Horse ovary follicle segmentation method and system

Abstract

The invention discloses a horse ovary follicle segmentation method and system, and relates to the technical field of image processing. Comprising the following steps: acquiring a preprocessed ultrasonic image of horse ovary follicles, and carrying out gray processing to obtain a gray image; a deep convolutional neural network DeepLabv3 + is combined with an SE module and multi-scale cavity convolution to carry out feature extraction on the grayscale image, and image features are obtained; combining a deformation level set method, a region growing algorithm and graph segmentation to obtain a fusion segmentation model; segmenting the extracted image features by using a fusion segmentation model to obtain a primary horse ovary follicle segmentation result; the primary horse ovary follicle segmentation result is obtained through morphological operation, and a final horse ovary follicle segmentation result is obtained. According to the method, the follicle region in the horse ovary ultrasound image with unclear follicle boundary and noise interference can be accurately identified and segmented.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method and system for segmenting horse ovarian follicles. Background Art

[0002] In the modern veterinary field, the reproduction and health monitoring of horses are very important. As a key organ in the horse reproduction system, the growth and development status of the internal follicles in the ovary are directly related to the reproductive ability of horses. To evaluate ovarian function and follicle status, detailed analysis of ovarian images is usually required.

[0003] However, traditional follicle segmentation methods mainly rely on manual operations. This method is not only time-consuming and laborious, but also easily affected by subjective factors, resulting in inaccurate and non-repeatable segmentation results. With the development of computer technology and image processing technology, intelligent segmentation methods based on image processing have gradually become a research hotspot. The application of advanced technologies such as deep learning and machine learning in the field of medical image processing provides new ideas for solving the above problems. In particular, the excellent performance of convolutional neural networks (CNNs) in image segmentation tasks makes it possible to intelligently segment horse ovarian follicles.

[0004] In actual ultrasound images, the contrast between follicles and surrounding tissues is often low, and there is noise interference, making it difficult to accurately identify and segment follicles. In addition, the shape and size of follicles vary greatly at different growth stages, and it is impossible to accurately identify and segment the follicle area. Summary of the Invention

[0005] In order to overcome the above-mentioned disadvantages of the prior art that it is difficult to accurately identify and segment due to unclear follicle boundaries and noise interference, the main object of the present invention is to provide a method and system for segmenting horse ovarian follicles.

[0006] To achieve the above object, the present invention adopts the following technical solutions. A method for segmenting horse ovarian follicles includes the following steps:

[0007] Obtain the ultrasound image of the preprocessed horse ovarian follicle, perform adaptive enhancement processing, and then convert it into a grayscale image to obtain the grayscale image of the horse ovarian follicle;

[0008] Select DeepLabv3+ as the basic network, add an SE module, and insert multi-scale atrous convolution into the ASPP module in the basic network to form a feature extraction module, and perform feature extraction on the grayscale image to obtain the image features of the horse ovarian follicle;

[0009] Build a fusion segmentation model to perform fusion segmentation on the image features of the equine ovarian follicles and obtain the segmentation result of primary equine ovarian follicles. Among them, the fusion segmentation model initially segments the image features of equine ovarian follicles by using region growing initialized with automatic seed points and graph cut respectively, and merges them by using weighted voting to obtain a preliminary segmentation map. The adjusted segmentation map is obtained by combining the deformable level set. The adjusted segmentation map is fed back to the region growing and graph cut algorithms for iteration. The seed points and growth criteria of region growing are updated through the adjusted segmentation map, and the edge weights and region partitioning in the graph cut algorithm are adjusted according to the adjusted segmentation map to obtain two iterated preliminary segmentation results. Until the optimal weight of weighted voting is determined according to cross-validation, the two preliminary segmentation results are weighted and voted by using the optimal weight to obtain the segmentation result of primary equine ovarian follicles;

[0010] Optimize the segmentation result of the primary equine ovarian follicles by using adaptive morphological operations to obtain the final image segmentation result of equine ovarian follicles.

[0011] Obtaining the segmentation result of primary equine ovarian follicles includes the following steps;

[0012] Use region growing initialized with automatic seed points to initially segment the image features of the equine ovarian follicles to obtain the first segmentation result;

[0013] Use graph cut to initially segment the image features of the equine ovarian follicles to obtain the second segmentation result;

[0014] Merge the first segmentation result and the second segmentation result by weighted voting to obtain a preliminary segmentation image;

[0015] Convert the preliminary segmentation map into the initial contour of the deformable level set method, and use the deformable level set method to adjust the preliminary segmentation image to obtain an adjusted segmentation map;

[0016] Feed the adjusted segmentation map back to the region growing and graph cut algorithms for iteration. Update the seed points and growth criteria of region growing through the adjusted segmentation map, adjust the edge weights and region partitioning in the graph cut algorithm according to the adjusted segmentation map, determine the optimal weight of weighted voting according to cross-validation, and weight and vote the two preliminary segmentation results by using the optimal weight to obtain the segmentation result of primary equine ovarian follicles.

[0017] The use of region growing initialized with automatic seed points to initially segment the image features of the equine ovarian follicles includes the following steps:

[0018] Automatically select appropriate seed points based on the image features of the equine ovarian follicles, and determine the criteria for region growing. The criteria include the gray value, texture, and gradient of the pixels of the equine ovarian follicle image;

[0019] Starting from the seed point, check whether the neighboring pixels around the seed point satisfy the growth criterion; when the neighboring pixels satisfy the criterion, merge them into the current region and use them as new seed points to expand the region until no more pixels can be merged, obtaining the first segmentation result.

[0020] The preliminary segmentation of the image features using graph cuts includes the following steps:

[0021] Convert the horse ovarian follicle image features into a graph model, where each pixel on the graph model corresponds to a node, and each pixel node is connected to its adjacent pixels to obtain connection edges. According to the similarity measure between nodes, obtain the weights of the connection edges and construct an adjacency matrix;

[0022] Based on the adjacency matrix, define the data term and smooth term of the energy function, set the graph cut parameters, and use the cut that minimizes the energy function to segment the horse ovarian follicle image features into the follicle region as the foreground and the background region as the background, obtaining the second segmentation result.

[0023] The obtaining of the horse ovarian follicle image features includes the following steps:

[0024] Normalize the grayscale image of the horse ovarian follicle to obtain the preprocessed grayscale image of the horse ovarian follicle;

[0025] Select DeepLabv3+ as the basic network, add the SE module, and insert dilated convolutions into the ASPP module in the basic network to obtain the feature extraction network;

[0026] Input the preprocessed grayscale image of the horse ovarian follicle into the feature extraction network to obtain the horse ovarian follicle image features.

[0027] The inputting of the preprocessed grayscale image of the horse ovarian follicle into the feature extraction network to obtain the horse ovarian follicle image features includes the following steps:

[0028] The preprocessed grayscale image of the horse ovarian follicle is subjected to feature extraction through the basic convolutional layer of the feature extraction network to obtain a low-level feature map, including edges, corners, and simple geometric shapes;

[0029] Use the SE module to perform channel-level calibration on the low-level feature map to obtain optimized low-level image features;

[0030] Using the dilated convolutional layers with different dilation rates and the global average pooling branch in the ASPP module, the optimized low-level image features are used to extract image features from different scales and capture the global context information of the entire image. Then, the results of the dilated convolutional layers with different dilation rates and the global average pooling branch are concatenated in the channel dimension to integrate the information from different scales and the global, obtaining enhanced image features, that is, obtaining the image features of the horse ovarian follicles.

[0031] The graph cut parameters include the labels of the foreground and background, the energy values of the foreground and background, the weight of the smooth term, the data term weight, the edge weight, and the iteration stop condition.

[0032] A system for a method of segmenting horse ovarian follicles, comprising:

[0033] A feature extraction module, configured to obtain the ultrasonic image of the preprocessed horse ovarian follicles, perform adaptive enhancement processing, convert it into a grayscale image, and obtain the grayscale image of the horse ovarian follicles; select DeepLabv3+ as the basic network, add the SE module, and insert multi-scale dilated convolutions into the ASPP module in the basic network to form a feature extraction module, and perform feature extraction on the grayscale image to obtain the image features of the horse ovarian follicles;

[0034] A segmentation model construction module, configured to construct a fusion segmentation model, and perform fusion segmentation on the image features of the horse ovarian follicles to obtain a primary horse ovarian follicle segmentation result;

[0035] A segmentation processing module, configured to input the image features of the horse ovarian follicles into the fusion segmentation model, and perform preliminary segmentation on the image features of the horse ovarian follicles by using region growing initialized with automatic seed points and graph cut respectively, and perform merging by using weighted voting to obtain a preliminary segmentation map, combine the deformable level set to obtain an adjusted segmentation map, feedback the adjusted segmentation map to the region growing and graph cut algorithms for iteration, update the seed points and growth criteria of the region growing according to the adjusted segmentation map, adjust the edge weights and region division in the graph cut algorithm according to the adjusted segmentation map, obtain two iterated preliminary segmentation results, until the optimal weight of the weighted voting is determined according to cross-validation, and the two preliminary segmentation results are weighted and voted by using the optimal weight to obtain a primary horse ovarian follicle segmentation result; optimize the primary horse ovarian follicle segmentation result by using adaptive morphological operations to obtain the final image segmentation result of the horse ovarian follicles.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. The present invention combines the DeepLabv3+ deep convolutional neural network with the SE module and the multi-scale dilated convolution module for feature extraction, which can effectively extract rich image features from the grayscale images of horse ovarian follicles, thereby improving the accuracy of image segmentation. This method can accurately distinguish the boundaries of horse ovarian follicles, reduce noise interference, and improve the reliability of the segmentation results. Moreover, this feature extraction module has strong versatility and compatibility and can adapt to different ovarian follicle image data. This method is not only applicable to the segmentation of horse ovarian follicles, but also can be extended to the segmentation problems of ovarian follicles of other types of animals or humans, with a wider scope of application.

[0038] 2. By performing multi-level fusion of the region growing algorithm and the graph cut algorithm, the present invention not only utilizes the flexibility of region growing, but also takes into account the global optimization ability of the graph cut algorithm in segmentation. Combined with the adjustment process of the deformable level set method, the segmentation accuracy and convergence speed are improved. In addition, the optimal weight is determined through cross-validation in the iterative optimization process, further enhancing the robustness and accuracy of the segmentation model.

[0039] 3. The present invention also optimizes the primary segmentation result using morphological operations, further refining the segmentation boundary and effectively avoiding the inaccurate segmentation or over-segmentation phenomena that occur in traditional methods. This optimization step ensures the high quality of the final segmentation result, especially suitable for processing horse ovarian follicle images with complex backgrounds and structures.

[0040] 4. The method of the present invention can quickly converge to the optimal solution through effective preliminary segmentation and graph cut optimization strategies. The refined segmentation boundary can accurately extract the target region, reducing the need for manual intervention and automatically adapting to the morphological and size changes of different horse ovarian follicles. Combined with the maximum flow minimum cut algorithm in graph cut, it ensures the globally optimal segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application.

[0042] Figure 1 is a schematic structural flow diagram of the present invention;

[0043] Figure 2 is an ultrasonic image of a stage 2 follicle of a horse ovary provided in an embodiment of the present invention;

[0044] Figure 3 is the segmentation result of the stage 2 follicle in an embodiment of the present invention;

[0045] Figure 4 is a schematic flow diagram of the fusion segmentation of the present invention;

[0046] Figure 5 It is a schematic flowchart for the present invention to obtain the first segmentation result;

[0047] Figure 6 It is a schematic flowchart for the present invention to obtain the second segmentation result. Specific embodiments

[0048] In the modern veterinary field, the reproduction and health monitoring of horses are very important. As a key organ in the horse reproduction system, the internal follicle growth and development status of the ovary are directly related to the reproductive ability of horses. To evaluate ovarian function and follicle status, it is usually necessary to conduct a detailed analysis of ovarian images.

[0049] With the development of computer technology and image processing technology, intelligent segmentation methods based on image processing have gradually become a research hotspot. The application of advanced technologies such as deep learning and machine learning in the field of medical image processing provides new ideas for solving the above problems. Especially the excellent performance of convolutional neural networks in image segmentation tasks makes it possible for the intelligent segmentation of horse ovarian follicles.

[0050] To solve the above problems, the present invention aims to provide a method for segmenting horse ovarian follicles. Refer to Figure 4 , including the following steps:

[0051] Obtain the ultrasonic image of the preprocessed horse ovarian follicles, perform adaptive enhancement processing, and then convert it into a grayscale image to obtain the grayscale image of the horse ovarian follicles;

[0052] Select DeepLabv3+ as the basic network, add the SE module, and insert multi-scale atrous convolution into the ASPP module in the basic network to form a feature extraction module, and perform feature extraction on the grayscale image to obtain the image features of the horse ovarian follicles;

[0053] Construct a fusion segmentation model, perform fusion segmentation on the image features of the horse ovarian follicles to obtain the primary segmentation result of the horse ovarian follicles; wherein, the fusion segmentation model respectively uses region growing initialized with automatic seed points and graph cut to perform preliminary segmentation on the image features of the horse ovarian follicles, and uses weighted voting for merging to obtain a preliminary segmentation map, combines the deformable level set to obtain an adjusted segmentation map, feeds the adjusted segmentation map back to the region growing and graph cut algorithms for iteration, updates the seed points and growth criteria of region growing according to the adjusted segmentation map, adjusts the edge weights and region partitioning in the graph cut algorithm according to the adjusted segmentation map to obtain two iterated preliminary segmentation results, until the optimal weight of weighted voting is determined according to cross-validation, and the two preliminary segmentation results are weighted and voted using the optimal weight to obtain the primary segmentation result of the horse ovarian follicles;

[0054] Optimize the segmentation result of the primary horse ovarian follicles using adaptive morphological operations to obtain the image segmentation result of the final horse ovarian follicles.

[0055] Use an ultrasound device to obtain an image of the horse ovary. Specifically, when segmenting the ultrasound image of the horse ovarian follicles, first perform adaptive enhancement processing on the obtained ultrasound image, and then perform grayscale conversion to obtain a grayscale image. Normalize the grayscale image to standardize the pixel values.

[0056] Input the preprocessed grayscale image into the feature extraction network. Through the forward propagation of the network, layer by layer extract the image features of the horse ovary to generate a feature map containing local and global information.

[0057] Refer to Figure 5 and Figure 6 Construct a fusion segmentation model. Specifically, respectively use region growing initialized by automatic seed points and graph cut to perform preliminary segmentation on the image features of the horse ovarian follicles, and use weighted voting for merging to obtain a preliminary segmentation map. Combine the deformable level set to obtain an adjusted segmentation map. Feed the adjusted segmentation map back to the region growing and graph cut algorithms for iteration. Update the seed points and growth criteria of region growing according to the adjusted segmentation map, and adjust the edge weights and region partitioning in the graph cut algorithm according to the adjusted segmentation map to obtain two preliminary segmentation results after iteration. Until the optimal weight of weighted voting is determined according to cross-validation, the two preliminary segmentation results are weighted and voted using the optimal weight to obtain the segmentation result of the primary horse ovarian follicles;

[0058] Automatically select appropriate seed points based on the characteristics of the horse ovarian follicle image, and determine the criteria for region growing. The criteria include the gray value, texture, and gradient of the pixels in the horse ovarian follicle image. Starting from the seed points, check whether the neighboring pixels around the seed points meet the growing criteria. When the neighboring pixels meet the criteria, merge them into the current region and use them as new seed points to expand the region until there are no more pixels to merge, obtaining the first segmentation result. Convert the characteristics of the horse ovarian follicle image into a graph model, where each pixel on the graph model corresponds to a node, and each pixel node is connected to its adjacent pixels to obtain connection edges. According to the similarity measure between the nodes, obtain the weights of the connection edges and construct an adjacency matrix. Based on the adjacency matrix, define the data term and smooth term of the energy function, set the graph cut parameters, and use the cut that minimizes the energy function to segment the characteristics of the horse ovarian follicle image into a follicle region as the foreground and a background region as the background, obtaining the second segmentation result. Combine the first segmentation result and the second segmentation result through weighted voting to obtain a preliminary segmentation image. Convert the preliminary segmentation map into the initial contour of the deformable level set method, and use the deformable level set method to adjust the preliminary segmentation result to obtain an adjusted segmentation map. Feed the adjusted segmentation map back to the region growing and graph cut algorithms, update the seed points and growing criteria of the region growing through the adjusted segmentation map, adjust the edge weights and region partitioning in the graph cut algorithm according to the adjusted segmentation map, obtain two preliminary segmentation results after iteration, determine the optimal weight of the weighted voting according to cross-validation, and use the optimal weight to perform weighted voting on the two preliminary segmentation results to obtain the primary horse ovarian follicle segmentation result.

[0059] Obtain the primary horse ovarian follicle segmentation result. Perform morphological operations on the primary segmentation result, such as removing small connected regions, smoothing edges, etc., to obtain the final horse ovarian follicle segmentation result.

[0060] The process of the entire horse ovarian follicle image segmentation method can be clearly seen, as well as how to obtain accurate segmentation results through multi-level fusion and optimization.

[0061] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0062] Example 1:

[0063] The present invention comprehensively utilizes deep learning and image processing technologies to ensure the high accuracy and effectiveness of image segmentation. The following is an example of horse ovarian follicle segmentation.

[0064] First, obtain the ultrasonic image of the horse ovarian follicle, perform gray processing on the ultrasonic image to obtain the gray image of the horse ovarian follicle. Use the deep convolutional neural network DeepLabv3+, combined with dilated convolution, to extract the features of the gray image in the ASPP module. This step is achieved through the forward propagation of the network, and dilated convolution helps to capture more extensive context information.

[0065] The preliminary segmentation of the extracted horse ovarian follicle feature images is performed using region growing algorithm and graph cut technology to obtain two sets of segmentation results. The two sets of preliminary segmentation results are merged through a voting system to form a preliminary segmentation map. Then, the preliminary segmentation map is used as the initial contour of the level set method to adjust the segmentation result and obtain a more refined adjusted segmentation map. The adjusted segmentation map is fed back to the region growing and graph cut algorithms for iterative optimization, and the optimal weight is determined through cross-validation to complete the construction of the fusion segmentation model.

[0066] The constructed fusion segmentation model is used to perform the ultimate segmentation on the feature images to obtain the segmentation results of primary horse ovarian follicles. Morphological operations such as erosion and dilation are performed on the primary segmentation results to clarify the boundaries and remove noise, and the final segmentation results of horse ovarian follicle images are obtained.

[0067] The final segmentation results are compared with the expert manual segmentation results to evaluate the accuracy and efficiency of the algorithm.

[0068] A detailed segmentation report is generated, including information such as segmentation accuracy, error rate, and visual effect of the segmentation boundary.

[0069] This embodiment can effectively improve the segmentation accuracy of horse ovarian follicles by combining the latest deep learning technology and traditional image processing methods, and is applicable to the veterinary clinical and research fields, especially suitable for processing large datasets to improve work efficiency.

[0070] Example 2:

[0071] To specifically illustrate the effectiveness and implementation details of the method, the following is a detailed example, referring to Figures 1 to 6 , which includes specific data and operation steps.

[0072] 100 cases of horse ovarian ultrasound images are selected, and the resolution of each image is 512×512 pixels.

[0073] All horse ovarian ultrasound images are grayscale processed, and histogram equalization technology is used to enhance the image contrast. Then, a pre-trained DeepLabv3+ model is used, combined with atrous convolution and ASPP module for feature extraction.

[0074] The network is trained on an NVIDIA Tesla V100 GPU, using the cross-entropy loss function, Adam optimizer, the learning rate is set to 0.001, the batch size is 16, and the training is iterated for 50 epochs.

[0075] The region growing algorithm and the graph cut algorithm initially process the extracted features of horse ovarian follicle images, generating two sets of segmentation results. The two sets of segmentation results are combined with the initial results through a weighted voting mechanism, and the deformed level set method is used to adjust the merged image to obtain an adjusted segmentation map. Then, the adjusted segmentation map is fed back to the region growing and graph cut algorithms. The seed points and growth criteria of the region growing are updated according to the adjusted segmentation map, and the edge weights and region partitioning in the graph cut algorithm are adjusted based on the adjusted segmentation map for iterative optimization to finely adjust the boundary. The fusion segmentation model undergoes 50 iterations of optimization, adjusting the edge weights and region partitioning in the graph cut algorithm each time to reduce the overall segmentation error.

[0076] Table 1 Performance Comparison

[0077] Method Precision Recall Specificity Fusion segmentation model 92％ 89％ 95％ Region growing algorithm 84％ 76％ 89％ Graph cut algorithm 87％ 80％ 91％

[0078] Compared with traditional single algorithms, such as separate region growing or graph cut, the fusion model performs better in processing complex and low-contrast images, especially in terms of accuracy and specificity.

[0079] The generated report details the parameter settings, result images, and performance metrics of each segmentation step.

[0080] This method has been successfully applied in multiple veterinary clinics for the monitoring and analysis of follicle development stages, improving the accuracy and efficiency of diagnosis.

[0081] Through this case, it can be seen that the application of deep learning combined with traditional image processing techniques in horse ovarian follicle segmentation can significantly improve the accuracy and efficiency of segmentation, especially suitable for the processing of large-scale datasets. This method not only optimizes the segmentation process but also ensures its reliability and practicality in actual applications through detailed data analysis and verification.

[0082] Example 3:

[0083] In this example, the size of the original ultrasound image of the 2nd stage horse ovarian follicle is 512x512 pixels, as Figure 2 shown, the image format is.png, and the image has been preprocessed to remove noise and enhance contrast.

[0084] The color image is converted into a single-channel grayscale image through a standard grayscale algorithm. When the color channels of the original image are RGB, the grayscale value is calculated using the common weighted average method:

[0085] I gray = 0.2989×R + 0.5870×G + 0.1140×B

[0086] The grayscale image of the 2nd stage follicle is a two-dimensional array of 512x512.

[0087] For the convenience of subsequent feature extraction, it is necessary to normalize the grayscale image of the secondary follicles so that the pixel value range is within [0, 1]. Then, using DeepLabv3+, a feature extraction module composed of multi-scale dilated convolutions is used to extract the grayscale image features. In this embodiment, ResNet-101 is used as the base network of DeepLabv3+. Multi-scale dilated convolutions with ratios of 6, 12, and 18 are inserted into the ASPP module to form the feature extraction module in this embodiment. The normalized grayscale image of the secondary follicles is input, and through forward propagation, a feature map of 512x512x256 is obtained. This feature map contains the local and global features of the equine ovarian follicles. The local features include edges and textures, and the global features include morphology and structure.

[0088] According to the grayscale value, color, texture, and gradient information of the image, a certain position in the follicle area of the image is selected as the seed point. The growth criterion is set that the pixel grayscale value is not much different from the seed point, and the grayscale value difference is less than the set threshold t = 5, and the pixel can be added to the current area. An initial seed point is selected at the center position of the equine ovarian follicle, with x = 250 and y = 250. The growth criterion is set as the image grayscale value, gradient, and color information. The threshold of the grayscale value is set to [0.4, 0.8], that is, pixels with grayscale values between 0.4 and 0.8 are selected for region growing. Starting from the seed point (250, 250), the neighboring pixels are checked and expanded, and it is checked whether the neighboring pixels meet the growth criterion. If the neighboring pixels meet the criterion, the pixel is added to the current area and continues to expand until no further expansion is possible, forming the first segmentation result. The first segmentation result obtained by region growing is a rough area containing the equine ovarian follicle, with pixel value 1 representing the follicle area and the background being 0.

[0089] The features of the equine ovarian follicle image are converted into a graph model. Each pixel on the graph model is regarded as a node in the graph, and there are connecting edges between the nodes. The weights of the connecting edges are calculated according to the similarity between pixels, such as grayscale difference, texture difference, etc. An adjacency matrix W is constructed, and the element W(i, j) represents the similarity between pixel i and pixel j.

[0090] The energy function includes a data term (the energy value of the foreground and background) and a smooth term (the edge smoothness connecting adjacent pixels), and weights are set. The formula is:

[0091]

[0092] where D i is the data term, representing the energy value of the foreground or background, and V ij is the smooth term, representing the similarity between adjacent pixels.

[0093] In this embodiment, the foreground energy value is 0.5, the background energy value is 1.0, the data item weight is 0.8, and the smoothing item weight is 0.2. Then, the minimum cut edge weight is used to optimize the characteristics of the horse ovarian follicle image, and the foreground and background of the characteristics of the horse ovarian follicle image are segmented to obtain the second segmentation result. At this time, the segmentation of the horse ovarian follicle region and the background region cut by graph cut is more refined.

[0094] By weighted merging of the results of region growing and graph cut algorithms, the final classification of each pixel point can be determined through a voting mechanism to form a preliminary segmentation map.

[0095] The preliminary segmentation map is converted into the initial contour of the deformable level set method. The initial contour is usually represented by a simple circle or ellipse for the follicle region. The deformable level set method is used to adjust the preliminary segmentation result, optimize the segmentation boundary, reduce noise and refine the shape. This process is usually carried out iteratively until convergence to obtain the adjusted segmentation map. Set the convergence condition. When the maximum number of iterations is 1000, calculate the error between the segmentation boundary and the true boundary after each iteration. If the error is less than the set threshold, stop the iteration. The adjusted segmentation map is fed back to the region growing and graph cut algorithms to update the seed points, region division criteria and edge weights in graph cut for iterative optimization.

[0096] Morphological operations such as closing and opening are performed on the preliminary segmentation result to remove noise and fill holes, and improve the coherence of the segmentation boundary. The closing operation is used to fill small holes inside the follicle. And the opening operation is used to remove small noise points in the segmentation result. Among them, the size of the structural element for closing and opening operations is 3x3.

[0097] Finally, the image obtained through morphological optimization will show a more accurate follicle segmentation result. Please refer to Figure 3 , and the result is represented as a binary image, where the follicle region is the foreground 1 and the background region is the background 0. The size of the final segmentation image is 512x512 and contains a clear follicle region.

[0098] It should be noted that in the present invention, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0099] The above embodiments are only illustrative examples of the present invention and do not constitute a limitation on the protection scope of the present invention. Any design identical or similar to the present invention falls within the protection scope of the present invention.

Claims

1. A method for segmenting horse ovarian follicles, characterized in that: The following steps are involved: Acquire the pre-processed ultrasonic image of the horse ovarian follicle, perform adaptive enhancement processing, and convert it into a grayscale image to obtain a grayscale image of the horse ovarian follicle; Select DeepLabv3+ as the basic network, add the SE module, and insert the multi-scale dilated convolution into the ASPP module in the basic network to form a feature extraction module, perform feature extraction on the grayscale image, and obtain the image features of the horse ovarian follicles; Constructing a fusion segmentation model, performing fusion segmentation on the image features of the horse ovarian follicles, and obtaining a primary horse ovarian follicle segmentation result; wherein the fusion segmentation model respectively uses region growing and graph cuts initialized with automatic seed points to perform preliminary segmentation on the image features of the horse ovarian follicles, and uses weighted voting to merge them to obtain a preliminary segmentation map, combines the deformation level set to obtain an adjusted segmentation map, feeds the adjusted segmentation map back to the region growing and graph cut algorithms for iteration, updates the seed points and growth criteria of the region growing by adjusting the segmentation map, adjusts the edge weights and region divisions in the graph cut algorithm according to the adjusted segmentation map, obtains preliminary segmentation results after two iterations, until the optimal weight of the weighted voting is determined according to cross-validation, and the two preliminary segmentation results are weighted voted using the optimal weight to obtain a primary horse ovarian follicle segmentation result; The primary horse ovarian follicle segmentation result is optimized by using adaptive morphological operation to obtain the final horse ovarian follicle image segmentation result.

2. The method for segmenting horse ovarian follicles according to claim 1, characterized in that: The method of obtaining the primary horse ovarian follicle segmentation result comprises the following steps: Performing preliminary segmentation on the horse ovarian follicle image features by using region growing initialized by automatic seed points to obtain a first segmentation result; Performing preliminary segmentation on the horse ovarian follicle image features by using graph cuts to obtain a second segmentation result; Merging the first segmentation result and the second segmentation result by weighted voting to obtain a preliminary segmented image; Converting the preliminary segmentation map into an initial contour of a deformed level set method, and adjusting the preliminary segmentation image using the deformed level set method to obtain an adjusted segmentation map; The adjusted segmentation map is fed back to the region growing and graph cut algorithms for iteration, the seed points and growth criteria of the region growing are updated through the adjusted segmentation map, the edge weights and region divisions in the graph cut algorithm are adjusted according to the adjusted segmentation map, the optimal weight of the weighted voting is determined according to cross-validation, and the two preliminary segmentation results are weighted voted using the optimal weight to obtain the primary horse ovarian follicle segmentation result.

3. The method for segmenting horse ovarian follicles according to claim 2, characterized in that: The method of performing preliminary segmentation of the horse ovarian follicle image features by region growing initialized by automatic seed points comprises the following steps: Automatically selecting appropriate seed points based on the horse ovarian follicle image features and determining a criterion for region growing, wherein the criterion includes grayscale values, textures and gradients of pixels of the horse ovarian follicle image; Starting from the seed point, check whether the neighborhood pixels around the seed point meet the growth criteria; when the neighborhood pixels meet the criteria, merge them into the current area and use them as new seed points to expand the area until there are no more pixels to merge, and obtain the first segmentation result.

4. The method for segmenting horse ovarian follicles according to claim 2, characterized in that: The method of performing preliminary segmentation of image features by using graph cuts comprises the following steps: Converting the horse ovarian follicle image features into a graph model, wherein each pixel on the graph model corresponds to a node, each pixel node is connected to its adjacent pixels to obtain connection edges, and according to the similarity measure between the nodes, the weights of the connection edges are obtained to construct an adjacency matrix; Based on the adjacency matrix, the data term and the smoothing term of the energy function are defined, the graph cut parameters are set, and the horse ovarian follicle image features are segmented into the follicle area as the foreground and the background area as the background by using the cut that minimizes the energy function to obtain the second segmentation result.

5. The method for segmenting horse ovarian follicles according to claim 1, characterized in that: The method of obtaining the image features of horse ovarian follicles comprises the following steps: Normalizing the grayscale image of the horse ovarian follicle to obtain a pre-processed grayscale image of the horse ovarian follicle; Select DeepLabv3+ as the base network, add the SE module, and insert the multi-scale dilated convolution into the ASPP module in the base network to obtain the feature extraction network; The preprocessed grayscale image of the horse ovarian follicle is input into the feature extraction network to obtain the image features of the horse ovarian follicle.

6. The method for segmenting horse ovarian follicles according to claim 5, characterized in that: The method of obtaining the image features of horse ovarian follicles comprises the following steps: The preprocessed grayscale image of the horse ovarian follicle is subjected to feature extraction through the basic convolutional layer of the feature extraction network to obtain a low-level feature map, including edges, corners, and simple geometric shapes; Using the SE module to perform channel-level calibration on the low-level feature map to obtain optimized low-level image features; The atrous convolution layers and global average pooling branches with different dilation rates in the ASPP module are used to extract image features from different scales of the optimized low-level image features and capture the global context information of the entire image. Then, the results of the atrous convolution layers and global average pooling branches with different dilation rates are spliced ​​in the channel dimension to integrate information from different scales and the global state to obtain enhanced image features, that is, the image features of horse ovarian follicles.

7. The method for segmenting horse ovarian follicles according to claim 3, characterized in that: The graph cut parameters include labels of foreground and background, energy values ​​of foreground and background, weights of smoothing items, weights of data items, edge weights and iteration stop conditions.

8. A system for segmenting horse ovarian follicles, characterized in that: include: A feature extraction module is used to obtain the pre-processed ultrasonic image of the horse ovarian follicle, convert it into a grayscale image after adaptive enhancement processing, and obtain the grayscale image of the horse ovarian follicle; DeepLabv3+ is selected as the basic network, the SE module is added, and the multi-scale hole convolution is inserted into the ASPP module in the basic network to form a feature extraction module, and the grayscale image is extracted to obtain the image features of the horse ovarian follicle; A segmentation model construction module is used to construct a fusion segmentation model, perform fusion segmentation on the image features of the horse ovarian follicles, and obtain a primary horse ovarian follicle segmentation result; The segmentation processing module is used to input the image features of the horse ovarian follicles into the fusion segmentation model, respectively use the region growing and graph cut initialized with automatic seed points to perform preliminary segmentation on the image features of the horse ovarian follicles, and use weighted voting to merge them to obtain a preliminary segmentation map, combine the deformation level set to obtain an adjusted segmentation map, feed the adjusted segmentation map back to the region growing and graph cut algorithms for iteration, update the seed points and growth criteria of the region growing by adjusting the segmentation map, adjust the edge weights and region division in the graph cut algorithm according to the adjusted segmentation map, obtain the preliminary segmentation results after two iterations, until the optimal weight of the weighted voting is determined according to cross-validation, and the two preliminary segmentation results are weighted voted using the optimal weight to obtain a primary horse ovarian follicle segmentation result; the primary horse ovarian follicle segmentation result is optimized by using adaptive morphological operations to obtain a final horse ovarian follicle image segmentation result.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method described in any one of claims 1 to 7 is implemented.

Images