A method, device and system for segmenting thyroid nodule ultrasound images

Abstract

The present invention discloses a segmentation method, device and system for ultrasound images of thyroid nodules. The segmentation device includes a data acquisition unit, a model training unit and a segmentation extraction unit. The segmentation system includes an image segmentation module and a data storage module. By training and evaluating a preset first visual segmentation model obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of a preset semantic segmentation network Unet to obtain a second visual segmentation model, and then segmenting the ultrasound image to be segmented by the second visual segmentation model, thereby obtaining a thyroid nodule ultrasound image segmentation result map, the segmentation method, device and system improve the accuracy of the segmentation of thyroid nodule images.

Description

Technical Field

[0001] The present invention relates to the field of segmentation of thyroid nodule ultrasonic images, and to a method, device and system for segmenting thyroid nodule ultrasonic images. Background Art

[0002] Thyroid nodules are important endocrine organs in the human body. They are located in the front of the neck below the thyroid cartilage and secrete thyroid hormones to regulate the body's metabolism. According to epidemiological studies, perceptible thyroid nodules occur in 4%-7% of the population, among which low-echo nodules with irregular boundaries are more likely to develop into malignant nodules. The incidence of malignant thyroid nodules is 0.1%-0.2%. The shape and boundaries of thyroid nodules are the key features to distinguish between benign and malignant nodules. Image segmentation is to classify each pixel in the image. Accurate image segmentation has always been a difficult problem, especially in the field of medical image segmentation. Accurate thyroid nodule segmentation is the basis for distinguishing between benign and malignant nodules. Therefore, achieving more accurate thyroid nodule ultrasound image segmentation is crucial for the diagnosis of thyroid diseases.

[0003] In the existing technology, the segmentation methods of thyroid nodules in ultrasound images are mainly divided into contour and shape-based methods, region-based methods, deep learning-based methods, and methods that use multiple technologies together. Among them, in the deep learning-based methods, most methods use technologies based on fully convolutional neural networks (such as Unet networks) to achieve automatic segmentation of nodules in images.

[0004] However, the prior art still has the following defects: the network model in the prior art often only considers the local information of the thyroid nodule ultrasound image, resulting in low accuracy in the segmentation of the thyroid nodule image.

[0005] Therefore, there is a need for a segmentation method, device and system for thyroid nodule ultrasound images to overcome the above-mentioned defects in the prior art. Summary of the invention

[0006] In view of the above-mentioned existing technical problems, the object of the present invention is to provide a method, device and system for segmenting thyroid nodule ultrasound images, so as to improve the accuracy of thyroid nodule image segmentation.

[0007] The present invention provides a method for segmenting an ultrasound image of a thyroid nodule, the method comprising: obtaining a preset model training data group and an ultrasound image data group to be segmented; training and evaluating a preset first visual segmentation model according to the model training data group, thereby obtaining a second visual segmentation model; the first visual segmentation model is obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of a preset semantic segmentation network Unet; and segmenting and extracting the ultrasound image data group according to the second visual segmentation model, thereby obtaining a thyroid nodule ultrasound image data group.

[0008] In one embodiment, after obtaining a preset model training data group and an ultrasound image data group to be segmented, the segmentation method also includes: replacing the downsampling method of the preset semantic segmentation network Unet with a two-branch downsampling method, thereby obtaining a first improved model; introducing a preset hierarchical cascade multi-head self-attention module in the third downsampling stage, the fourth downsampling stage, the bottom stage and the corresponding upsampling stage of the first improved model, thereby obtaining a first visual segmentation model.

[0009] In one embodiment, according to the model training data group, a preset first visual segmentation model is trained and evaluated to obtain a second visual segmentation model, which specifically includes: according to a preset data cleaning method, the data in the model training data group is cleaned to obtain a first training data group; according to a preset cropping and equalization method, the data in the first training data group is cropped and equalized to obtain a second training data group; the second training data group is divided into a training set, a validation set and a test set according to a preset ratio; the training set is input into the preset first visual segmentation model for model training to obtain a first training model; according to the validation set, the parameters of the first training model are adjusted to obtain a second training model; the second training model is used as a second visual segmentation model, and the second visual segmentation model is evaluated according to the test set and the preset model evaluation method to obtain a corresponding evaluation result.

[0010] In one embodiment, after segmenting and extracting the ultrasound image data group according to the second visual segmentation model to obtain the thyroid nodule ultrasound image data group, it also includes: generating a segmentation analysis report based on the thyroid nodule ultrasound image data group, and sending the segmentation analysis report to a user.

[0011] The present invention also provides a segmentation device for thyroid nodule ultrasound images, the segmentation device comprising a data acquisition unit, a model training unit and a segmentation extraction unit, the data acquisition unit is used to acquire a preset model training data group and an ultrasound image data group to be segmented; the model training unit is used to train and evaluate a preset first visual segmentation model according to the model training data group, thereby obtaining a second visual segmentation model; the first visual segmentation model is obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of a preset semantic segmentation network Unet; the segmentation extraction unit is used to segment and extract the ultrasound image data group according to the second visual segmentation model, thereby obtaining a thyroid nodule ultrasound image data group.

[0012] In one embodiment, the segmentation device also includes a model acquisition unit, which is used to: replace the downsampling method of the preset semantic segmentation network Unet with a two-branch downsampling method to obtain a first improved model; introduce a preset hierarchical cascade multi-head self-attention module in the third downsampling stage, the fourth downsampling stage, the bottom stage and the corresponding upsampling stage of the first improved model to obtain a first visual segmentation model.

[0013] In one embodiment, the segmentation device further includes an analysis report unit, which is used to generate a segmentation analysis report based on the thyroid nodule ultrasound image data set and send the segmentation analysis report to a user.

[0014] In one embodiment, the model training unit is specifically used to: perform data cleaning on the data in the model training data group according to a preset data cleaning method, so as to obtain a first training data group; perform cropping and equalization on the data in the first training data group according to a preset cropping and equalization method, so as to obtain a second training data group; divide the second training data group into a training set, a validation set and a test set according to a preset ratio; input the training set into a preset first visual segmentation model for model training, so as to obtain a first training model; adjust the parameters of the first training model according to the validation set, so as to obtain a second training model; use the second training model as a second visual segmentation model, and evaluate the second visual segmentation model according to the test set and a preset model evaluation method, so as to obtain a corresponding evaluation result.

[0015] The present invention also provides a thyroid nodule ultrasound image segmentation system, which includes an image segmentation module and a data storage module, wherein the image segmentation module is communicatively connected to the data storage module, wherein the image segmentation module is used to execute the thyroid nodule ultrasound image segmentation method as described above, and the data storage module is used to store all data.

[0016] In one embodiment, the segmentation system further includes a user interaction module, which is communicatively connected to the image segmentation module and the data storage module respectively, and the user interaction module includes a touchable / non-touchable display screen, an input keyboard, a virtual keyboard, a microphone or a speaker.

[0017] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0018] The present invention provides a method, device and system for segmenting thyroid nodule ultrasound images. A preset first visual segmentation model obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of a preset semantic segmentation network Unet and modifying the original pooling layer downsampling to a dual-branch downsampling method is trained and evaluated to obtain a second visual segmentation model, and then the ultrasound image to be segmented is segmented by the second visual segmentation model, thereby obtaining a thyroid nodule ultrasound image. The segmentation method, device and system improve the accuracy of thyroid nodule image segmentation.

[0019] Furthermore, the thyroid nodule ultrasound image segmentation method, device and system provided by the present invention also improve the preset semantic segmentation network Unet by replacing the downsampling method of the preset semantic segmentation network Unet with a two-branch downsampling method when improving the preset semantic segmentation network Unet to obtain the first visual segmentation model, thereby further improving the segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be further described below in conjunction with the accompanying drawings, wherein:

[0021] Figure 1 A flow chart showing an embodiment of a method for segmenting a thyroid nodule ultrasound image according to the present invention is shown;

[0022] Figure 2 A flowchart showing another embodiment of a method for segmenting a thyroid nodule ultrasound image according to the present invention is shown;

[0023] Figure 3 A structural diagram showing an embodiment of a thyroid nodule ultrasound image segmentation device according to the present invention;

[0024] Figure 4 A structural diagram of an embodiment of a thyroid nodule ultrasound image segmentation system according to the present invention is shown. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific embodiment 1

[0027] The embodiment of the present invention firstly describes a segmentation method of a thyroid nodule ultrasound image. Figure 1 A flow chart of an embodiment of a method for segmenting a thyroid nodule ultrasound image according to the present invention is shown.

[0028] like Figure 1 As shown, the method comprises the following steps:

[0029] S1: Obtain a preset model training data set and an ultrasound image data set to be segmented.

[0030] In order to improve the segmentation accuracy of thyroid nodule areas in ultrasound images, an embodiment of the present invention proposes to improve the existing semantic segmentation network Unet, and introduces a preset hierarchical cascade multi-head self-attention module (H-MHSA module) on its basis, and modifies the original pooling layer downsampling to a dual-branch downsampling method, and trains the improved model. Subsequently, the trained model is used to identify and segment the ultrasound image to be segmented (for ease of description and understanding, the embodiment of the present invention describes the ultrasound image as an "ultrasound image data group").

[0031] Therefore, in order to successfully realize the identification and segmentation of thyroid nodules in the future, the embodiment of the present invention must first obtain a preset model training data group to train the improved model, and needs to obtain an ultrasound image data group to be segmented as the object of identification and segmentation.

[0032] The preset model training data set is obtained by obtaining thyroid nodule ultrasound images, nodule region ultrasound images and their corresponding segmentation annotation files from the database. Specifically, the thyroid nodule ultrasound images are first obtained from the database, and professionals are asked to manually annotate the thyroid nodule region in the ultrasound image. Subsequently, the annotated json file is converted into a black and white Mask label to obtain the corresponding segmentation standard file.

[0033] S2: According to the model training data group, the preset first visual segmentation model is trained and evaluated to obtain a second visual segmentation model.

[0034] Among them, the first visual segmentation model is obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of the preset semantic segmentation network Unet. Specifically, referring to the encoder-decoder structure of the Unet network, a H-MHSA module (hierarchical cascade multi-head self-attention module) is introduced at the beginning of the third downsampling of the Unet model framework (including the bottom stage) and its corresponding upsampling stage, so as to obtain the first visual segmentation model; the first visual segmentation model is used to capture long-distance information, thereby solving the shortcomings of the existing preset Unet network model in capturing long-distance information, thereby improving the accuracy of segmentation. At the same time, the first visual segmentation model in the embodiment of the present invention has the advantage of end-to-end, can automatically segment thyroid nodules, can be used in thyroid nodule ultrasound image-assisted diagnosis system, assist doctors in diagnosis, has clinical use value, and can be used as the basis for subsequent classification of benign and malignant thyroid nodules.

[0035] In one embodiment, this step specifically includes: according to a preset data cleaning method, cleaning the data in the model training data group to obtain a first training data group; according to a preset cropping and equalization method, cropping and equalizing the data in the first training data group to obtain a second training data group; dividing the second training data group into a training set, a validation set and a test set in a preset ratio; inputting the training set into a preset first visual segmentation model for model training to obtain a first training model; adjusting the parameters of the first training model according to the validation set to obtain a second training model; using the second training model as a second visual segmentation model, and evaluating the second visual segmentation model according to the test set and a preset model evaluation method to obtain a corresponding evaluation result.

[0036] In one embodiment, parameter adjustment includes the following steps: inputting the data in the validation set into a first training model and calculating the first accuracy of the first training model; saving the first training model and the corresponding first accuracy, adjusting the parameters of the first training model according to the first accuracy and a preset parameter adjustment method, and repeating the above steps; using the first training model with the highest first accuracy as the second training model.

[0037] In one embodiment, the model performance of the second training model and the preset semantic segmentation network Unet are evaluated on the test set respectively, so as to obtain the second accuracy and the third accuracy respectively, and the evaluation result is obtained according to the quantitative relationship between the second accuracy and the third accuracy.

[0038] Among them, the data cleaning step can clean up data with obvious labeling errors, missing labels, and particularly poor image quality; the trimming step can achieve data desensitization; the equalization step refers to the method of limiting contrast histogram equalization to improve the quality of ultrasound images and enhance the contrast, so as to facilitate distinction; during the training process, the pixel values ​​of (0-255) need to be normalized to between (0-1) before training. In one embodiment, the preset ratio is 6:2:2 for training set: validation set: test set.

[0039] S3: According to the second visual segmentation model, the ultrasound image data group is segmented and extracted to obtain a thyroid nodule ultrasound image data group.

[0040] After the second visual segmentation model is obtained through training and evaluation, the second visual segmentation model can be used to perform segmentation and extraction on the ultrasound image data set to be segmented.

[0041] An embodiment of the present invention describes a method for segmenting an ultrasound image of a thyroid nodule. The method improves the accuracy of segmentation of thyroid nodule images by training and evaluating a preset first visual segmentation model obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of a preset semantic segmentation network Unet and modifying the original pooling layer downsampling to a dual-branch downsampling method. The second visual segmentation model is then used to segment the ultrasound image to be segmented. Specific embodiment 2

[0043] Furthermore, an embodiment of the present invention also describes a method for segmenting a thyroid nodule ultrasound image. Figure 2 A flow chart of another embodiment of a method for segmenting a thyroid nodule ultrasound image according to the present invention is shown.

[0044] like Figure 2 As shown, the method comprises the following steps:

[0045] A1: Replace the downsampling method of the preset semantic segmentation network Unet with a two-branch downsampling method to obtain the first improved model.

[0046] Among them, the two-branch downsampling method refers to one branch downsampling through a typical 3x3 convolution with a step size of 2, and the other branch downsampling through maximum pooling combined with 1x1 convolution for channel transformation. The information of the two branches is fused by adding element-by-element summation, which can retain more context information. In one embodiment, the RelU activation function of the preset semantic segmentation network Unet is also replaced with a smoother SiLU activation function.

[0047] In order to successfully outline the edge of the thyroid nodule area later, in one embodiment, the number of channels is set according to 64, 128, 256, 512, and 1024, and channel splicing is performed through skip connection in the encoding stage and the decoding stage to make full use of low-level information and high-level information. Finally, in the decoding stage to restore the original image size, a 1x1 convolution map is used to map the output of 1 channel to obtain a mask image as the edge of the thyroid nodule area.

[0048] A2: A preset hierarchical cascade multi-head self-attention module is introduced into the third downsampling stage, the fourth downsampling stage, the bottom stage and the corresponding upsampling stage of the first improved model to obtain the first visual segmentation model.

[0049] In order to solve the insufficiency of the existing preset Unet network model in capturing long-distance information, thereby improving the accuracy of segmentation, the embodiment of the present invention also introduces a preset hierarchical cascade multi-head self-attention module in the third downsampling stage, the fourth downsampling stage, the bottom stage and the corresponding upsampling stage of the first improved model. Among them, the third downsampling stage corresponds to the downsampling of the third layer, the fourth downsampling stage corresponds to the downsampling of the fourth layer; the corresponding upsampling stage refers to the upsampling of the bottom and the upsampling of the corresponding fourth layer during the decoding process.

[0050] In one embodiment, in the encoding stage for downsampling feature extraction and the decoding stage for restoring the image size, the low-level information and high-level information of the corresponding stages are fused with features in the form of skip connections.

[0051] In one embodiment, this step is specifically as follows in the application: in the downsampling stage of the third and fourth layers and the bottom middle stage, in the layer where upsampling needs to start and in the decoding layer corresponding to the jump connection feature fusion in the fourth encoding stage, two H-MHSA modules and a depth-separable convolution of an inverted residual structure are stacked to capture long-distance information and feature learning; in the aforementioned downsampling stage, two-branch downsampling is used after stacking two H-MHSAs and a depth-separable convolution of an inverted residual structure, and deconvolution is used in the corresponding upsampling stage; two 3x3 convolutions are stacked in the first two layers of the encoding stage , BN normalization layer and SiLU activation function, and down-sampling is performed in the form of two branches (using Maxpool and convolution with a step size of 2), and the feature maps obtained from the two branches are summed and fused; while the upsampling (decoding) stage of the first and second layers still uses two stacked 3x3 convolutions, BN normalization layers and SiLU activation functions, in the form of deconvolution, and in the final output layer, two stacked 3x3 convolutions, BN normalization and SiLU activation functions are used, and then a 1x1 convolution is used to map the number of channels of the feature map to 1, and the Sigmoid function is used to activate the output feature map.

[0052] A3: Obtain a preset model training data set and an ultrasound image data set to be segmented.

[0053] In order to improve the segmentation accuracy of thyroid nodule areas in ultrasound images, an embodiment of the present invention proposes to improve the existing semantic segmentation network Unet, introduce a preset hierarchical cascade multi-head self-attention module (H-MHSA module) on its basis, and train the improved model. Subsequently, the trained model is used to identify and segment the ultrasound image to be segmented (for ease of description and understanding, the embodiment of the present invention describes the ultrasound image as an "ultrasound image data group").

[0054] Therefore, in order to successfully realize the identification and segmentation of thyroid nodules in the future, the embodiment of the present invention must first obtain a preset model training data group to train the improved model, and needs to obtain an ultrasound image data group to be segmented as the object of identification and segmentation.

[0055] The preset model training data set is obtained by obtaining thyroid nodule ultrasound images, nodule region ultrasound images and their corresponding segmentation annotation files from the database. Specifically, the thyroid nodule ultrasound images are first obtained from the database, and professionals are asked to manually annotate the thyroid nodule region in the ultrasound image. Subsequently, the annotated json file is converted into a black and white Mask label to obtain the corresponding segmentation standard file.

[0056] A4: Based on the model training data group, the preset first visual segmentation model is trained and evaluated to obtain the second visual segmentation model.

[0057] Among them, the first visual segmentation model is obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of the preset semantic segmentation network Unet. Specifically, referring to the encoder-decoder structure of the Unet network, a H-MHSA module (hierarchical cascade multi-head self-attention module) is introduced in the bottom stage and its corresponding upsampling (decoding) stage starting from the third downsampling (encoding) stage of the Unet model framework, so as to obtain the first visual segmentation model; the first visual segmentation model is used to capture long-distance information, so as to solve the problem that the existing preset Unet network model is insufficient in capturing long-distance information, thereby improving the accuracy of segmentation. At the same time, the first visual segmentation model in the embodiment of the present invention has the advantage of end-to-end, can automatically segment thyroid nodules, can be used in thyroid nodule ultrasound image-assisted diagnosis system, assist doctors in diagnosis, has clinical use value, and can be used as the basis for subsequent classification of benign and malignant thyroid nodules.

[0058] In one embodiment, this step specifically includes: according to a preset data cleaning method, cleaning the data in the model training data group to obtain a first training data group; according to a preset cropping and equalization method, cropping and equalizing the data in the first training data group to obtain a second training data group; dividing the second training data group into a training set, a validation set and a test set in a preset ratio; inputting the training set into a preset first visual segmentation model for model training to obtain a first training model; adjusting the parameters of the first training model according to the validation set to obtain a second training model; using the second training model as a second visual segmentation model, and evaluating the second visual segmentation model according to the test set and a preset model evaluation method to obtain a corresponding evaluation result.

[0059] In one embodiment, parameter adjustment includes the following steps: inputting the data in the validation set into a first training model and calculating the first accuracy of the first training model; saving the first training model and the corresponding first accuracy, adjusting the parameters of the first training model according to the first accuracy and a preset parameter adjustment method, and repeating the above steps; using the first training model with the highest first accuracy as the second training model.

[0060] In one embodiment, the model performance of the second training model and the preset semantic segmentation network Unet are evaluated on the test set respectively, so as to obtain the second accuracy and the third accuracy respectively, and the evaluation result is obtained according to the quantitative relationship between the second accuracy and the third accuracy.

[0061] Among them, the data cleaning step can clean up data with obvious labeling errors, missing labels, and particularly poor image quality; the trimming step can achieve data desensitization; the equalization step refers to the method of limiting contrast histogram equalization to improve the quality of ultrasound images and enhance the contrast, so as to facilitate distinction; during the training process, the pixel values ​​of (0-255) need to be normalized to between (0-1) before training. In one embodiment, the preset ratio is 6:2:2 for training set: validation set: test set.

[0062] A5: According to the second visual segmentation model, the ultrasound image data group is segmented and extracted to obtain a thyroid nodule ultrasound image data group.

[0063] After the second visual segmentation model is obtained through training and evaluation, the second visual segmentation model can be used to perform segmentation and extraction on the ultrasound image data set to be segmented.

[0064] A6: Generate a segmentation analysis report based on the thyroid nodule ultrasound image data set, and send the segmentation analysis report to the user.

[0065] After obtaining the thyroid nodule ultrasound image data set, in order to improve the visibility of the segmentation result, a segmentation analysis report is generated according to the thyroid nodule ultrasound image data set, and the segmentation analysis report is sent to the user.

[0066] In one embodiment, the segmentation analysis report includes the position and edge line of the thyroid nodule region on the ultrasound image. In one embodiment, the content included in the segmentation analysis report is selected and set by the user.

[0067] An embodiment of the present invention describes a method for segmenting thyroid nodule ultrasound images. The preset first visual segmentation model obtained by introducing a preset hierarchical cascade multi-head self-attention module and modifying the original pooling layer downsampling to a two-branch downsampling method on the basis of a preset semantic segmentation network Unet is trained and evaluated to obtain a second visual segmentation model, and then the ultrasound image to be segmented is segmented by the second visual segmentation model to obtain a thyroid nodule ultrasound image. The segmentation method improves the accuracy of segmentation of the thyroid nodule image; further, the method for segmenting thyroid nodule ultrasound images described in an embodiment of the present invention further improves the segmentation accuracy by replacing the downsampling method of the preset semantic segmentation network Unet with a two-branch downsampling method when improving the preset semantic segmentation network Unet to obtain the first visual segmentation model. Specific embodiment three

[0069] In addition to the above method, the embodiment of the present invention also describes a device for segmenting thyroid nodule ultrasound images. Figure 3A structural diagram of an embodiment of a thyroid nodule ultrasound image segmentation device according to the present invention is shown.

[0070] As shown in the figure, the segmentation device includes a data acquisition unit 11, a model training unit 12 and a segmentation extraction unit 13.

[0071] The data acquisition unit 11 is used to acquire a preset model training data set and an ultrasound image data set to be segmented.

[0072] The model training unit 12 is used to train and evaluate the preset first visual segmentation model according to the model training data group, so as to obtain the second visual segmentation model. The first visual segmentation model is obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of the preset semantic segmentation network Unet and modifying the original pooling layer downsampling to a dual-branch downsampling method.

[0073] The segmentation and extraction unit 13 is used to segment and extract the ultrasound image data set according to the second visual segmentation model, so as to obtain a thyroid nodule ultrasound image data set.

[0074] When it is necessary to segment the thyroid nodules in the ultrasound image data group, the data acquisition unit 11 is first used to acquire a preset model training data group and an ultrasound image data group to be segmented; then, the model training unit 12 is used to train and evaluate the preset first visual segmentation model based on the model training data group, so as to obtain a second visual segmentation model; finally, the segmentation extraction unit 13 is used to segment and extract the ultrasound image data group based on the second visual segmentation model, so as to obtain a thyroid nodule ultrasound image data group.

[0075] In one embodiment, the segmentation device also includes a model acquisition unit, which is used to: replace the downsampling method of the preset semantic segmentation network Unet with a two-branch downsampling method to obtain a first improved model; introduce a preset hierarchical cascade multi-head self-attention module in the third downsampling stage, the fourth downsampling stage and the corresponding upsampling stage of the first improved model to obtain a first visual segmentation model.

[0076] In one embodiment, the segmentation device further includes an analysis report unit, which is used to generate a segmentation analysis report based on the thyroid nodule ultrasound image data set and send the segmentation analysis report to a user.

[0077] In one embodiment, the model training unit 12 is specifically used to: perform data cleaning on the data in the model training data group according to a preset data cleaning method, so as to obtain a first training data group; perform cropping and equalization on the data in the first training data group according to a preset cropping and equalization method, so as to obtain a second training data group; divide the second training data group into a training set, a validation set and a test set according to a preset ratio; input the training set into a preset first visual segmentation model for model training, so as to obtain a first training model; adjust the parameters of the first training model according to the validation set, so as to obtain a second training model; use the second training model as a second visual segmentation model, and evaluate the second visual segmentation model according to the test set and a preset model evaluation method, so as to obtain a corresponding evaluation result.

[0078] The embodiment of the present invention describes a device for segmenting thyroid nodule ultrasound images. The second visual segmentation model is obtained by training and evaluating a preset first visual segmentation model obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of a preset semantic segmentation network Unet and modifying the original pooling layer downsampling to a two-branch downsampling method. The ultrasound image to be segmented is then segmented by the second visual segmentation model to obtain a thyroid nodule ultrasound image. The segmentation method, device and system improve the accuracy of segmentation of thyroid nodule images. Furthermore, the method, device and system for segmenting thyroid nodule ultrasound images described in the embodiment of the present invention further improves the segmentation accuracy by replacing the downsampling method of the preset semantic segmentation network Unet with a two-branch downsampling method when improving the preset semantic segmentation network Unet to obtain the first visual segmentation model. Specific embodiment 4

[0080] In addition to the above method and apparatus, the present invention also describes a segmentation system for thyroid nodule ultrasound images. Figure 4 A structural diagram of an embodiment of a thyroid nodule ultrasound image segmentation system according to the present invention is shown.

[0081] like Figure 4 As shown, the segmentation system includes an image segmentation module 1 and a data storage module 2, and the image segmentation module 1 is communicatively connected with the data storage module 2.

[0082] The image segmentation module 1 is used to execute the above-mentioned segmentation method of thyroid nodule ultrasound image, so as to segment a thyroid nodule ultrasound image data group from the ultrasound image data group to be segmented, and the data storage module 2 is used to store all data.

[0083] In one embodiment, the segmentation system further comprises a user interaction module, which is respectively connected to the image segmentation module and the data storage module for communicating with the user. In one embodiment, the user interaction module comprises a touchable / non-touchable display screen, an input keyboard, a virtual keyboard, a microphone or a speaker.

[0084] The embodiment of the present invention describes a method, device and system for segmenting thyroid nodule ultrasound images. The preset first visual segmentation model obtained by introducing a preset hierarchical cascade multi-head self-attention and modifying the original pooling layer downsampling to a two-branch downsampling mode module on the basis of a preset semantic segmentation network Unet is trained and evaluated to obtain a second visual segmentation model, and then the ultrasound image to be segmented is segmented by the second visual segmentation model to obtain a thyroid nodule ultrasound image. The segmentation method, device and system improve the accuracy of segmentation of thyroid nodule images; further, the method, device and system for segmenting thyroid nodule ultrasound images described in the embodiment of the present invention also improves the preset semantic segmentation network Unet to obtain the first visual segmentation model by replacing the downsampling mode of the preset semantic segmentation network Unet with a two-branch downsampling mode, thereby further improving the segmentation accuracy.

[0085] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for segmenting a thyroid nodule ultrasound image, characterized in that: The segmentation method comprises: Acquire a preset model training data set and an ultrasound image data set to be segmented; According to the model training data group, a preset first visual segmentation model is trained and evaluated to obtain a second visual segmentation model; the first visual segmentation model is obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of a preset semantic segmentation network Unet; Segmenting and extracting the ultrasound image data set according to the second visual segmentation model, thereby obtaining a thyroid nodule ultrasound image data set; Before obtaining the preset model training data set and the ultrasound image data set to be segmented, the segmentation method further includes: The downsampling method of the preset semantic segmentation network Unet is replaced by a two-branch downsampling method, thereby obtaining a first improved model; A preset hierarchical cascade multi-head self-attention module is introduced into the third downsampling stage, the fourth downsampling stage, the bottom stage and the corresponding upsampling stage of the first improved model to obtain a first visual segmentation model.

2. The method for segmenting thyroid nodules ultrasonic images according to claim 1, characterized in that: According to the model training data group, a preset first visual segmentation model is trained and evaluated to obtain a second visual segmentation model, specifically including: According to a preset data cleaning method, data in the model training data group is cleaned to obtain a first training data group; According to a preset trimming and balancing method, trimming and balancing the data in the first training data group, so as to obtain a second training data group; Dividing the second training data set into a training set, a validation set, and a test set in a preset ratio; Inputting the training set into a preset first visual segmentation model to perform model training, thereby obtaining a first training model; According to the verification set, adjusting parameters of the first training model to obtain a second training model; The second training model is used as a second visual segmentation model, and the second visual segmentation model is evaluated according to the test set and a preset model evaluation method to obtain a corresponding evaluation result.

3. The method for segmenting thyroid nodules ultrasonic images according to claim 2, characterized in that: After segmenting and extracting the ultrasound image data set according to the second visual segmentation model to obtain a thyroid nodule ultrasound image data set, the method further includes: A segmentation analysis report is generated according to the thyroid nodule ultrasound image data set, and the segmentation analysis report is sent to a user.

4. A thyroid nodule ultrasound image segmentation device, characterized in that: The segmentation device includes a data acquisition unit, a model training unit and a segmentation extraction unit. The data acquisition unit is used to acquire a preset model training data group and an ultrasound image data group to be segmented; The model training unit is used to train and evaluate the preset first visual segmentation model according to the model training data group, so as to obtain a second visual segmentation model; the first visual segmentation model is obtained by introducing a preset hierarchical cascade multi-head self-attention module on the basis of a preset semantic segmentation network Unet; The segmentation and extraction unit is used to segment and extract the ultrasound image data set according to the second visual segmentation model, so as to obtain a thyroid nodule ultrasound image data set; The segmentation device further includes a model acquisition unit, and the model acquisition unit is used to: The downsampling method of the preset semantic segmentation network Unet is replaced by a two-branch downsampling method, thereby obtaining a first improved model; A preset hierarchical cascade multi-head self-attention module is introduced into the third downsampling stage, the fourth downsampling stage, the bottom stage and the corresponding upsampling stage of the first improved model to obtain a first visual segmentation model.

5. The device for segmenting thyroid nodules ultrasonic images according to claim 4, characterized in that: The segmentation device further comprises an analysis report unit, which is used to generate a segmentation analysis report according to the thyroid nodule ultrasound image data set and send the segmentation analysis report to a user.

6. The device for segmenting thyroid nodules ultrasonic images according to claim 5, characterized in that: The model training unit is specifically used for: According to a preset data cleaning method, data in the model training data group is cleaned to obtain a first training data group; According to a preset trimming and balancing method, trimming and balancing the data in the first training data group, so as to obtain a second training data group; Dividing the second training data set into a training set, a validation set, and a test set in a preset ratio; Inputting the training set into a preset first visual segmentation model to perform model training, thereby obtaining a first training model; According to the verification set, adjusting parameters of the first training model to obtain a second training model; The second training model is used as a second visual segmentation model, and the second visual segmentation model is evaluated according to the test set and a preset model evaluation method to obtain a corresponding evaluation result.

7. A thyroid nodule ultrasound image segmentation system, characterized in that: The segmentation system includes an image segmentation module and a data storage module, wherein the image segmentation module is communicatively connected to the data storage module, wherein the image segmentation module is used to execute the thyroid nodule ultrasound image segmentation method as described in any one of claims 1 to 3, and the data storage module is used to store all data.

8. The thyroid nodule ultrasound image segmentation system according to claim 7, characterized in that: The segmentation system further includes a user interaction module, which is respectively connected to the image segmentation module and the data storage module in communication, and includes a touchable / non-touchable display screen, an input keyboard, a virtual keyboard, a microphone or a speaker.

Images