and multiple optimized fusion directional analysis results of each of the initial lesion area images are output; The optimized fusion direction analysis results are concatenated to generate an optimized multi-directional collaborative result, and the optimized multi-directional collaborative result is input into the attention model to output a directional weighted feature; The directional weighted feature is weighted with the initial adaptive topological constraint to obtain a new initial adaptive topological constraint, and the number of updates is counted in real time; If the update times are equal to the preset times threshold, the initial adaptive topology constraint is used as the optimized adaptive topology constraint; If the update number is less than the preset number threshold, multiple new directional analysis results of each initial lesion area image are determined under the constraint of the initial adaptive topological constraint until any directional analysis results in the same direction meet the preset consistency difference range or the update number is equal to the preset number threshold.

5. The collaborative segmentation method for vaginal canal according to claim 1, characterized in that: The step of constructing a priori curves according to each of the initial lesion segmentation boundary curves under the constraint of the optimized adaptive topological constraint, and determining a plurality of priori segmentation curves according to the priori curves by an adaptive grid refinement strategy includes: Extracting a plurality of discrete points from each of the initial lesion segmentation boundary curves respectively, and generating a priori curves based on each of the discrete points using a Bezier curve under the constraint of the optimized adaptive topological constraint; Inputting the priori curve into a convolutional neural network and outputting a curve residual; Using the curve residual as a regularization term to correct the priori curve to obtain a calibration curve; filtering the calibration curve to output scale curves of multiple scales; Calculating the curvature of each scale curve to determine the curvature distribution of each scale; Extracting high curvature points and inflection points from each of the curvature distributions, and performing adaptive grid refinement on the regions where each of the high curvature points and each of the inflection points are located in the calibration curve to determine a refined curve; Using a minimum graph cut algorithm to perform image segmentation on the thinning curve to determine a thin segmentation curve; The priori curve is used to calculate errors respectively with each of the refined segmentation curves, and a plurality of the errors are used to form an error distribution; When the error distribution does not meet the preset error condition, adaptive grid refinement and control point density adjustment are performed in the refinement curve according to the error distribution, and a new refinement curve is determined until the error distribution meets the preset error condition, and the refined segmentation curve is determined as the priori segmentation curve.

6. The collaborative segmentation method for vaginal canal according to claim 1, characterized in that: The method of performing image segmentation on the target vaginal canal three-dimensional morphological model based on the prior curve and the prior curvature distribution of each prior segmentation curve, and performing multi-directional collaborative analysis under the constraint of the optimized adaptive topological constraint to determine multiple target lesion area images includes: Performing multi-scale curvature analysis on the prior curve and each of the prior segmentation curves to obtain a plurality of prior curvature distributions; The priori curve and the priori curvature distribution of each priori segmentation curve are fitted respectively by spline curve fitting technology to obtain a plurality of spline curves; The curvature change rate of each spline curve is used for fitting to obtain a curvature change function, and the curvature change function is used as a regularization term to construct a curvature energy functional; Performing image segmentation on the target vaginal canal three-dimensional morphological model based on the curvature energy functional by using a minimum graph cut algorithm to obtain a plurality of curvature lesion area images; Under the constraint of optimizing the adaptive topological constraint, performing feature extraction in different directions on each of the curvature lesion region images, and outputting a plurality of curvature direction analysis results of each of the curvature lesion region images; If the preset consistency difference range is satisfied between the analysis results of curvature direction in any same direction, the optimized adaptive topological constraint is used as a regularization term to construct the target energy functional; If the directional analysis results in any same direction do not meet the preset consistency difference range, the curvature directional analysis results of each of the curvature lesion area images are respectively input into the cross-latitude interaction model for feature processing, and multiple curvature fusion directional analysis results of each of the curvature lesion area images are output and spliced ​​to generate a curvature multi-directional collaborative result; The curvature direction weighted feature obtained by inputting the curvature multi-directional coordination result into the attention model is weighted with the optimized adaptive topological constraint to obtain a new optimized adaptive topological constraint, and the number of optimization updates is counted in real time; If the number of optimization updates is equal to a preset optimization number threshold, the new optimization adaptive topological constraint is used as a regularization term to construct a target energy functional; If the number of optimization updates is less than the preset optimization number threshold, multiple new curvature direction analysis results of each curvature lesion area image are determined under the constraint of the optimization adaptive topological constraint, until any curvature direction analysis results in the same direction meet the preset consistency difference range or the number of optimization updates is equal to the preset optimization number threshold; Using a minimum graph cut algorithm to perform image segmentation on the target vaginal canal three-dimensional morphological model based on the target energy functional to obtain a plurality of optimized lesion area images; If the standard deviation of the curvature distribution of any of the optimized lesion area images does not meet the convergence conditions, a new optimized adaptive topological constraint is determined based on the multi-directional collaborative results of the curvature of each of the curvature lesion area images until the standard deviation of the curvature distribution of any of the optimized lesion area images meets the preset convergence conditions or the number of optimization updates is equal to the preset optimization number threshold, then the optimized lesion area image is determined as the target lesion area image.

7. The collaborative segmentation method for the vaginal canal according to claim 2, 4 or 6, characterized in that: The feature processing process of the cross-dimensional interaction model includes: Perform full permutation and combination of multiple input features of the associated input to obtain multiple feature combinations; Dividing the input features in each of the feature combinations into a plurality of weighted features and fusion features according to the sorting; After weighted summing each weighted feature in each feature combination, channel splicing is performed with the corresponding fusion feature to obtain the splicing feature of each feature combination; The key area features of each spliced ​​feature are extracted through a cross-dimensional attention mechanism to obtain multiple output features.

8. The collaborative segmentation method for vaginal canal according to claim 1, characterized in that: The step of performing image enhancement rendering on each of the target lesion area images to convert them into a plurality of vaginal canal lesion visual images includes: Performing Gaussian smoothing filtering and morphological opening and closing operations on each of the target lesion area images, and outputting a plurality of corrected lesion area images; Calculating the segmentation accuracy using each of the corrected lesion region images and the corresponding standard lesion region images; If any of the segmentation accuracy does not meet the expected segmentation standard, each of the corrected lesion area images is used as a new target lesion area image until the segmentation accuracy meets the expected segmentation standard and is determined as the target vaginal canal segmentation lesion image; Each of the target vaginal canal segmentation lesion images is converted into a vaginal canal lesion visual image through image rendering technology.

9. A collaborative segmentation system for a vaginal canal, characterized in that: include: An image processing module, used for acquiring a plurality of vaginal canal medical images according to a preset step length and performing preprocessing to determine a plurality of vaginal canal target images; A feature extraction module, used for performing multi-channel feature extraction and cross-dimensional feature interaction based on each of the vaginal canal target images, and outputting comprehensive features of multiple vaginal canal regions; A preliminary segmentation module is used to perform parametric modeling of the vaginal canal according to each of the comprehensive features to generate a target vaginal canal three-dimensional morphological model, perform lesion area segmentation on the target vaginal canal three-dimensional morphological model, and generate a plurality of initial lesion area images and corresponding initial lesion segmentation boundary curves; A constraint optimization module, used for determining an initial adaptive topological constraint based on each of the initial lesion region images, performing a multi-directional collaborative analysis on each of the initial lesion region images under the constraints of the initial adaptive topological constraint, and determining an optimized adaptive topological constraint; A curve optimization module, used for constructing a priori curves according to each of the initial lesion segmentation boundary curves under the constraint of the optimization adaptive topological constraint, and determining a plurality of priori segmentation curves according to the prior curves through an adaptive grid refinement strategy; A target segmentation module, used for performing image segmentation on the target vaginal canal three-dimensional morphological model based on the prior curve and the prior curvature distribution of each prior segmentation curve, and performing multi-directional collaborative analysis under the constraint of the optimized adaptive topological constraint to determine multiple target lesion area images; The visualization module is used to perform image enhancement rendering on each of the target lesion area images and convert them into multiple vaginal canal lesion visual images.

10. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the collaborative segmentation method for the vaginal canal as described in any one of claims 1-8.

Images