Medical image segmentation device based on improved unet neural network and uterine body ultrasound image analysis method using the device

Abstract

The application discloses a medical image segmentation device based on an improved UNET neural network, which is based on an improved UNET USNET network and comprises multiple convolution subnets and an output convolution layer; the multiple convolution subnets have a convolution kernel with a size not less than 3*3, and the size gradually increases from input to output; each of the convolution subnets is provided with a pooling layer at the end, the maximum pooling operation step length of the pooling layer is set to 1, and no down-sampling and up-sampling is used; and the application further discloses a uterine body ultrasound image analysis method for performing ultrasound image segmentation by using the aforementioned USNET network. Different from a common UNET deep neural network, the convolution subnets are of the same structure, only the convolution kernel size and the operation size of the pooling layer are different, the number of the convolution subnets, the convolution kernel size and the convolution layer number can be adaptively adjusted without reducing the spatial resolution, the ultrasound image segmentation is realized, and thus the segmentation precision of the serosa and the endometrium in the ultrasound image is improved.

Description

Technical Field

[0001] This invention relates to a method for analyzing medical images, and more particularly to a medical image segmentation device based on an improved UNET neural network and a method for analyzing uterine ultrasound images using the device. Background Technology

[0002] Endometrial diseases are common female reproductive system disorders, including endometrial polyps, submucosal fibroids, typical or atypical endometrial hyperplasia, and endometrial cancer. Benign endometrial lesions have a high incidence rate, and the resulting symptoms such as spotting outside of menstruation, prolonged menstrual periods, and heavy menstrual flow severely impact women's daily lives. The five-year survival rate for stage I endometrial cancer can be as high as 90%, while the five-year survival rate for advanced metastatic patients is less than 20%. Therefore, early diagnosis and treatment of endometrial diseases are the most effective ways to reduce the harm caused by malignant endometrial tumors.

[0003] Ultrasound imaging of the uterine body is the most important tool for screening and diagnosing endometrial diseases. However, the accurate analysis of uterine body ultrasound images depends on the experience of the ultrasound physician. The analysis results of different ultrasound physicians or technicians vary greatly, which seriously affects the gynecologist's accurate diagnosis and treatment of endometrial diseases.

[0004] Deep learning algorithms are a type of machine learning method in the field of artificial intelligence. They contain multilayer perceptrons with multiple hidden layers. By combining low-level features to form more abstract high-level representations of attribute categories or features, they discover distributed feature representations of data. Deep learning methods have already proven their effectiveness in application fields such as image recognition and signal recognition, and can significantly improve the recognition accuracy of traditional methods.

[0005] Compared to CT and MRI images, endometrial ultrasound images have lower resolution and contrast. Existing deep learning networks such as HRNet are unable to effectively analyze uterine ultrasound images of patients with various endometrial diseases and obtain good results. Summary of the Invention

[0006] The purpose of this invention is to provide a medical image segmentation device based on an improved UNET neural network that can effectively process low-resolution, low-contrast medical images, and a method for analyzing uterine ultrasound images using this device.

[0007] To achieve the above objectives, the present invention provides a medical image segmentation device based on an improved UNET neural network. This device implements a USNET network (USNET is a custom name used in this invention) based on the improved UNET in hardware or software. The USNET network includes multiple convolutional subnetworks and an output convolutional layer. The multiple convolutional subnetworks have convolutional kernels with a size of not less than 3×3, and their size gradually increases from input to output. Each convolutional subnetwork ends with a pooling layer, and the maximum pooling operation stride of the pooling layer is set to 1, without using downsampling or upsampling.

[0008] Preferably, there are 3 convolutional subnets connected end to end; each convolutional subnet includes 6 convolutional layers and 1 pooling layer, and the kernel sizes of the convolutional layers of the 3 convolutional subnets are 3×3, 5×5 and 7×7 respectively; the output convolutional layer includes 1 convolutional layer with a kernel size of 1×1.

[0009] Preferably, the convolutional layers of each convolutional subnet are divided into 3 groups, and each group includes 1 to 4 convolutional layers.

[0010] Preferably, the USNET network also uses residual construction to connect different scales and extracts feature maps at different levels to obtain mixed semantic information for image segmentation, thereby accelerating the training process.

[0011] Preferably, the maximum pooling operation size of each pooling layer is equal to (the kernel size of the corresponding convolutional subnet + 1).

[0012] The present invention also provides a method for analyzing uterine ultrasound images using any of the aforementioned medical image segmentation devices, including an ultrasound image segmentation process for segmenting the uterine serosa and endometrium, the process comprising the following steps:

[0013] 1.1) Convert the ultrasound image to be segmented, which includes the uterine serosa and endometrium, into grayscale format;

[0014] 1.2) Input the obtained grayscale ultrasound image into the trained segmentation neural network to segment the uterine serosa and endometrium, specifically using any of the following methods:

[0015] a. A segmentation neural network A0 is used to simultaneously segment the uterine serosa and endometrium;

[0016] b. The uterine serosa is segmented using a segmentation neural network A1, and the uterine endometrium is segmented using a segmentation neural network A2;

[0017] All of the aforementioned segmentation neural networks employ the USNET network;

[0018] 1.3) Post-process the segmentation results to remove incorrectly segmented endometrial tissue.

[0019] Preferably, in step 1.2), the training of each segmentation neural network includes the following steps:

[0020] a. Input an ultrasound image containing a feature region, which is the uterine serosa and / or endometrium;

[0021] b. Extract feature regions by manually marking and drawing; then convert the image to grayscale format.

[0022] c. Divide the image data obtained in step b into a training dataset and a test dataset. Perform image enhancement processing on the training dataset, namely, rotation, mirroring, and illumination adjustment, to increase the amount of data in the training dataset.

[0023] d. Input the training dataset into the segmentation neural network for training;

[0024] e. Test the segmentation neural network trained in step d using the test dataset. If the test results do not meet the requirements, return to step a, increase the original training data, adjust the training parameters to improve the segmentation neural network, and retrain the segmentation neural network until the test results meet the requirements.

[0025] Preferably, in step 1.3), the post-processing specifically includes the following steps:

[0026] a. A digital morphological filter is used to perform erosion and dilation processes to remove erroneous segmentation results of the uterine serosal;

[0027] b. A digital morphological filter is used for erosion and dilation to remove erroneous endometrial segmentation results within the segmentation range of the uterine serosa.

[0028] Preferably, the method further includes an image measurement process for measuring the size of the endometrium or space-occupying tumor, which includes the following steps:

[0029] 2.1) Extract the outer contour of the endometrium from the endometrial segmentation results obtained in step 1.3);

[0030] 2.2) Find the point within the segmented endometrium that has the largest average distance from the outer contour of the endometrium, and use it as the center of the largest inscribed circle;

[0031] 2.3) The diameter of the largest inscribed circle is the size of the endometrium or space-occupying tumor (corresponding to non-cancerous endometrium and cancerous endometrium, respectively).

[0032] Preferably, the method further includes an automatic classification process for classifying ultrasound images of the endometrium, which includes the following steps:

[0033] 3.1) Convert the ultrasound image to be classified, which contains the endometrium, into grayscale format;

[0034] 3.2) Input the obtained grayscale ultrasound image into the trained classification neural network for image classification, specifically using any of the following methods:

[0035] a. A classification neural network B0 is used to simultaneously identify ultrasound images of normal endometrium, ultrasound images of endometrium containing polyps, and ultrasound images of endometrium containing cancer.

[0036] b. A classification neural network B1 is used to identify ultrasound images of normal endometrium and ultrasound images of diseased endometrium. For the latter, a classification neural network B2 is used to identify ultrasound images of endometrium containing polyps and ultrasound images of endometrium containing cancer.

[0037] c. A classification neural network B3 is used to identify ultrasound images of cancerous endometrium and non-cancer endometrium. For the latter, a classification neural network B4 is used to identify ultrasound images of normal endometrium and endometrium containing polyps.

[0038] Preferably, in step 3.2), each classification neural network uses a ResNet50 network. The ResNet50 network includes a residual loop and has better feature extraction capabilities and convergence characteristics. Its training process includes the following steps:

[0039] a. Input an ultrasound image containing the endometrium;

[0040] b. Manually label different categories of endometrial ultrasound images; and convert the images to grayscale format;

[0041] c. Divide the grayscale image data obtained in step b into a training dataset and a test dataset. Perform image enhancement processing on the training dataset, namely, image rotation, mirroring, and illumination adjustment, to increase the amount of data in the training dataset.

[0042] d. Use the training dataset to train the pre-trained classification neural network and introduce a loss function to overcome the classification error caused by data imbalance;

[0043] e. Test the classification neural network trained in step d using the test dataset. If the test results do not meet the requirements, return to step a, increase the original training data, improve the classification neural network by adjusting the training parameters, and retrain the classification neural network until the test results meet the requirements.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1) Based on the original UNET network, the USNET network uses multiple progressively larger kernels to extract image features, which can ensure that the network has the same size real receptive field as traditional convolutional neural networks, while avoiding the loss of spatial information in the processing results.

[0046] 2) The USNET network uses max-pooling to extract the maximum feature. The pooling step size is set to 1, so that downsampling and upsampling are no longer needed. The final feature maps have the same size, which facilitates the fusion of information at different scales.

[0047] 3) This invention also uses the USNET network to analyze uterine ultrasound images, and by reasonably combining the characteristics of endometrial ultrasound images, it can significantly improve the accuracy of ultrasound image analysis.

[0048] 4) For medical images with low resolution, such as ultrasound images, the spatial resolution of features would be greatly reduced if deep neural networks used the usual pooling operation. However, unlike the typical UNET deep neural network, the USNET network can achieve ultrasound image segmentation without reducing spatial resolution, thereby improving the segmentation accuracy of the serosal and endometrium in ultrasound images.

[0049] 5) In addition to ultrasound images of the uterus, the USNET network is also applicable to the segmentation and analysis of ultrasound images of other organs, as well as medical images such as MRI images, CT images, X-ray images and endoscopic images. It is significantly superior to existing neural networks such as HRNet, especially when processing low-resolution, low-contrast medical images. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the structure of the USNET network provided in Embodiment 1 of the present invention.

[0051] Figure 2 This is a schematic diagram of the training process (including preprocessing, testing and improvement) of the classification neural network in Embodiment 2 of the present invention.

[0052] Figures 3-5 These are schematic diagrams of the three classification methods provided in Example 2.

[0053] Figures 6-8 These are schematic diagrams illustrating the training process (including preprocessing, testing, and improvement) of the segmentation neural networks A0 to A2 provided in Example 3.

[0054] Figure 9 This is a schematic diagram illustrating the application of corrosion and expansion treatments to remove erroneous segmentation results in Example 3.

[0055] Figure 10 This is a schematic diagram of the measurement of the inner membrane thickness in Example 4.

[0056] Figure 11 This is a graph plotted in Example 4 with the axial coordinate as the horizontal axis and the inner membrane thickness (size) as the vertical axis.

[0057] Figure 12This is a flowchart of the ultrasound image analysis method for the uterine body provided in Example 5.

[0058] Wherein: Input layer 1, Convolutional subnetwork 2, Pooling layer 3, Feature layer 4 Figure 4 Output convolutional layer 5. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0060] Example 1

[0061] The existing UNET network is the main network used for medical image segmentation. However, the pooling operation used in UNET discards spatial information, resulting in the loss of low-level semantics. In addition, because UNET uses downsampling and upsampling, the spatial resolution of feature maps at different scales is different. Therefore, it cannot directly fuse feature maps of different scales for image segmentation, making it difficult to effectively process low-resolution, low-contrast medical images such as uterine ultrasound images.

[0062] The medical image segmentation device based on the improved UNET neural network provided in this embodiment implements the USNET network based on the improved UNET in hardware (using known hardware capable of implementing neural networks, such as general-purpose computers, Raspberry Pi, etc.) or software (generally stored in the form of a readable storage medium or running on a hardware carrier). Moreover, each convolutional subnet has the same structure, only the convolutional kernel size is different. When the resolution of the image to be processed changes, convolutional subnets with different kernel sizes can be easily combined to establish a new segmentation network.

[0063] The USNET network is an improvement upon the UNET network, consisting of one input layer (1), 19 convolutional layers (Conv), and three max-pooling layers (3). The specific structure and connections of each layer are as follows: Figure 1 As shown.

[0064] The 19 convolutional layers, from input to output, consist of 6 convolutional layers with a kernel size of 3×3, 6 convolutional layers with a kernel size of 5×5, 6 convolutional layers with a kernel size of 7×7, and 1 output convolutional layer with a kernel size of 1×1.

[0065] The above convolutional layers are divided into three convolutional sub-networks 2 (excluding the output convolutional layer) based on different kernel sizes. Each convolutional sub-network 2 has three groups of convolutional layers, with each group consisting of two convolutional layers. Figure 1 The output displayed between two adjacent convolutional layer groups is the output of the previous convolutional layer group, i.e., the features in the figure. Figure 4 .

[0066] The USNET network also uses residual construction to connect different scales and extracts feature maps at different levels to obtain mixed semantic information for image segmentation, thus accelerating the training process.

[0067] Each convolutional subnet 2 is finally equipped with a pooling layer 3. The maximum pooling size of each pooling layer 3 is equal to (the kernel size of the corresponding convolutional subnet + 1). In this embodiment, the input of the pooling layer 3 of the first convolutional subnet 2 is the output of a convolutional layer with a 3×3 kernel, and the maximum pooling size is 2×2; and so on, the maximum pooling size of the pooling layer of the second convolutional subnet 2 is 3×3, and the maximum pooling size of the pooling layer of the third convolutional subnet 2 is 4×4.

[0068] The improved USNET network has the following advantages:

[0069] a. By using kernels of sizes 3×3, 4×4, and 5×5 instead of just 3×3, a larger receptive field is obtained. Furthermore, by avoiding downsampling and upsampling, feature maps of uniform size are obtained, facilitating the fusion of feature maps at different scales for feature extraction and image segmentation.

[0070] b. Max pooling operations of 2×2, 3×3, and 4×4 are used to extract the features with the strongest amplitude. A step size of 1 is used, and the sub-images overlap during the pooling operation to ensure the continuity of spatial features.

[0071] c. Using residual construction, different scales are connected, and feature maps at different levels are extracted to obtain mixed semantic information for image segmentation, thus accelerating the training process.

[0072] One implementation of d.USNET has a total of approximately 0.35M parameters, which is much less than HRNet's 28.5M and also less than ResUNet's 8.98M. This effectively avoids the overfitting problem during network training.

[0073] The training method for the USNET network will be described later in conjunction with an example.

[0074] Example 2

[0075] This embodiment provides a method for analyzing uterine ultrasound images, including an automatic classification process, which includes the following steps:

[0076] 1) Convert the ultrasound images containing the endometrium to be classified into grayscale format;

[0077] 2) Input the obtained grayscale ultrasound image into the trained classification neural network for image classification, specifically using any of the following methods:

[0078] a. A classification neural network B0 was used to simultaneously identify normal endometrial ultrasound images, endometrial ultrasound images containing polyps, and endometrial ultrasound images containing cancer. (See...) Figure 3 ;

[0079] b. A classification neural network B1 was used to distinguish between normal endometrial ultrasound images and diseased endometrial ultrasound images. For the latter, a classification neural network B2 was then used to distinguish between endometrial ultrasound images containing polyps and those containing cancerous endometrium. (See...) Figure 4 ;

[0080] c. A classification neural network B3 was used to distinguish between ultrasound images of cancerous and non-cancer endometrial tissue. For the latter, a classification neural network B4 was then used to distinguish between ultrasound images of normal endometrium and endometrium containing polyps. (See...) Figure 5 .

[0081] In step 2) above, all classification neural networks use the ResNet50 network. For example... Figure 2 As shown, its training process (including preprocessing and test improvements) includes the following steps:

[0082] a. Input an ultrasound image containing the endometrium;

[0083] b. Manually annotate different categories of endometrial ultrasound images; ultrasound images usually also contain color Doppler signals that show blood flow. To avoid the influence of color signals on image segmentation, the images are converted to grayscale format; manual annotation is usually done before converting to grayscale format, but this is not absolute, and annotation can also be done after converting to grayscale format;

[0084] c. Divide the grayscale image data obtained in step b into a training dataset and a test dataset. Perform image enhancement processing on the training dataset, namely, image rotation, mirroring, and illumination adjustment, to increase the amount of data in the training dataset.

[0085] d. Use the training dataset to train the pre-trained classification neural network and introduce a loss function (focal loss) to overcome the classification error caused by data imbalance;

[0086] e. Test the classification neural network trained in step d using the test dataset. If the test results do not meet the requirements (overfitting or poor segmentation results occur), return to step a, increase the original training data, improve the classification neural network by adjusting the training parameters, and retrain the classification neural network until the test results meet the requirements.

[0087] f. The final trained classification neural network is obtained through the above steps.

[0088] The training process for classification neural networks B0 to B4 is largely the same, with the main difference being the different labeled objects. The training can be carried out by referring to the steps described above.

[0089] Example 3

[0090] This embodiment provides a method for analyzing ultrasound images of the uterus, including an ultrasound image segmentation process for segmenting the uterine serosa and endometrium. This process includes the following steps:

[0091] 1) Convert the ultrasound images to be segmented, which contain the uterine serosa and endometrium, into grayscale format;

[0092] 2) Input the obtained grayscale ultrasound image into the trained segmentation neural network to segment the uterine serosa and endometrium, specifically using any of the following methods:

[0093] a. A segmentation neural network A0 is used to simultaneously segment the uterine serosa and endometrium. The training process is described below. Figure 6 ;

[0094] b. The uterine serosa was segmented using segmentation neural network A1, and the endometrium was segmented using segmentation neural network A2. The training process is described in [link to training process]. Figure 7 , Figure 8 The advantage of using two neural networks, A1 and A2, for segmentation is that artifacts caused by excessive ultrasonic reflection outside the serous membrane layer are easily missegmented as the inner membrane. If the inner membrane is segmented first or simultaneously with the serous membrane, the segmentation accuracy of the inner membrane is low. If the serous membrane is segmented first, the segmentation accuracy of the inner membrane can be greatly improved because the serous membrane segmentation is more accurate.

[0095] All of the segmentation neural networks described herein use the USNET network described in Example 1;

[0096] 3) Post-process the segmentation results to obtain more accurate endometrial segmentation results. This specifically includes:

[0097] a. A digital morphological filter was used to perform erosion and dilation processes sequentially to remove erroneous segmentation results of the uterine serosal. (See...) Figure 9 When segmenting the slurry film, occasional missegmentation outside the slurry film layer may occur. However, the area of ​​the erroneously segmented slurry film is small. The erosion and dilation processing of the digital morphology filter can effectively remove the erroneously segmented area.

[0098] b. A digital morphological filter is used to perform erosion and dilation processes sequentially to remove erroneous endometrial segmentation results within the segmented area of ​​the uterine serosa. During endometrial segmentation, pseudo-endometrium may appear between the true endometrium and the serosa due to ultrasound artifacts. However, its area is small, and it can also be effectively removed using the erosion and dilation processes of the digital morphological filter.

[0099] like Figures 6-8 As shown, in step 2) above, the training (including preprocessing and testing improvements) of each segmentation neural network includes the following steps:

[0100] a. Input an ultrasound image containing a feature region, which is the uterine serosa and / or endometrium;

[0101] b. The boundaries of the serosal membrane and the intima are manually drawn by the ultrasound physician, and the characteristic regions are extracted. The ultrasound image usually also contains color Doppler ultrasound signals that show blood flow. In order to avoid the influence of color signals on image segmentation, the image is converted to grayscale format. Manual drawing is usually done before converting to grayscale format, but it is not absolute. It is also possible to convert to grayscale format and then perform manual drawing.

[0102] c. Divide the image data obtained in step b into a training dataset and a test dataset. Perform image enhancement processing on the training dataset, namely, rotation, mirroring, and illumination adjustment, to increase the amount of data in the training dataset.

[0103] d. Input the training dataset into the segmentation neural network for training;

[0104] e. Test the segmentation neural network trained in step d using the test dataset. If the test results do not meet the requirements (overfitting or poor segmentation results occur), return to step a, increase the original training data, adjust the training parameters to improve the segmentation neural network, and retrain the segmentation neural network until the test results meet the requirements.

[0105] f. The final trained segmentation neural network is obtained through the above steps.

[0106] The training process for segmentation neural networks A0 to A2 is largely the same, with the main difference being the different labeled objects. The training can be performed by referring to the steps described above.

[0107] Example 4

[0108] This embodiment, based on the uterine ultrasound image analysis method in Embodiment 3, further adds an image measurement process, such as... Figure 10 As shown, the process includes the following steps:

[0109] 1) In the endometrial segmentation results obtained in Example 3, the outer contour B of the endometrium was extracted;

[0110] 2) Search for the point within the segmented endometrium that has the largest average distance from the outer contour of the endometrium, and use it as the center of the largest inscribed circle;

[0111] 3) The diameter D of the largest inscribed circle C is the size of the endometrium or space-occupying tumor (according to the classification results of Example 2, corresponding to non-cancerous endometrium and cancerous endometrium respectively).

[0112] like Figure 11 As shown, with the central axis L of the intima (see...) Figure 10 A graph is plotted with the endometrial thickness (size) D on the vertical axis and the graph plotted with the x-axis. Based on the shape of the graph, the endometrium can be classified into the following three conditions:

[0113] (a) Normal, see Figure (a);

[0114] (b) Local abnormalities in endometrial thickness may indicate the presence of local polyps or benign or malignant tumors, as shown in Figure (b);

[0115] (c) Abnormal overall endometrial thickness may indicate the premenstrual bleeding phase or the presence of a wide distribution of benign or malignant tumors. This provides data support for further diagnosis and treatment, as shown in Figure (c).

[0116] Example 5

[0117] This embodiment, combining embodiments 2 to 4, provides a method for analyzing uterine ultrasound images, including processes such as automatic classification, ultrasound image segmentation, and image measurement.

[0118] like Figure 12 As shown, the method includes the following steps:

[0119] 1) According to the method in Example 2, the ResNet50 classification neural network is used to classify the ultrasound images of the uterine body;

[0120] 2) According to the method in Example 3, the segmentation neural network USNET is used to segment and post-process the endometrium and serosa to obtain the endometrial region after removing the incorrect segmentation results;

[0121] 3) The endometrial thickness was measured according to the method in Example 4, and this was used as a reference for further diagnosis and treatment.

Claims

1. A medical image segmentation device based on an improved UNET neural network, wherein the device implements a USNET network based on an improved UNET in hardware or software, characterized in that: The USNET network includes multiple convolutional subnetworks (2) and an output convolutional layer (5); The multiple convolutional subnets (2) have convolutional kernels with a size of not less than 3×3, and the size of the kernels gradually increases from the input to the output; Each of the convolutional sub-networks (2) ends with a pooling layer (3), the maximum pooling operation stride of the pooling layer (3) is set to 1, and downsampling and upsampling are not used; the feature maps output by each convolutional sub-network (2) are of the same size; The number of the convolutional subnets (2) is 3, which are connected end to end; each convolutional subnet includes 6 convolutional layers and 1 pooling layer (3), and the kernel sizes of the convolutional layers of the 3 convolutional subnets (2) are 3×3, 5×5 and 7×7 respectively; the output convolutional layer includes 1 convolutional layer with a kernel size of 1×1; Each convolutional subnetwork (2) has three groups of convolutional layers, each group including 1 to 4 convolutional layers; the USNET network also uses residual construction to connect different scales and extracts feature maps at different levels to obtain mixed semantic information for image segmentation, thus accelerating the training process. The maximum pooling operation size of each pooling layer (3) = (the kernel size of the corresponding convolutional subnet + 1) .

2. A method for analyzing uterine ultrasound images using the medical image segmentation device of claim 1, characterized in that: This includes an ultrasound image segmentation process for segmenting the uterine serosa and endometrium, comprising the following steps: 1.1) Convert the ultrasound image to be segmented, which includes the uterine serosa and endometrium, into grayscale format; 1.2) Input the obtained grayscale ultrasound image into the trained segmentation neural network to segment the uterine serosa and endometrium, specifically using any of the following methods: a. A segmentation neural network A0 is used to simultaneously segment the uterine serosa and endometrium; b. The uterine serosa is segmented using a segmentation neural network A1, and the uterine endometrium is segmented using a segmentation neural network A2; All of the aforementioned segmentation neural networks employ the USNET network; 1.3) Post-process the segmentation results to remove incorrectly segmented endometrial tissue.

3. The method for analyzing uterine ultrasound images according to claim 2, characterized in that: In step 1.2), the training of each segmentation neural network includes the following steps: a. Input an ultrasound image containing a feature region, which is the uterine serosa and / or endometrium; b. Extract feature regions by manually marking and drawing; then convert the image to grayscale format; c. Divide the image data obtained in step b into a training dataset and a test dataset. Perform image enhancement processing on the training dataset, namely, rotation, mirroring, and illumination adjustment, to increase the amount of data in the training dataset. d. Input the training dataset into the segmentation neural network for training; e. Test the segmentation neural network trained in step d using the test dataset. If the test results do not meet the requirements, return to step a, increase the original training data, adjust the training parameters to improve the segmentation neural network, and retrain the segmentation neural network until the test results meet the requirements.

4. The method for analyzing uterine ultrasound images according to claim 2, characterized in that: In step 1.3), the post-processing specifically includes the following steps: a. A digital morphological filter is used for erosion and dilation to remove erroneous segmentation results of the uterine serosal; b. A digital morphological filter is used for erosion and dilation to remove erroneous endometrial segmentation results within the segmentation range of the uterine serosa.

5. The method for analyzing uterine ultrasound images according to claim 4, characterized in that: The method also includes an image measurement process for measuring the size of the endometrium or space-occupying tumor, which includes the following steps: 2.1) Extract the outer contour of the endometrium from the endometrial segmentation results obtained in step 1.3); 2.2) Search for the point within the segmented endometrium that has the largest average distance from the outer contour of the endometrium, and use it as the center of the largest inscribed circle; 2.3) The diameter of the largest inscribed circle is the size of the endometrium or space-occupying tumor.

6. The method for analyzing uterine body ultrasound images according to any one of claims 2 to 4, characterized in that: The method also includes an automatic classification process for classifying ultrasound images of the endometrium, which includes the following steps: 3.1) Convert the ultrasound image to be classified, which contains the endometrium, into grayscale format; 3.2) Input the obtained grayscale ultrasound image into the trained classification neural network for image classification, specifically using any of the following methods: a. A classification neural network B0 is used to simultaneously identify ultrasound images of normal endometrium, endometrium containing polyps, and endometrium containing cancer; b. A classification neural network B1 is used to identify ultrasound images of normal endometrium and ultrasound images of diseased endometrium. For the latter, a classification neural network B2 is used to identify ultrasound images of endometrium containing polyps and ultrasound images of endometrium containing cancer. c. A classification neural network B3 is used to identify ultrasound images of cancerous endometrium and non-cancer endometrium. For the latter, a classification neural network B4 is used to identify ultrasound images of normal endometrium and endometrium containing polyps.

7. The method for analyzing uterine ultrasound images according to claim 6, characterized in that: In step 3.2), each classification neural network uses a ResNet50 network, and its training process includes the following steps: a. Input an ultrasound image containing the endometrium; b. Manually label different categories of endometrial ultrasound images; and convert the images to grayscale format or keep the original format; c. Divide the grayscale image data obtained in step b into a training dataset and a test dataset. Perform image enhancement processing on the training dataset, namely, image rotation, mirroring, and illumination adjustment, to increase the amount of data in the training dataset. d. Using the training dataset, train the pre-trained classification neural network and introduce a loss function to overcome the classification error caused by data imbalance; e. Test the classification neural network trained in step d using the test dataset. If the test results do not meet the requirements, return to step a, increase the original training data, improve the classification neural network by adjusting the training parameters, and retrain the classification neural network until the test results meet the requirements.

Images