Hysteroscopy image lesion segmentation method, system and device and medium

Abstract

The invention discloses a hysteroscopy image lesion segmentation method, system and device and a medium. The method comprises the following steps: constructing a hysteroscopy image training set and a hysteroscopy image test set; improving the UNet network to obtain an initial focus segmentation model; performing unsupervised training on the initial lesion segmentation model according to the training set to obtain a middle lesion segmentation model; performing supervised training on the middle lesion segmentation model according to the test set to obtain a target lesion segmentation model; obtaining a to-be-segmented hysteroscopy image, and performing feature extraction on the to-be-segmented hysteroscopy image to obtain a hysteroscopy image feature map; and inputting the feature map into a target lesion segmentation model, and outputting a lesion segmentation result. According to the invention, the segmentation effect of the focus segmentation model is greatly improved; therefore, the focus segmentation model can be trained by using the unlabeled data, and the effect of performing semi-supervised training by using a large amount of unlabeled data of a hospital is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image segmentation, and in particular to a method, system, device and medium for segmenting lesions in hysteroscopic examination images. Background Art

[0002] In recent years, with the development of computer science, artificial intelligence technology has been widely used in the medical field, mainly focusing on disease diagnosis, risk prediction, medical knowledge graph construction, electronic medical record data integration and medical image analysis. Among them, artificial intelligence technology has a significant effect in disease diagnosis, especially in the identification of lesions in medical image analysis.

[0003] Hysteroscopy is currently a very routine examination with visualization characteristics. As a minimally invasive examination method, hysteroscopy can directly observe the uterine cavity and study the panoramic uterine cavity. However, there are still problems that the results of hysteroscopy are very dependent on the personal level of the physician, the physician takes a long time to read the film, and the examination efficiency is low. Moreover, since the training of the semantic segmentation network requires pixel-level labels, and the annotation of medical images requires professionally trained physicians, it is very difficult to collect a large number of medical image semantic segmentation data sets. After years of accumulation in the hospital's database system, a large amount of data that has not been labeled by professional physicians is often saved. Therefore, how to use this unlabeled data information for semi-supervised machine learning is of great research significance. Summary of the invention

[0004] The purpose of the present invention is to perform artificial intelligence image lesion segmentation on hysteroscopic images to obtain lesion areas, thereby intelligently assisting the hysteroscopic examination diagnosis results. In order to achieve the above purpose, the present invention provides a hysteroscopic examination image lesion segmentation method, system, device and medium.

[0005] In a first aspect, an embodiment of the present invention provides a method for segmenting lesions in a hysteroscopic examination image, comprising:

[0006] Construct a hysteroscopic image training set and a hysteroscopic image test set;

[0007] The UNet network is improved to obtain an initial lesion segmentation model, wherein the improvement includes replacing each layer of the encoder branch of the UNet network with a target layer and adding a decoder branch so that two different decoder branches share the encoder branch, wherein the target layer includes two branches, wherein one branch includes three first convolution blocks with a convolution kernel size of 3×3 connected in sequence, and the other branch includes a second convolution block with a convolution kernel size of 1×1, and the output results of the two branches pass through another second convolution block together;

[0008] Performing unsupervised training on the initial lesion segmentation model according to the training set to obtain an intermediate lesion segmentation model;

[0009] Performing supervised training on the intermediate lesion segmentation model according to the test set to obtain a target lesion segmentation model;

[0010] Acquire a hysteroscopic examination image to be segmented, and perform feature extraction on the hysteroscopic examination image to be segmented to obtain a hysteroscopic examination image feature map;

[0011] The feature map is input into the target lesion segmentation model, and a lesion segmentation result is output, where the lesion segmentation result is a weighted value of the output results of the two decoder branches.

[0012] Preferably, the step of constructing a hysteroscopic image training set and a hysteroscopic image test set comprises:

[0013] Multiple hysteroscopic images were obtained from the hospital database;

[0014] screening unlabeled images in the hysteroscopic images to form a hysteroscopic image training set;

[0015] The labeled images in the hysteroscopic examination images are screened to form a hysteroscopic examination image test set.

[0016] Preferably, the first convolution module includes a convolution layer with a convolution kernel size of 3×3 and a normalization layer.

[0017] Preferably, the second convolution module includes a convolution layer with a convolution kernel size of 1×1, a normalization layer and an activation layer.

[0018] Preferably, the two decoder branches are any two decoder branches in a convolutional neural network set, and the convolutional neural network set includes UNet, VNet, DeConvNet, SegNet, UNet++ and DeepLabv3++.

[0019] Preferably, performing unsupervised training on the initial lesion segmentation model according to the training set to obtain an intermediate lesion segmentation model comprises:

[0020] Performing data enhancement on the training set to obtain an enhanced training set;

[0021] Characterizing the Dice loss function as the loss function of the initial lesion segmentation model to minimize the difference in lesion segmentation results output by the two decoder branches;

[0022] The initial lesion segmentation model is unsupervisedly trained according to the enhanced training set until the loss function is less than a first preset threshold and a preset number of training times is reached, thereby obtaining an intermediate lesion segmentation model.

[0023] Preferably, the performing supervised training on the intermediate lesion segmentation model according to the test set to obtain a target lesion segmentation model comprises:

[0024] The intermediate lesion segmentation model is supervised and trained according to the test set until the prediction accuracy of the lesion segmentation result is higher than a second preset threshold, thereby obtaining a target lesion segmentation model.

[0025] In a second aspect, an embodiment of the present invention provides a hysteroscopic examination image lesion segmentation system, comprising:

[0026] A data set construction module, used to construct a hysteroscopy image training set and a hysteroscopy image test set;

[0027] An initial model building module, used for improving the UNet network to obtain an initial lesion segmentation model, wherein the improvement includes replacing each layer of the encoder branch of the UNet network with a target layer and adding a decoder branch so that two different decoder branches share the encoder branch, wherein the target layer includes two branches, wherein one branch includes three first convolution blocks with a convolution kernel size of 3×3 connected in sequence, and the other branch includes a second convolution block with a convolution kernel size of 1×1, and the output results of the two branches pass through another second convolution block together;

[0028] An intermediate model construction module, used for performing unsupervised training on the initial lesion segmentation model according to the training set to obtain an intermediate lesion segmentation model;

[0029] A target model construction module, used for performing supervised training on the intermediate lesion segmentation model according to the test set to obtain a target lesion segmentation model;

[0030] A feature extraction module is used to obtain a hysteroscopic examination image to be segmented, and perform feature extraction on the hysteroscopic examination image to be segmented to obtain a hysteroscopic examination image feature map;

[0031] The lesion segmentation module is used to input the feature map into the target lesion segmentation model and output a lesion segmentation result, wherein the lesion segmentation result is a weighted value of the output results of the two decoder branches.

[0032] In a third aspect, an embodiment of the present invention provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the hysteroscopic image lesion segmentation method as described above when executing the computer program.

[0033] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the hysteroscopic image lesion segmentation method as described above.

[0034] Compared with the prior art, the embodiments of the present invention provide a method, system, device and medium for segmenting lesions in hysteroscopic examination images, and have the following beneficial effects: by adjusting the convolutional network structure of the encoder branch, the lesion segmentation model can fully learn low-level local and detail information, enrich the learning information of each layer, and reduce the number of parameters of the lesion segmentation model; at the same time, the depth of the lesion segmentation model is increased, redundant calculations are reduced, and the segmentation effect of the lesion segmentation model is greatly improved; by setting two different decoder branches to utilize more image features, and by minimizing the difference in segmentation results of the two decoder branches, the lesion segmentation model can be trained using unlabeled data, thereby achieving the effect of semi-supervised training using a large amount of unlabeled data in the hospital; and by sharing the encoder branch, computer resources can be saved. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of a method for segmenting lesions in a hysteroscopic examination image according to an embodiment of the present invention;

[0036] Figure 2 is a schematic diagram of a process of constructing a data set according to an embodiment of the present invention;

[0037] Figure 3 is a schematic diagram of a convolutional network structure modified by an embodiment of the present invention;

[0038] Figure 4 is a schematic diagram of a process for constructing an intermediate lesion segmentation model according to an embodiment of the present invention;

[0039] Figure 5 It is a structural schematic diagram of a hysteroscopic examination image lesion segmentation system according to an embodiment of the present invention;

[0040] Figure 6 It is a schematic diagram of the structure of a terminal device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0042] like Figure 1 As shown, an embodiment of the present invention provides a method for segmenting lesions in hysteroscopic examination images, comprising the steps of:

[0043] S1. Construct a hysteroscopy image training set and a hysteroscopy image test set;

[0044] Specifically, Figure 2 As shown, step S1 includes:

[0045] S101, obtaining a plurality of hysteroscopic examination images from a hospital database;

[0046] This embodiment obtains 3,000 hysteroscopic examination images stored in the hospital database.

[0047] S102, screening unlabeled images in hysteroscopic examination images to form a hysteroscopic examination image training set;

[0048] Most hospitals do not have enough manpower to perform a large number of image reading and annotation, so a large amount of unlabeled data is stored in the hospital database. In this embodiment, unlabeled images in 3,000 hysteroscopic images are screened to form a hysteroscopic image training set.

[0049] S103, screening labeled images in hysteroscopic examination images to form a hysteroscopic examination image test set.

[0050] In this embodiment, labeled images are screened from 3,000 hysteroscopic images to form a hysteroscopic image test set.

[0051] S2, improve the UNet network to obtain the initial lesion segmentation model;

[0052] Hysteroscopic images may be affected by bleeding, tissue reflections, or lens blur, which may reduce image quality and increase the difficulty of lesion segmentation. Some lesions may be similar to the surrounding normal tissues in color and texture, making it difficult for conventional automatic segmentation algorithms to distinguish. The UNet network has achieved good results in medical image segmentation with its bilaterally symmetrical encoder-decoder structure.

[0053] In the encoder stage, the main operation of each layer of the UNet network is: first, a convolution operation with a convolution kernel size of 3×3 is performed, and then the ReLU activation function is used to make the output result nonlinear, which is called a convolution block. Since no padding operation is performed during the convolution operation, the width and height of the feature map will be reduced by 2 after each convolution. The encoder structure uses convolution operations and downsampling operations to enable the network to learn deeper features step by step. When extracting features, the UNet network has a total of five layers. The first layer includes two convolution block operations; the second to fifth layers first downsample the output feature map of the previous layer, and the size of the feature map will become half of the original size, and then two convolution block operations are performed, a total of 4 downsampling times. In the decoder stage, the operations performed correspond to those in the encoder stage, and a total of 4 deconvolution operations are performed for upsampling. Each layer of the decoder first undergoes an upsampling, and the size of the feature map will be expanded to twice the original size. Before the convolution operation, the UNet network will also use a long jump connection to splice the feature maps of the same size obtained in the encoder stage and the decoder stage on the channel. In order to ensure the consistency of the number of channels in the encoder stage and the decoder stage, the number of channels is reduced to half of the original number each time an upsampling operation is performed, and then the number of channels is spliced ​​to utilize the shallow information features.

[0054] In each layer of the UNet network encoder, only two convolutional layers with a convolution kernel size of 3×3 were used in succession to obtain the output of each layer. However, different hysteroscopic equipment and the technical level of the operator will also affect the quality of the image and the recognizability of the lesions. For tiny polyps or early cancers, due to their small size, it may be difficult to clearly identify and accurately segment them in the image, and the lesions in the uterine cavity are diverse, including polyps, tumors, adhesions, and inflammation. It is easy to ignore targets with a small proportion during the UNet network learning process, and the segmentation accuracy is not high for different types of lesions. Therefore, step S2 improves the traditional UNet network.

[0055] Specifically, the improvement includes replacing each layer of the encoder branch of the UNet network with the target layer and adding a decoder branch so that two different decoder branches share the encoder branch. The target layer includes two branches, one of which includes three first convolution blocks with a convolution kernel size of 3×3 connected in sequence, and the other includes a second convolution block with a convolution kernel size of 1×1. The output results of the two branches pass through another second convolution block together. For details, see Figure 3 Furthermore, the first convolution module includes a convolution layer with a convolution kernel size of 3×3 and a normalization layer, and the second convolution module includes a convolution layer with a convolution kernel size of 1×1, a normalization layer, and an activation layer.

[0056] In the encoder stage, in order to enable the lesion segmentation model to fully learn the low-level local and detail information of the image in training, each layer of the encoder in the UNet network is modified as follows Figure 3 The convolutional network structure shown in Figure 1. The convolutional network structure performs two operations: one is to first perform a convolution operation with a convolution kernel size of 3×3, and then pass through a normalization layer, which is called the first convolution block; the other is to first perform a convolution operation with a convolution kernel size of 1×1, and then pass through a normalization layer, and finally use the ReLU activation function through the activation layer to make the output result nonlinear, which is called the second convolution block. Specifically, after the feature map is input, it is divided into two branches for convolution operation. One branch passes through three consecutive first convolution blocks to make the network's receptive field reach 7; the other branch passes through a second convolution block to adjust the number of output channels, and retains the original image features as a residual structure. Moreover, the output obtained by each first convolution block is spliced ​​in dimension, and then passes through a second convolution block together with the output of the second convolution block, and then enters the next layer. Such an encoder network structure increases the reuse of features, so that the lesion segmentation model can fully learn the local and detail features in the hysteroscopic examination image, and on this basis, reduce redundant calculations. Therefore, in the encoder stage, by modifying the convolutional network structure of the encoder, the lesion segmentation model can fully learn the low-level local and detail information, enrich the learning information of each layer, and reduce the number of parameters of the lesion segmentation model; at the same time, the depth of the lesion segmentation model is increased, which greatly improves the segmentation effect of the lesion segmentation model.

[0057] In the decoder stage, the two decoder branches are any two decoder branches in the convolutional neural network set, and the convolutional neural network set includes UNet, VNet, DeConvNet, SegNet, UNet++ and DeepLabv3++. Since two different decoder branches will output two segmentation results, if the difference between the two segmentation results is very large, it means that the lesion segmentation model does not have a high confidence in the two results, so this embodiment minimizes the difference between the two segmentation results to make the two decoder branches produce consistent output results, that is, the two decoder branches perform mutually supervised machine learning. Then, the output result weighted value of the two decoder branches is used as the final lesion segmentation result, thereby improving the confidence of the output result. Specifically, the output result weighted value of the two decoder branches is calculated using the following formula:

[0058] Output=w1×Input1+w2×Input2

[0059] Among them, Input1 and Input2 represent the output results of two decoder branches, w1 and w2 represent weight coefficients, which are preset according to specific rules. The weight coefficients set in this embodiment are all 0.5.

[0060] It can be appreciated that, compared to having two decoder branches use two independent encoder branches, sharing the encoder branch can save video memory.

[0061] S3, performing unsupervised training on the initial lesion segmentation model according to the training set to obtain an intermediate lesion segmentation model;

[0062] Specifically, Figure 4 As shown, step S3 includes:

[0063] S301, performing data enhancement on the training set to obtain an enhanced training set;

[0064] Specifically, data augmentation operations such as random flipping, rotation, and cropping can be used to synthesize more training data to obtain an enhanced training set.

[0065] S302, characterizing the Dice loss function as the loss function of the initial lesion segmentation model, so as to minimize the difference in lesion segmentation results output by the two decoder branches;

[0066] For the unlabeled data in the training set, the two decoder branches learn from each other by minimizing the difference between the lesion segmentation results output by the two decoder branches. For the unlabeled data, the lesion segmentation model can output two segmentation results. The Dice loss function can measure the similarity between the two segmentation results, and the Dice loss function can calculate the gradient of the two segmentation results, so the Dice loss function is calculated to minimize the lesion segmentation results output by the two decoder branches.

[0067] Specifically, the following formula is used to characterize the Dice loss function:

[0068] Dice_loss=1-Dice

[0069]

[0070] Wherein, X represents the set of positive pixels in the lesion segmentation result output by one decoder branch, Y represents the set of positive pixels in the lesion segmentation result output by another decoder branch, |X| and |Y| represent the sizes of pixel sets respectively, and |X∩Y| represents the size of the intersection of two pixel sets.

[0071] S303. Perform unsupervised training on the initial lesion segmentation model according to the enhanced training set until the loss function is less than a first preset threshold and reaches a preset number of training times, thereby obtaining an intermediate lesion segmentation model.

[0072] The initial lesion segmentation model is unsupervisedly trained according to the enhanced training set, and the Dice loss function is calculated until the loss function is less than the first preset threshold and the preset number of training times is reached, so as to obtain the parameters of the lesion segmentation model, that is, the intermediate lesion segmentation model.

[0073] S4, performing supervised training on the intermediate lesion segmentation model according to the test set to obtain a target lesion segmentation model;

[0074] Specifically, supervised training is performed on the intermediate lesion segmentation model according to the test set until the prediction accuracy of the lesion segmentation result is higher than a second preset threshold, thereby obtaining a target lesion segmentation model.

[0075] S5, obtaining a hysteroscopic examination image to be segmented, and performing feature extraction on the hysteroscopic examination image to be segmented to obtain a hysteroscopic examination image feature map;

[0076] Specifically, a hysteroscopic device is used to collect images of the inside of the patient's uterus to obtain a hysteroscopic examination image to be segmented; a convolutional neural network is used to extract features of the obtained hysteroscopic examination image to be segmented to obtain a hysteroscopic examination image feature map.

[0077] S6. Input the feature map into the target lesion segmentation model and output the lesion segmentation result.

[0078] It can be understood that the target lesion segmentation model obtained by steps S1 to S4 has its input as the hysteroscopic image feature map and its output as the lesion segmentation result, that is, the weighted value of the output results of the two decoder branches.

[0079] The embodiment of the present invention provides a method for segmenting lesions in hysteroscopic examination images. By adjusting the convolutional network structure of the encoder branch, the lesion segmentation model is enabled to fully learn low-level local and detail information, enrich the learning information of each layer, and reduce the number of parameters of the lesion segmentation model; at the same time, the depth of the lesion segmentation model is increased, redundant calculations are reduced, and the segmentation effect of the lesion segmentation model is greatly improved; by setting two different decoder branches to utilize more image features, and by minimizing the difference in segmentation results of the two decoder branches, the lesion segmentation model can be trained using unlabeled data, thereby achieving the effect of using a large amount of unlabeled data in the hospital for semi-supervised training; by sharing the encoder branch, computer resources can be saved.

[0080] Based on the above-mentioned hysteroscopic image lesion segmentation method, Figure 5 As shown, an embodiment of the present invention provides a hysteroscopic examination image lesion segmentation system, comprising:

[0081] Data set construction module 1, used to construct a hysteroscopy image training set and a hysteroscopy image test set;

[0082] An initial model building module 2 is used to improve the UNet network to obtain an initial lesion segmentation model, wherein the improvement includes replacing each layer of the encoder branch of the UNet network with a target layer and adding a decoder branch so that two different decoder branches share the encoder branch, wherein the target layer includes two branches, wherein one branch includes three first convolution blocks with a convolution kernel size of 3×3 connected in sequence, and the other branch includes a second convolution block with a convolution kernel size of 1×1, and the output results of the two branches pass through another second convolution block together;

[0083] An intermediate model construction module 3 is used to perform unsupervised training on the initial lesion segmentation model according to the training set to obtain an intermediate lesion segmentation model;

[0084] The target model construction module 4 is used to perform supervised training on the intermediate lesion segmentation model according to the test set to obtain the target lesion segmentation model;

[0085] The feature extraction module 5 is used to obtain the hysteroscopic examination image to be segmented, and perform feature extraction on the hysteroscopic examination image to be segmented to obtain a hysteroscopic examination image feature map;

[0086] The lesion segmentation module 6 is used to input the feature map into the target lesion segmentation model and output the lesion segmentation result, which is the weighted value of the output results of the two decoder branches.

[0087] It should be noted that each module in the above-mentioned hysteroscopic image lesion segmentation system can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules. For the specific definition of a hysteroscopic image lesion segmentation system, please refer to the definition of a hysteroscopic image lesion segmentation method above. The two have the same functions and effects, which will not be repeated here.

[0088] An embodiment of the present invention further provides a terminal device, the terminal device comprising:

[0089] processor, memory, and bus;

[0090] The bus is used to connect the processor and the memory;

[0091] The memory is used to store operation instructions;

[0092] The processor is used to call the operation instruction, and the executable instruction enables the processor to perform the operation corresponding to the hysteroscopic examination image lesion segmentation method as described above in the present invention.

[0093] In an optional embodiment, a terminal device is provided, such as Figure 6 As shown, Figure 6 The terminal device 5000 shown includes: a processor 5001 and a memory 5003. The processor 5001 and the memory 5003 are connected, such as through a bus 5002. Optionally, the terminal device 5000 may also include a transceiver 5004. It should be noted that in actual applications, the transceiver 5004 is not limited to one, and the structure of the terminal device 5000 does not constitute a limitation on the embodiments of the present invention.

[0094] Processor 5001 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA or other programmable logic device, a transistor logic device, a hardware component or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of the present invention. Processor 5001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0095] The bus 5002 may include a path to transmit information between the above components. The bus 5002 may be a PCI bus or an EISA bus, etc. The bus 5002 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0096] The memory 5003 may be a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, or an EEPROM, a CD-ROM or other optical disk storage, an optical disk storage (including a compressed optical disk, a laser disk, an optical disk, a digital versatile disk, a Blu-ray disk, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited to these.

[0097] The memory 5003 is used to store application code for executing the solution of the present invention, and the execution is controlled by the processor 5001. The processor 5001 is used to execute the application code stored in the memory 5003 to implement the content shown in any of the above method embodiments.

[0098] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the method for segmenting lesions in hysteroscopic examination images described above is implemented.

[0099] Yet another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed on a computer, enables the computer to execute the corresponding contents of the aforementioned method embodiments.

[0100] In addition, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the steps of the above method are implemented.

[0101] In summary, the embodiments of the present invention provide a method, system, device and medium for lesion segmentation in hysteroscopic examination images. By adjusting the convolutional network structure of the encoder branch, the lesion segmentation model can fully learn low-level local and detail information, enrich the learning information of each layer, and reduce the number of parameters of the lesion segmentation model; at the same time, the depth of the lesion segmentation model is increased, redundant calculations are reduced, and the segmentation effect of the lesion segmentation model is greatly improved; by setting two different decoder branches to utilize more image features, and by minimizing the difference in segmentation results of the two decoder branches, the lesion segmentation model can be trained using unlabeled data, thereby achieving the effect of semi-supervised training using a large amount of unlabeled data in the hospital; by sharing the encoder branch, computer resources can be saved.

[0102] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.

Claims

1. A method for segmenting lesions in hysteroscopic images, characterized in that: include: Construct a hysteroscopic image training set and a hysteroscopic image test set; The UNet network is improved to obtain an initial lesion segmentation model, wherein the improvement includes replacing each layer of the encoder branch of the UNet network with a target layer and adding a decoder branch so that two different decoder branches share the encoder branch, wherein the target layer includes two branches, wherein one branch includes three first convolution blocks with a convolution kernel size of 3×3 connected in sequence, and the other branch includes a second convolution block with a convolution kernel size of 1×1, and the output results of the two branches pass through another second convolution block together; Performing unsupervised training on the initial lesion segmentation model according to the training set to obtain an intermediate lesion segmentation model; Performing supervised training on the intermediate lesion segmentation model according to the test set to obtain a target lesion segmentation model; Acquire a hysteroscopic examination image to be segmented, and perform feature extraction on the hysteroscopic examination image to be segmented to obtain a hysteroscopic examination image feature map; The feature map is input into the target lesion segmentation model, and a lesion segmentation result is output, where the lesion segmentation result is a weighted value of the output results of the two decoder branches.

2. The method for segmenting lesions in hysteroscopic images according to claim 1, characterized in that: The step of constructing a hysteroscopic examination image training set and a hysteroscopic examination image test set comprises: Multiple hysteroscopic images were obtained from the hospital database; screening unlabeled images in the hysteroscopic images to form a hysteroscopic image training set; The labeled images in the hysteroscopic examination images are screened to form a hysteroscopic examination image test set.

3. The method for segmenting lesions in hysteroscopic images according to claim 1, characterized in that: The first convolution module includes a convolution layer with a convolution kernel size of 3×3 and a normalization layer.

4. The method for segmenting lesions in hysteroscopic images according to claim 1, characterized in that: The second convolution module includes a convolution layer with a convolution kernel size of 1×1, a normalization layer and an activation layer.

5. The method for segmenting lesions in hysteroscopic images according to claim 1, characterized in that: The two decoder branches are any two decoder branches in a convolutional neural network set, and the convolutional neural network set includes UNet, VNet, DeConvNet, SegNet, UNet++ and DeepLabv3++.

6. The method for segmenting lesions in hysteroscopic images according to claim 1, characterized in that: The step of performing unsupervised training on the initial lesion segmentation model according to the training set to obtain an intermediate lesion segmentation model comprises: Performing data enhancement on the training set to obtain an enhanced training set; Characterizing the Dice loss function as the loss function of the initial lesion segmentation model to minimize the difference in lesion segmentation results output by the two decoder branches; The initial lesion segmentation model is unsupervisedly trained according to the enhanced training set until the loss function is less than a first preset threshold and a preset number of training times is reached, thereby obtaining an intermediate lesion segmentation model.

7. The method for segmenting lesions in hysteroscopic images according to claim 1, characterized in that: The step of performing supervised training on the intermediate lesion segmentation model according to the test set to obtain a target lesion segmentation model comprises: The intermediate lesion segmentation model is supervised and trained according to the test set until the prediction accuracy of the lesion segmentation result is higher than a second preset threshold, thereby obtaining a target lesion segmentation model.

8. A hysteroscopic image lesion segmentation system, characterized in that: include: A data set construction module, used to construct a hysteroscopy image training set and a hysteroscopy image test set; An initial model building module, used for improving the UNet network to obtain an initial lesion segmentation model, wherein the improvement includes replacing each layer of the encoder branch of the UNet network with a target layer and adding a decoder branch so that two different decoder branches share the encoder branch, wherein the target layer includes two branches, wherein one branch includes three first convolution blocks with a convolution kernel size of 3×3 connected in sequence, and the other branch includes a second convolution block with a convolution kernel size of 1×1, and the output results of the two branches pass through another second convolution block together; An intermediate model construction module, used for performing unsupervised training on the initial lesion segmentation model according to the training set to obtain an intermediate lesion segmentation model; A target model construction module, used for performing supervised training on the intermediate lesion segmentation model according to the test set to obtain a target lesion segmentation model; A feature extraction module is used to obtain a hysteroscopic examination image to be segmented, and perform feature extraction on the hysteroscopic examination image to be segmented to obtain a hysteroscopic examination image feature map; The lesion segmentation module is used to input the feature map into the target lesion segmentation model and output a lesion segmentation result, wherein the lesion segmentation result is a weighted value of the output results of the two decoder branches.

9. A terminal device, characterized in that: The invention comprises a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for segmenting lesions in hysteroscopic images as claimed in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the hysteroscopic image lesion segmentation method according to any one of claims 1 to 7.

Images