Thyroid automatic segmentation method and system based on uncertain region perception

Abstract

The application belongs to the technical field of thyroid image segmentation, and specifically discloses a thyroid automatic segmentation method and system based on uncertain region perception. The method is used for pre-processing an original ultrasound image, constructing a double-resolution network model by using a deep learning library, training the double-resolution network model, obtaining an optimized double-resolution network model, inputting the pre-processed ultrasound image into the optimized double-resolution network model, and obtaining segmented feature information. A convolutional pyramid uncertain region identification module is constructed, the obtained feature information is input into the convolutional pyramid uncertain region identification module, a missed region in the feature information is repaired, and a multi-region is suppressed. The features output by the convolutional pyramid uncertain region identification module are input into a residual module, and final thyroid segmentation features are obtained. By using the technical scheme, the missed region is repaired, the multi-region is suppressed, and the segmentation accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of thyroid image segmentation, and relates to an automatic thyroid segmentation method and system based on uncertain region perception. BACKGROUND

[0002] As the largest endocrine gland in the human body, the thyroid mainly functions to secrete thyroid hormones. Thyroid hormones play an important role in promoting growth and development, metabolic regulation, and maintaining the excitability of the nervous system. In clinical practice, accurate segmentation of the thyroid plays an important role in the diagnosis of thyroid-related diseases and the planning of subsequent surgeries. The current clinical method for thyroid segmentation is based on manual outlining. However, this method is time-consuming and labor-intensive, and the quality of manual outlining of the thyroid is easily affected by human factors such as experience, mood, and fatigue level.

[0003] At present, image segmentation methods based on deep learning have replaced traditional methods and become the mainstream of medical image segmentation due to their strong feature expression ability, and have made great progress in thyroid segmentation. However, there are still some problems with the segmentation of the thyroid using this method: 1) a serious multi-part problem, which greatly increases the screening work of medical personnel; 2) a relatively serious missing part phenomenon, which greatly hinders the clinical application of the method; and 3) an uneven edge segmentation, which poses a great challenge to subsequent measurement and surgery planning. SUMMARY

[0004] The purpose of the present application is to provide an automatic thyroid segmentation method and system based on uncertain region perception to improve the accuracy of segmentation.

[0005] To achieve the above purpose, the basic scheme of the present application is as follows: an automatic thyroid segmentation method based on uncertain region perception, comprising the following steps:

[0006] Obtain an original ultrasound image and pre-process it;

[0007] Construct a dual-resolution network model using a deep learning library, train the dual-resolution network model, obtain an optimized dual-resolution network model, input the pre-processed ultrasound image into the optimized dual-resolution network model, and obtain segmented feature information;

[0008] Construct a convolutional pyramid uncertain region identification module, input the obtained feature information into the convolutional pyramid uncertain region identification module, repair the missing part region in the feature information, and suppress the multi-part region;

[0009] Input the feature output by the convolutional pyramid uncertain region identification module into a residual module to obtain the final thyroid segmentation feature.

[0010] The working principle and beneficial effects of the present basic scheme are that the dual-resolution network performs dual-branch feature extraction at the original image scale, which can effectively alleviate the loss of detail information, which plays an important role in improving the segmentation effect of the network, especially the segmentation effect of the edge. There is no information interaction between different resolutions in the network, which can improve the preservation of detail information as much as possible. The missing region is repaired, the multiple region is suppressed, and the segmentation accuracy is improved.

[0011] Further, the method for pre-processing the original ultrasound image is as follows:

[0012] The target region of interest is segmented from the original ultrasound image by threshold segmentation, and then the target region is adjusted to 256x256 size, and the image is normalized:

[0013]

[0014] where x∈R w×h represents the original ultrasound image, w and h represent the length and width of the image respectively; μ and σ are the mean and variance of the original image data respectively, and R represents the set of real numbers.

[0015] Pre-processing the image is beneficial for subsequent use.

[0016] Further, the method for constructing a dual-resolution network model using a deep learning library is as follows:

[0017] The dual-resolution network model uses an Encoder-Decoder framework, and the Encoder module in the Encoder-Decoder framework uses a stack of convolutional blocks for feature extraction and uses an average pooling operation with a step size of 2 to reduce the spatial resolution of the features.

[0018] Using a trilinear spline interpolation method, the deep semantic features extracted by the main path of the dual-resolution network model are upsampled, and then the rich detail features extracted by the branch path of the dual-resolution network model are fused using an addition method. The fused features are sent to the convolutional pyramid uncertain region identification module.

[0019] Simple operation, easy to use.

[0020] Further, the sample ultrasound images are obtained and divided into a training set and a test set according to a ratio of 5:1, and the GAN method is used for image enhancement of the training set;

[0021] In the training set, the SGD optimizer with kinetic energy of 0.9 is used to optimize the dual-resolution network model, the learning rate is set to 0.001, the batch is set to 6, and the maximum number of iterations is 1000 epochs.

[0022] The control of training the dual-resolution network model is performed by using an early-stopping mechanism, and the tolerance limit is set to 5 epochs, that is, when the performance of the dual-resolution network model on the test set does not improve after 5 epochs, the training of the model is terminated;

[0023] The accuracy of the model segmentation is evaluated on the test set by using the Dice index, and the value range of the Dice index is [0, 1], and the higher the value, the better the segmentation effect. The calculation of the Dice is as follows:

[0024]

[0025] Wherein, y∈R w×h represents the label of the original image, w and h represent the length and width of the image respectively, M:p=f(θ,x) represents the segmentation model, M is the abbreviation of Model, f is the abbreviation of function, θ represents the network parameter, p∈R n×w×h represents the output probability feature map, n represents the number of output classes, and the last output is the prediction result The operation of obtaining the prediction result through the feature map is argmax, which is a function of parameter (set) of a function, that is, a function of the maximum independent variable.

[0026] Image enhancement is performed on the sample ultrasound image, which is beneficial to image segmentation and optimization of segmentation effect.

[0027] Further, the method for repairing the missed region in the feature information by the convolution pyramid uncertain region identification module and suppressing the multi-region is as follows:

[0028] The convolution pyramid uncertain region identification module includes a missed region repair sub-module and a multi-region suppression sub-module.

[0029] Repair the missed region:

[0030] The feature information f in output by the dual-resolution network model is input into the missed region repair sub-module, and the feature information f in is inverted to obtain the missed region feature information and input into the convolution pyramid module for extraction of the missed region feature.

[0031] The extracted missed region feature information is added to the feature information f in to complete the repair operation of the missed region:

[0032] f MRSM =f in +α×CPM(1-f in )

[0033] Wherein, fMRSM represents the output feature of the leak repair sub-module, f in represents the original input feature, CPM represents the convolution pyramid operation; a represents the weight coefficient, which is a learnable parameter in the network;

[0034] suppressing multi-part region:

[0035] The original input feature f in is sent to the residual suppression module for multi-part region feature extraction;

[0036] The extracted multi-part region feature is added to the feature f MRSM , and the added feature is sent to the convolution module for feature fusion;

[0037] The original input feature f in is subtracted from the fused feature to obtain the output of the multi-part suppression sub-module:

[0038] f OSSM = f in - β × Conv((Res(f in )+ f MRSM ))

[0039] Wherein, f OSSM represents the output feature of the multi-part suppression sub-module, f in represents the original input feature, Res represents the residual operation, Conv represents the convolution operation, f MRSM represents the output feature of the leak repair sub-module, and β represents the learnable weight coefficient.

[0040] The leak repair sub-module learns the features of the leak region and repairs the leak region. The multi-part suppression sub-module improves the overall segmentation effect and handles possible misclassification problems.

[0041] Further, the method for obtaining the final thyroid segmentation feature is:

[0042] The original input feature f in , the output feature f MRSM of the leak repair sub-module, and the output feature f OSSM of the multi-part suppression sub-module are fused by vector splicing, and the fused feature is sent to the residual module to obtain the final feature f CPUARM :

[0043] f CPUARM = Res(cat[f in ,f MRSM ,f OSSM ])

[0044] Here, cat represents vector concatenation operation, and Res represents residual operation.

[0045] Obtain segmentation features to improve segmentation results.

[0046] Furthermore, the optimization method for the dual-resolution network model is as follows:

[0047] A deep supervision strategy is employed to optimize the dual-resolution network model. The loss function for enhancing the uncertain region is as follows:

[0048]

[0049] Where, y∈R w×h The labels represent the original image, w and h represent the length and width of the image respectively, and e is the natural constant; y i and Representing the label y and the prediction result respectively. In the image, the pixel is represented by M:p=f(θ,x), which denotes the segmentation model, θ represents the network parameters, and p∈R. n×w×h This represents the output probabilistic feature map, where n represents the number of output categories, and the final output prediction result.

[0050] In the proposed uncertainty region enhancement loss function, the following is utilized: To dynamically enhance the weights of uncertain regions, y i and The greater the difference between them, the greater their weight in the loss function, and the more the network will focus on learning their corresponding regions.

[0051] The loss function is:

[0052] in, The loss function representing deep supervision; and UUREL fusion These represent the main path, the branch path, and the loss function after fusing the main path and the branch path, respectively.

[0053] An edge loss function is introduced, and the Sobel operator is used to calculate the edges. Then, the extracted edges are dilated by 5 pixels to obtain the edge region.

[0054] y edge =Imdilate(Sobel(y),5)

[0055] wherein Imdilate and Sobel represent the dilation operation and the edge operation calculated by the Sobel operator respectively, y and y edge represent the input image and the output edge region respectively;

[0056] The overall loss function is:

[0057]

[0058] wherein UREL super represents the deep supervision loss, UREL edge represents the edge loss.

[0059] The uncertain region enhancement loss is used as the loss function to optimize the model, so as to improve the segmentation effect of the network.

[0060] The application further provides a thyroid automatic segmentation system based on uncertain region perception, which comprises a data acquisition unit and a processing unit.

[0061] The system is used to realize thyroid automatic segmentation and extract segmentation features with better segmentation effect. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is a structural schematic diagram of a convolutional pyramid module of the thyroid automatic segmentation method based on uncertain region perception of the application;

[0063] Figure 2 is a structural schematic diagram of a residual module of the thyroid automatic segmentation method based on uncertain region perception of the application;

[0064] Figure 3 is a structural schematic diagram of a convolutional pyramid uncertain region identification module of the thyroid automatic segmentation method based on uncertain region perception of the application;

[0065] Figure 4 is a structural schematic diagram of a dual-resolution network model of the thyroid automatic segmentation method based on uncertain region perception of the application. DETAILED DESCRIPTION

[0066] Embodiments of the application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the application, and cannot be understood as a limitation of the application.

[0067] In the description of the present application, it should be understood that the terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0068] In the description of the present application, unless otherwise specified and limited, it should be noted that the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be mechanical connection or electrical connection, it can be the communication between two elements, it can be direct connection or indirect connection through intermediate medium, and the specific meaning of the above terms can be understood by those skilled in the art according to the specific circumstances.

[0069] The application discloses a thyroid automatic segmentation method based on uncertain area perception, comprising the following steps:

[0070] An original ultrasound image is acquired and preprocessed;

[0071] A dual-resolution network model (DRN) is constructed by using a deep learning library, the dual-resolution network model is trained, an optimized dual-resolution network model is obtained, the preprocessed ultrasound image is input into the optimized dual-resolution network model, and segmented feature information is obtained;

[0072] In an image segmentation task, there are often missing segmentation and multiple segmentation problems. In order to improve the segmentation accuracy, it is hoped that the missing segmentation area is repaired and the multiple segmentation area is inhibited, a convolution pyramid uncertain area recognition module (CPUARM) is constructed based on this, the obtained feature information is input into the convolution pyramid uncertain area recognition module, the missing segmentation area in the feature information is repaired, and the multiple segmentation area is inhibited.

[0073] The features output by the convolution pyramid uncertain area recognition module are input into a residual module, and final thyroid segmentation features are obtained.

[0074] In a preferred scheme of the present application, the method for preprocessing the original ultrasound image is as follows:

[0075] The target area of interest is segmented from the original ultrasound image in a threshold segmentation manner, then the target area is adjusted to 256*256 size, and the image is normalized.

[0076]

[0077] where x e R w×h denotes the original ultrasound image, y e R w×h denotes the label, w and h represent the length and width of the image respectively; μ and σ are the mean and variance of the original image data respectively, and R represents the set of real numbers.

[0078] In a preferred scheme of the present application, as shown in Figure 4 The method for constructing a dual-resolution network model using a deep learning library is as follows:

[0079] The dual-resolution network model adopts an Encoder-Decoder framework, and the Encoder module in the Encoder-Decoder framework adopts a stack of convolutional blocks for feature extraction and utilizes an average pooling operation with a step size of 2 to reduce the spatial resolution of the features.

[0080] The main path is designed to extract rich semantic features, and has 4 stages, each stage having a different resolution, from top to bottom being 1 / 1, 1 / 2, 1 / 4 and 1 / 8 of the original image. An average pooling operation is adopted in the main path to adjust the resolution, and the channel numbers of the main path are set to be 64, 128, 256 and 512 respectively. Like the main path, the branch path also has 4 stages, and the channel numbers of each stage of the branch path are all 64 to reduce the computational load of the model. However, each stage has the same resolution as the original image to alleviate the loss of detail information caused by continuous downsampling. Figure 4 Residual convolution module; Avg pooling represents average pooling; main path represents the main path, which is used for the extraction of rich semantic features; Branch path represents the branch path, which is used for the extraction of detail information; Stage is a stage; Edge loss is an edge loss.

[0081] The three-linear spline interpolation method is used to gradually restore the feature map rich in semantic information extracted by the Encoder module to the original image size, and each pixel in the image is classified. The deep semantic features extracted by the main path of the dual-resolution network model are up-sampled, and then the rich detail features extracted by the branch path of the dual-resolution network model are fused in an additive manner, and the fused features are sent to the convolution pyramid uncertain region identification module. In order to improve the convergence speed of the model, auxiliary loss is added to the main path and branch path of the network. At the same time, in order to improve the edge segmentation effect, edge loss is added to the branch path.

[0082] In a preferred scheme of the present application, the collection and standard comparison of ultrasound images are time-consuming and laborious, resulting in fewer samples available for model training in clinical applications, so it is necessary to enhance the image. For example, 120 ultrasound images are collected, 100 of which are selected as a training set, and the remaining 20 are used as a test set for model verification. GAN (Generative adversarial networks, generative adversarial networks) is used to enhance the training data set by random rotation, mirroring, Gaussian noise, etc. to alleviate the overfitting problem that may occur during training.

[0083] The sample ultrasound images are obtained and divided into a training set and a test set in a ratio of 5:1, and the GAN method is used to enhance the training set;

[0084] In the training set, the SGD optimizer with kinetic energy of 0.9 is used to optimize the dual-resolution network model, the learning rate is set to 0.001, the batch is set to 6, and the maximum number of iterations is 1000 epochs;

[0085] The early-stopping mechanism is used to control the training of the dual-resolution network model, and the tolerance limit is set to 5 epochs, that is, when the performance of the dual-resolution network model does not improve on the test set for 5 epochs, the training of the model is terminated;

[0086] The Dice index is used to evaluate the accuracy of the model segmentation on the test set. The Dice index has a value range of [0, 1], and the higher the value, the better the segmentation effect. The calculation result of the Dice index is shown in Table 1:

[0087]

[0088] where y is R w×hwhere Y represents the label of the original image, w and h represent the length and width of the image respectively, M:p=f(θ,x) represents the trained segmentation model, M is the abbreviation of Model, f is the abbreviation of function, θ represents the network parameters, and p∈R n ×w×h where Y represents the label of the original image, w and h represent the length and width of the image respectively, M:p=f(θ,x) represents the trained segmentation model, M is the abbreviation of Model, f is the abbreviation of function, θ represents the network parameters, and p∈R where Y represents the label of the original image, w and h represent the length and width of the image respectively, M:p=f(θ,x) represents the trained segmentation model, M is the abbreviation of Model, f is the abbreviation of function, θ represents the network parameters, and p∈R

[0089] Table 1 thyroid segmentation effect dice (%)

[0090] Methods mean max min UNet 88.75 94.89 25.47 FCN 86.39 95.32 46.32 TransUNet 90.60 94.53 70.68 DRN 92.48 95.22 59.75

[0091] In Table 1, UNet is a method proposed by Ronneberger O et al. in "Convolutional networks for biomedical image segmentation"; FCN is a method proposed by Long J et al. in "Fully convolutional networks for semantic segmentation"; and TransUNet is a method proposed by Chen J et al. in "TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation".

[0092] In a preferred scheme of the present application, as shown in Figure 3 The method for repairing the missing region and suppressing the over-segmentation region in the feature information by the convolution pyramid uncertain region identification module is as follows:

[0093] The convolution pyramid uncertain region identification module comprises a missing repair sub-module (MRSM) and an over-segmentation suppression sub-module (OSSM);

[0094] Repairing the missing region:

[0095] The missing repair sub-module is used to learn the features of the missing region, and the feature information f in The missing repair sub-module is used to learn the features of the missing region, and the feature information f in The missing repair sub-module is used to learn the features of the missing region, and the feature information f Figure 1The missing region feature extraction is performed on the missing region;

[0096] The extracted missing region feature information is added to the feature information f in to complete the missing region repair operation:

[0097] f MRSM = f in + α × CPM (1-f in )

[0098] wherein f MRSM represents the output feature after enhancement by the missing repair sub-module, f in represents the original input feature, CPM represents the convolution pyramid operation, and α represents the weight coefficient, which is a learnable parameter in the network;

[0099] The multiple region is suppressed:

[0100] The multiple region suppression sub-module is used to learn the multiple region feature, and the original input feature f in is sent into the suppression residual module to perform the multiple region feature extraction.

[0101] The extracted multiple region feature is added to the feature f MRSM , and the added feature is sent into the convolution module to perform the feature fusion.

[0102] The original input feature f in is subtracted from the fused feature to obtain the output of the multiple region suppression sub-module:

[0103] f OSSM = f in - β × Conv ((Res (f in ) + f MRSM ))

[0104] wherein f OSSM represents the output feature of the multiple region suppression sub-module, f in represents the original input feature, Res represents the residual operation, Conv represents the convolution operation, f MRSM represents the output feature of the missing repair sub-module, and β represents the learnable weight coefficient.

[0105] The method for obtaining the final thyroid segmentation feature is:

[0106] The original input feature f in , the output feature f MRSM of the missing repair sub-module, and the output feature f OSSM of the multiple region suppression sub-module are fused in a vector splicing manner, and the fused feature is sent into the residual module (as shown in Figure 2 to obtain the final feature fCPUARM :

[0107] f CPUARM = Res(cat[f in ,f MRSM ,f OSSM ])

[0108] Wherein, cat represents vector splicing operation, Res represents residual operation.

[0109] In a preferred scheme of the application, the thyroid automatic segmentation method further comprises an optimization method for the dual-resolution network model as follows:

[0110] A deep supervision strategy is adopted to optimize the dual-resolution network model, and an uncertain region enhance loss (UREL) is as follows:

[0111]

[0112] Wherein, y is in R w×h , w and h respectively represent the length and width of the image, and e is a natural constant; y i and respectively represent the pixels in the label y and the prediction result , and the segmentation model is represented as M:p=f (theta, x), wherein theta represents network parameters, and p is in R n×w×h , represents the output probability feature map, n represents the number of output categories, and the final output prediction result

[0113] In the proposed uncertain region enhance loss function, the weight of the uncertain region is dynamically enhanced by , the greater the difference between y i and , the greater the weight in the loss function, and the network will pay more attention to the corresponding region learning;

[0114] The loss function is as follows:

[0115] Wherein, represents the loss function of deep supervision; UREL fusion respectively represent the main path, branch path, and loss function after the fusion of the main path and branch path;

[0116] An edge loss function is introduced, an edge is calculated by using a Sobel operator, and then the extracted edge is dilated by 5 pixels to obtain an edge region:

[0117] y edge = Imdilate (Sobel (y), 5)

[0118] wherein Imdilate and Sobel represent a dilate operation and an edge calculation operation by using a Sobel operator respectively, y and y edge represent an input image and an output edge region respectively;

[0119] Then, the overall loss function is:

[0120]

[0121] wherein UREL super represents a deep supervision loss, and UREL edge represents an edge loss.

[0122] The application also provides a thyroid automatic segmentation system based on uncertain region perception, comprising a data acquisition unit and a processing unit, the data acquisition unit is used for acquiring an original ultrasound image, an output end of the data acquisition unit is electrically connected with an input end of the processing unit, and the processing unit executes the method and performs thyroid automatic segmentation.

[0123] The dual-resolution network provided in the technical solution is used for extracting features in a dual branch from a first stage, and high-resolution features are extracted in the original image scale, which can effectively alleviate the loss of detail information and improve the segmentation effect of the network, especially the segmentation effect of edges. In the dual-resolution network, no information interaction is performed between different resolutions, and the processing is mainly used to improve the retention of detail information as much as possible. In order to handle the misclassification problem, a multi-suppression sub-module is designed to improve the overall segmentation effect.

[0124] The model provided in the application uses a trilinear spline interpolation method in the decoder to up-sample deep semantic features extracted by the main path, and then fuses the up-sampled features with rich detail features of the branch path by using an addition method, and finally sends the fused features into the proposed convolution pyramid uncertain region identification module to suppress misclassified regions and repair missed regions, and outputs the final segmentation result. In this way, the calculation amount of the model can be reduced, and the loss of detail information in the layer-by-layer up-sampling of the model can be alleviated.

[0125] In the description of the specification, reference to "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that a particular feature, structure, material, or characteristic being described is included in at least one embodiment or example of the application. The appearances of the above expressions in various places in the specification are not necessarily referring to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0126] Although embodiments of the application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and alterations can be made hereto without departing from the spirit and scope of the application, which is defined by the following claims and their equivalents.

Claims

1. An automatic thyroid segmentation method based on uncertain region perception, characterized in that, Includes the following steps: Acquire raw ultrasound images and preprocess them; A dual-resolution network model was constructed using a deep learning library. The dual-resolution network model was trained to obtain an optimized dual-resolution network model. The preprocessed ultrasound image was then input into the optimized dual-resolution network model to obtain segmented feature information. A convolutional pyramid uncertain region identification module is constructed. The obtained feature information is input into the convolutional pyramid uncertain region identification module to repair the missing regions in the feature information and suppress the over-segmented regions. The features output by the convolutional pyramid uncertain region identification module are input into the residual module to obtain the final thyroid segmentation features; The method for constructing a dual-resolution network model using deep learning libraries is as follows: The dual-resolution network model adopts an Encoder-Decoder framework. The Encoder module in the Encoder-Decoder framework uses stacked convolutional blocks for feature extraction and uses average pooling with a stride of 2 to reduce the spatial resolution of the features. The deep semantic features extracted by the main path of the dual-resolution network model are upsampled using the trilinear spline interpolation method. Then, they are fused with the rich detail features extracted by the branch path of the dual-resolution network model by addition. The fused features are then fed into the convolutional pyramid uncertain region identification module. The method used by the convolutional pyramid uncertain region identification module to repair missing regions in feature information and suppress over-segmented regions is as follows: The convolutional pyramid uncertainty region identification module includes a missing segment repair submodule and a multi-segment suppression submodule; Repairing areas where points were missed: Feature information output by the dual-resolution network model The input omission repair submodule processes feature information. The missing region feature information is obtained by inverting the value and then input into the convolutional pyramid module to extract the missing region features. The extracted missing region feature information and feature information Add them together to complete the repair operation for the missing areas: )， in, This indicates the output characteristics after enhancement by the omission repair submodule. Represents the original input features. This represents the convolution pyramid operation; These represent weight coefficients, which are learnable parameters in the network. Suppress multi-partition areas: Original input features The data is fed into the residual suppression module for multi-region feature extraction. Extracted multi-region features and features The features are added together, and the resulting features are fed into the convolution module for feature fusion. Original input features Subtracting the fused features yields the output of the multi-suppression submodule: ， in, This represents the output characteristics of the multi-suppression submodule. Represents the original input features. This indicates residual operations. This represents the convolution operation. This indicates the output characteristics of the missing score repair submodule. This represents the learnable weight coefficients.

2. The automatic thyroid segmentation method based on uncertain region perception as described in claim 1, characterized in that, The method for preprocessing the raw ultrasound images is as follows: Thresholding segmentation was used to segment the target region of interest from the original ultrasound image. The target region was then resized to 256×256 pixels and the image was normalized. ， in, Represents the original ultrasound image. and These represent the length and width of the image, respectively. , Here, denoted as the mean and variance of the original image data, and R represents the set of real numbers.

3. The automatic thyroid segmentation method based on uncertain region perception as described in claim 1, characterized in that, Sample ultrasound images were acquired and divided into training and test sets in a 5:1 ratio. The training set was then augmented using the GAN method. In the training set, the dual-resolution network model was optimized using the SGD optimizer with a kinetic energy of 0.9, the learning rate was set to 0.001, the batch size was set to 6, and the maximum number of iterations was 1000 epochs. An early-stopping mechanism is used to control the training of the dual-resolution network model. The tolerance limit is set to 5 epochs, that is, if the performance of the dual-resolution network model does not improve after 5 epochs on the test set, the training of the model is terminated. The Dice metric is used to evaluate the segmentation accuracy of the model on the test set. The Dice metric ranges from [0, 1], with higher values ​​indicating better segmentation performance. The Dice metric is calculated as follows: ， in, Labels representing the original image, and Let the length and width of the image be represented by , respectively. This represents a segmentation model, where M is short for Model and f is short for function. Represents network parameters, This represents the output probability feature map. This indicates the number of categories in the output, and the final output prediction result. , The operation of obtaining prediction results through feature maps, argmax is a function that calculates the parameters of a function, that is, the function that maximizes the independent variable.

4. The automatic thyroid segmentation method based on uncertain region perception as described in claim 1, characterized in that, The method for obtaining the final thyroid segmentation features is as follows: The original input features Output characteristics of the missing score repair submodule Output characteristics of the multi-suppression submodule The feature fusion is performed using vector concatenation, and the fused features are then fed into the residual module to obtain the final features. : ， in, This represents a vector concatenation operation. This indicates residual operations.

5. The automatic thyroid segmentation method based on uncertain region perception as described in claim 1, characterized in that, The optimization methods for dual-resolution network models are also included as follows: A deep supervision strategy is employed to optimize the dual-resolution network model. The loss function for enhancing the uncertain region is as follows: ， in, Labels representing the original image, and These represent the length and width of the image, respectively. It is a natural constant; and Representing tags respectively and prediction results The pixels in the middle, using Represents the segmentation model. Represents network parameters, This represents the output probability feature map. This indicates the number of categories in the output, and the final output prediction result. , ; In the proposed uncertainty region enhancement loss function, the following is utilized: To dynamically enhance the weights of uncertain regions, and The greater the difference between them, the greater their weight in the loss function, and the more the network will focus on learning their corresponding regions. The loss function is: , in, The loss function representing deep supervision; , ,and These represent the main path, the branch path, and the loss function after fusing the main path and the branch path, respectively. An edge loss function is introduced, and the Sobel operator is used to calculate the edges. Then, the extracted edges are dilated by 5 pixels to obtain the edge region. ， in, and These represent the dilation operation and the edge computation operation using the Sobel operator, respectively. and These represent the input image and the output edge region, respectively. The overall loss function is then: ， in, This indicates damage caused by deep surveillance. This indicates marginal loss.

6. An automatic thyroid segmentation system based on uncertain region perception, characterized in that, The device includes a data acquisition unit and a processing unit. The data acquisition unit is used to acquire raw ultrasound images. The output end of the data acquisition unit is connected to the input end of the processing unit. The processing unit performs the method described in any one of claims 1-5 to automatically segment the thyroid gland.

Images