A semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features

Abstract

This paper proposes a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features. The method includes five steps: 1) image preprocessing; 2) feature extraction based on convolution and wavelet transform; 3) construction of a semi-supervised deep learning model based on Mean-Teacher; 4) edge feature evaluation based on convexity; and 5) edge feature evaluation based on active contours. The present invention uses wavelet transform to suppress speckle noise in ultrasound images, addressing the issue of speckle noise in thyroid ultrasound images. Convexity loss function and active contour loss function are used to evaluate edge features for semi-supervised semantic segmentation, enabling high-precision segmentation of thyroid organs from ultrasound images.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology and relates to a semi-supervised thyroid ultrasound image segmentation method for ultrasound images. Specifically, it relates to research in the fields of image processing, medical image analysis, discrete wavelet transform, semi-supervised learning and thyroid ultrasound image segmentation. The present invention is a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features. Background Art

[0002] Goiter is a common condition with complex causes, including malnutrition, genetic factors, hormonal imbalances, and immune system abnormalities. Goiter can be categorized into two types: diffuse goiter and nodular goiter. Diffuse goiter refers to irregular growth of the entire thyroid gland, while nodular goiter refers to irregular growth of one or more nodules within the thyroid gland. Goiter varies in severity; mild cases may not require treatment, while severe cases may require surgery.

[0003] Ultrasound examination is a widely used technique for diagnosing goiters, helping doctors determine their location, size, and shape. However, characteristics of ultrasound images, such as unclear organ outlines and image noise, pose challenges to subjective diagnosis. Furthermore, the morphology of goiters varies greatly, and overlap between the thyroid gland and adjacent tissues complicates diagnosis. Consequently, automated thyroid segmentation methods have become a hot topic of research in the medical community. In addition to deep learning-based methods, other medical image segmentation approaches exist, such as rule-based methods and traditional machine learning-based methods. Rule-based methods leverage the knowledge of human experts to design rules that separate the thyroid region from other regions in ultrasound images. While these methods can achieve high accuracy, they require significant expert input in rule design and are difficult to adapt to the variability of thyroid morphology. Traditional machine learning-based methods typically use handcrafted features to represent ultrasound images and then employ classifiers or clustering algorithms for segmentation. These methods require manual feature selection and extraction, and rely on expert experience and knowledge.

[0004] However, with the continuous development and optimization of deep learning methods, deep learning-based medical image segmentation methods have surpassed traditional methods and gradually become the mainstream methods for medical image segmentation. Deep learning-based methods have the following advantages: first, they do not require manual feature design, but instead automatically extract features by learning from a large amount of data, thus being able to adapt to different thyroid morphologies; second, they can process large amounts of data, thereby improving accuracy and robustness; finally, they can learn more complex features, thus providing more comprehensive medical information.

[0005] In recent years, deep learning methods have achieved tremendous success in the field of medical image semantic segmentation. Convolutional neural networks (CNNs), one of the most widely used models in deep learning, have been shown to perform exceptionally well in this area. However, deep learning methods require a large number of annotated ultrasound images to train the models. Preparing this data requires significant human and time resources, even requiring specialized expert input. This limits the application of these methods.

[0006] To address this issue, semi-supervised learning methods have been proposed. They utilize a combination of labeled and unlabeled data to train models, reducing the need for labeled data and improving the model's generalization ability. Pseudo-labeling and co-training are two commonly used semi-supervised learning methods. Pseudo-labeling methods use a pre-trained model to pseudo-label unlabeled data and then use this pseudo-labeled data to train the model. Co-training methods use two or more different models, training them separately on labeled and unlabeled data, then alternately updating these models so that they learn from each other. Semi-supervised learning is a machine learning method that uses a combination of labeled and unlabeled data to train models. It has been widely used in the field of medical image segmentation. Semi-supervised learning methods are of great significance in medical image segmentation because preparing labeled data requires significant human resources and time, even requiring the participation of professional experts. Medical image segmentation is a task that often requires a large amount of labeled data. Therefore, semi-supervised learning methods can effectively utilize limited labeled data while simultaneously leveraging abundant unlabeled data, thereby improving the model's generalization ability.

[0007] Pseudo-labeling is one of the most commonly used methods in semi-supervised learning. It pseudo-labels unlabeled data using a pre-trained model, and then uses this pseudo-labeled data to train the model. Pseudo-labeled data can be the model's predictions on the unlabeled data or a subset of manually labeled data. Pseudo-labeling methods have the advantages of being simple and easy to implement, and can improve model accuracy and robustness. However, they also have drawbacks, such as the unreliable quality of pseudo-labeled data and shifts in the distribution of training data, which can lead to overfitting of the model or decreased generalization performance.

[0008] Co-training is another commonly used semi-supervised learning method. It utilizes multiple models trained on different datasets, iteratively updating the models in alternating fashion. In co-training, each model is trained using both labeled and unlabeled data. Their predictions are then exchanged and used to enhance the training of the other models. The advantage of co-training is that it leverages the strengths of multiple models, improving their accuracy and robustness. However, it also has drawbacks, such as strong coupling between models, which can easily lead to misleading predictions.

[0009] In general, semi-supervised learning methods are of great significance in medical image segmentation. With the continuous development and optimization of deep learning technology, the application of semi-supervised learning methods is also expanding. In the future, researchers can expect to further develop semi-supervised learning methods in medical image segmentation, thereby better utilizing limited labeled data and improving the accuracy and robustness of medical image segmentation. Summary of the Invention

[0010] This invention aims to improve the accuracy and robustness of medical image segmentation. It provides a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features. This method incorporates frequency domain feature extraction and edge profile feature evaluation. A computer-assisted medical system based on this method can improve the accuracy and robustness of medical image segmentation.

[0011] The present invention is achieved through the following technical solutions. The present invention proposes a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features, the method comprising the following steps:

[0012] Step 1: ROI selection and preprocessing. The ROI size was set to 224*224. When selecting the ROI, the prominent part of the thyroid region with clear texture was selected, and the uniform part that could highlight the characteristics of the thyroid parenchyma was selected. The grayscale value of the ROI image was mapped to the range of 0-255.

[0013] Step 2: Feature extraction based on convolution and wavelet transform;

[0014] Step 3: Build a semi-supervised deep learning model based on Mean-Teacher;

[0015] Step 4: Edge feature evaluation based on convexity;

[0016] Step 5: Edge feature evaluation based on active contour.

[0017] Furthermore, in step 2, the encoder operation process is specifically as follows:

[0018] Step 2: Perform convolution operation on ROI to obtain spatial domain feature map X;

[0019] Step 22: Perform two-dimensional discrete wavelet transform (2D-DWT) on X to obtain a low-frequency subgraph LL and high-frequency subgraphs LH, HL, and HH;

[0020] Step 2: Perform convolution operation on the low-frequency sub-image LL to reduce the impact of high-frequency noise on feature extraction and obtain the feature image Y.

[0021] Furthermore, in step 2, the decoder operation process is specifically as follows:

[0022] Step 24: Concatenate the feature map Z with the high-frequency sub-map obtained by the encoder at the same level, integrate the frequency domain information, and more comprehensively represent the texture information of the image to obtain the feature map T;

[0023] Step 25: Perform inverse discrete wavelet transform IDWT on the feature map T to obtain the spatial domain feature map R;

[0024] Step 26: Perform a convolution operation on the spatial feature map R, decode the information of the feature map, and obtain a fused image that contains the frequency characteristics and semantic information of the original image.

[0025] Furthermore, the semi-supervised deep learning model includes a student model and a teacher model; the network parameters of the student model are obtained through learning and gradient descent; the network parameters of the teacher model are obtained through the moving average of the network parameters of the student model; and the network parameters of the teacher model are updated by the moving average calculation of the network parameters of the student model.

[0026] Furthermore, the network parameter update of the teacher model is specifically as follows:

[0027] Suppose there are a batch of training samples X1 and X2, where X1 is labeled data and the corresponding label is Z1, and X2 is unlabeled data; the specific process is as follows:

[0028] (1) Take this batch of samples as the input of the student network, and then get the output labels: Ys1, Ys2 respectively;

[0029] (2) Construct the loss function for labeled data X1, the labeled classification loss function L1(Z1, Ys1);

[0030] (3) Take this batch of data as the input of the teacher model and obtain the output labels Yt1, Yt2;

[0031] (4) Construct an unsupervised loss function L2 and use the mean square error loss function;

[0032] (5) The total loss function L1+L2 gradient descent is used to update the network parameters of the student model, and the network parameters of the teacher model are updated by moving average.

[0033] Furthermore, in step 4, convexity is the relative amount by which the image object differs from the convex hull, and the measure of convexity can be obtained by the ratio of the perimeter of the object's convex hull to the perimeter of the object itself:

[0034]

[0035] The closer the convexity is to 1, the smoother the image boundary is, and the closer the convexity is to 0, the more burrs there are on the image boundary.

[0036] Define the loss function Convexity Loss:

[0037]

[0038] Mean Convexity is the average convexity of all thyroid ultrasound image labels, which is approximately the standard value of thyroid convexity.

[0039] Furthermore, the step 4 is specifically as follows:

[0040] (1) Calculate the coordinates of the thyroid center of mass;

[0041] (2) Divide the annotated image into two parts: above the centroid and below the centroid;

[0042] (3) Initialize an empty list BoundList;

[0043] (4) For each column of pixels above the centroid, traverse from left to right to find the pixel point belonging to the thyroid boundary, calculate the distance between the point and the centroid, and add the distance to BoundList;

[0044] (5) For each column of pixels below the centroid, traverse from right to left to find the pixel point belonging to the thyroid boundary, calculate the distance between the point and the centroid, and add the distance to BoundList;

[0045] (6) Initialize the concavity counter to 0;

[0046] (7) Traverse BoundList and find convex and concave points. If the left and right elements of an element are both smaller than the element, it is a convex point; if both are larger than the element, it is a concave point. Whenever the convexity and concaveness change, the counter is incremented by one.

[0047] (8) Calculate convexity: Divide the value of the counter by the length of BoundList.

[0048] Furthermore, in step five, a two-dimensional active contour loss function is introduced, which is defined as follows:

[0049] Loss AC =Length+λ·Region

[0050]

[0051] Region=∫ Ω ((c1-v)2-(c2-v)2)udx

[0052] Therefore, Length and Region can be written pixel-wise as follows:

[0053]

[0054]

[0055] The present invention proposes an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features are implemented.

[0056] The present invention proposes a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features are implemented.

[0057] Compared with the existing technology, the present invention has the following advantages: the present invention uses the active contour energy function as the loss function in semi-supervised learning, calculates the global spatial information of the image through the energy function, and solves the semantic segmentation problem by minimizing the energy function for edge feature evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0059] Figure 1 This is a flow chart of a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features according to the present invention;

[0060] Figure 2 It is the structure diagram of wavelet U-type network;

[0061] Figure 3 This is a flowchart of the Mean-Teacher semi-supervised learning strategy;

[0062] Figure 4 This is the flow chart of the convexity loss function algorithm. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0064] Combine Figure 1-Figure 4 The present invention proposes a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features, the method comprising the following steps:

[0065] Step 1: ROI selection and preprocessing. The ROI size was set to 224*224. When selecting the ROI, the prominent part of the thyroid region with clear texture was selected, and the uniform part that could highlight the characteristics of the thyroid parenchyma was selected. The grayscale value of the ROI image was mapped to the range of 0-255.

[0066] 1) Select a ROI in the ultrasound image with a prominent thyroid region and clear texture, and set the size of the ROI to a 224*224 square area;

[0067] 2) To avoid the uneven brightness of the selected ROI image due to human factors, the grayscale value of the ROI image is mapped to the range of 0-255, and the grayscale stretching is performed according to the following formula:

[0068]

[0069] Where G(x, y) represents the grayscale of the original image, W min and W max are the maximum and minimum values ​​of G(x, y) respectively, and G(x, y)' is the grayscale value of the stretched image.

[0070] Step 2: Feature extraction based on convolution and wavelet transform;

[0071] Ultrasonic images have the problem of speckle noise, which makes the texture unclear and affects the segmentation effect. Therefore, when extracting image features, it is necessary to consider both the low-frequency grayscale information and high-frequency detail information of the image texture. This paper proposes a wavelet U-net, which uses wavelet transform to suppress the influence of ultrasonic inherent speckle noise on the image, and combines the low-frequency grayscale information and high-frequency detail information of the image to more comprehensively represent the texture information of the image. The network structure is as follows: Figure 2 shown.

[0072] In step 2, the encoder operation process is as follows:

[0073] Step 2: Perform convolution operation on ROI to obtain spatial domain feature map X;

[0074] Step 22: Perform two-dimensional discrete wavelet transform (2D-DWT) on X to obtain a low-frequency subgraph LL and high-frequency subgraphs LH, HL, and HH;

[0075] Step 2: Perform convolution operation on the low-frequency sub-image LL to reduce the impact of high-frequency noise on feature extraction and obtain the feature image Y.

[0076] In step 2, the decoder operation process is as follows:

[0077] Step 24: Concatenate the feature map Z with the high-frequency sub-map obtained by the encoder at the same level, integrate the frequency domain information, and more comprehensively represent the texture information of the image to obtain the feature map T;

[0078] Step 25: Perform inverse discrete wavelet transform IDWT on the feature map T to obtain the spatial domain feature map R;

[0079] Step 26: Perform a convolution operation on the spatial feature map R, decode the information of the feature map, and obtain a fused image that contains the frequency characteristics and semantic information of the original image.

[0080] Step 3: Build a semi-supervised deep learning model based on Mean-Teacher;

[0081] In semi-supervised learning, the Mean-Teacher method improves the robustness and generalization performance of the model by introducing the concepts of teacher model and student model. In this method, the teacher model and the student model share the same network structure, but their parameters are different. The parameters of the teacher model are obtained by weighted average of the parameters of the last few steps of multiple student models. The student model uses the target generated by the teacher model as its learning target. The overall architecture of the Mean-Teacher network consists of two parts: the student model and the teacher model: the network parameters of the student model are obtained through learning and gradient descent. The network parameters of the teacher model are obtained by the moving average of the network parameters of the student model. The network parameters of the teacher model are updated by the moving average calculation of the network parameters of the student model. The specific algorithm flow is as follows Figure 3 shown.

[0082] The network parameter update of the teacher model is specifically as follows:

[0083] Suppose there are a batch of training samples X1 and X2, where X1 is labeled data and the corresponding label is Z1, and X2 is unlabeled data; the specific process is as follows:

[0084] (1) Take this batch of samples as the input of the student network, and then get the output labels: Ys1, Ys2 respectively;

[0085] (2) Construct the loss function for labeled data X1, the labeled classification loss function L1(Z1, Ys1);

[0086] (3) Take this batch of data as the input of the teacher model and obtain the output labels Yt1, Yt2;

[0087] (4) Construct an unsupervised loss function L2 and use the mean square error loss function;

[0088] (5) The total loss function L1+L2 gradient descent is used to update the network parameters of the student model, and the network parameters of the teacher model are updated by moving average.

[0089] Step 4: Edge feature evaluation based on convexity;

[0090] In step 4, convexity is the relative amount by which the image object differs from the convex hull. This measure of convexity can be obtained by taking the ratio of the perimeter of the object's convex hull to the perimeter of the object itself:

[0091]

[0092] The closer the convexity is to 1, the smoother the image boundary is, and the closer the convexity is to 0, the more burrs on the image boundary are; therefore, convexity can be used to more accurately describe the edge information of the image.

[0093] Define the loss function Convexity Loss:

[0094]

[0095] Mean Convexity is the average convexity of all thyroid ultrasound image labels, which is approximately the standard value of thyroid convexity.

[0096] Convexity is a feature of an image that describes the degree of convexity or concavity of the edge in the image. When performing image segmentation tasks, convexity is usually used to evaluate the edge quality of the segmentation results, and convexity is used as a training loss to guide the learning process of the model. For the thyroid image segmentation task, convexity loss can help the model better deal with edge clarity and artifact problems, thereby improving the accuracy and stability of the segmentation results. The convexity index is used as a loss function to solve problems such as unclear edges or pseudo edges in thyroid ultrasound images. The convexity loss function algorithm process is as follows: Figure 4 shown.

[0097] The step 4 is specifically as follows:

[0098] (1) Calculate the coordinates of the thyroid center of mass;

[0099] (2) Divide the annotated image into two parts: above the centroid and below the centroid;

[0100] (3) Initialize an empty list BoundList;

[0101] (4) For each column of pixels above the centroid, traverse from left to right to find the pixel point belonging to the thyroid boundary, calculate the distance between the point and the centroid, and add the distance to BoundList;

[0102] (5) For each column of pixels below the centroid, traverse from right to left to find the pixel point belonging to the thyroid boundary, calculate the distance between the point and the centroid, and add the distance to BoundList;

[0103] (6) Initialize the concavity counter to 0;

[0104] (7) Traverse BoundList and find convex and concave points. If the left and right elements of an element are both smaller than the element, it is a convex point; if both are larger than the element, it is a concave point. Whenever the convexity and concaveness change, the counter is incremented by one.

[0105] (8) Calculate convexity: Divide the value of the counter by the length of BoundList.

[0106] Step 5: Edge feature evaluation based on active contour.

[0107] An active contour model is a contour that moves and deforms under the influence of internal forces within the curve itself and external constraints from image data. The forces acting on the active contour model vary locally in space, depending on the position and shape of the contour. Internal and external forces act differently: internal forces act as smoothing constraints, while external forces guide the active contour model toward image features. External forces applied to the active contour model originate from image or higher-level processing forces, pushing the active contour model away from undesirable properties.

[0108] In order to efficiently find an active contour that globally minimizes the active contour energy and thus achieve image segmentation, a two-dimensional active contour loss function is introduced. The two-dimensional active contour loss function of the present invention is defined as follows:

[0109] Loss AC =Length+λ·Region

[0110]

[0111] Region=∫ Ω ((c1-v) 2 -(c2-v) 2 )udx

[0112] Therefore, Length and Region can be written pixel-wise as follows:

[0113]

[0114]

[0115] The present invention uses the active contour energy function as a loss function in semi-supervised learning. The energy function is used to calculate the global spatial information of the image, and the semantic segmentation problem is solved by minimizing the energy function for edge feature evaluation.

[0116] Example

[0117] The following describes the specific implementation of the present invention with reference to the following examples and drawings: A semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features is applied to goiter image segmentation.

[0118] All thyroid ultrasound images and their segmentation annotations used in this experiment were provided by the National Center for Endemic Disease Control, Chinese Center for Disease Control and Prevention. After several semi-supervised learning experiments comparing labeled and unlabeled images at different ratios, a 1:1 ratio was found to be the optimal dataset distribution. There were 1,362 annotated thyroid images and 1,362 unannotated thyroid images, resulting in a total of 2,724 training images and 444 test images.

[0119] Perform step 1: Select a prominent thyroid region with clear texture in the thyroid ultrasound image, and a relatively uniform region that highlights the characteristics of the thyroid parenchyma. Select a 224x224 area within this region as the target ROI. Then, preprocess all ROIs and map the grayscale values ​​of each ROI image to a range of 0-255 using the grayscale mapping formula.

[0120] Step 2: Feature extraction based on convolution and wavelet transform. The preprocessed ROI image undergoes convolution and a two-dimensional discrete wavelet transform (2D-DWT) to extract a spatial feature map. Convolution is also performed on the low-frequency sub-map to reduce the impact of high-frequency noise. Next, the decoder concatenates the feature map with the high-frequency sub-map obtained by the encoder at the same level and performs an inverse discrete wavelet transform (IDWT) to integrate frequency domain information and more comprehensively represent image texture information. Finally, a convolution operation is performed on the spatial feature map to decode the feature map information, generating an output image that combines the frequency characteristics and semantic information of the original image.

[0121] Execute step three: Build a semi-supervised deep learning model based on Mean-Teacher. This model consists of two sub-networks: a student network and a teacher network. The student network is responsible for learning the feature representation of the input image and generating segmentation results, while the teacher network is responsible for generating high-quality pseudo-labels to guide the student network's training. During training, the parameters of the teacher network are the exponential moving average of the student network's parameters. The model includes components such as convolutional layers, activation functions, and pooling layers. The input is the fused image, and the output is the segmentation result. Through this semi-supervised learning approach, the model can achieve good segmentation performance even with limited labeled data.

[0122] Proceed to step 4: Convexity-based edge feature evaluation. Calculate the convexity loss function to evaluate boundary features. The convexity loss function measures the difference in convexity between the segmentation result and the actual boundary. By calculating the distance from each pixel to the centroid and combining it with a convexity counter, the concavity and convexity of the image boundary can be effectively measured. The convexity loss function guides the model to focus more on boundary information, helping to improve the accuracy of boundary segmentation.

[0123] Execute step five: Active Contour-Based Edge Feature Evaluation. Calculate the active contour loss function to enhance boundary details. The active contour model achieves dynamic contour evolution by internally enforcing the contour curve and externally constraining the image data. The internal enforcement smoothes the contour, while the external enforcement guides the contour toward image features. By calculating the active contour loss function, the model can better fit the shape of the target area while preserving boundary details.

[0124] By running and debugging the above five steps, the semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features is completed. The thyroid semantic segmentation results based on the nnU-Net benchmark network are shown in the following table:

[0125] Loss Function DICE IOU PA convexity Cross entropy only 87.83％ 78.74％ 91.99％ 88.74％ Cross entropy + convexity 88.34％ 80.21％ 92.01％ 90.13％ Cross Entropy + Active Contour 88.56％ 80.34％ 92.17％ 90.24％ Cross entropy + convexity + active contour 88.49％ 80.73％ 92.09％ 90.37％

[0126] The results of thyroid semantic segmentation based on wavelet U-type network are shown in the following table:

[0127] Loss Function DICE IOU PA convexity Cross entropy only 90.00％ 82.14％ 95.57％ 89.93％ Cross entropy + convexity 92.15％ 86.08％ 95.19％ 90.24％ Cross Entropy + Active Contour 91.54％ 85.03％ 95.71％ 90.10％ Cross entropy + convexity + active contour 91.39％ 87.69％ 96.65％ 91.17％

[0128] This paper discloses a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features. The proposed method was used to segment 3168 thyroid ultrasound images, achieving a Distance-Intensity-of-Intersection (DICE) of 93.39%, an Interconnection Under One Unit (IOU) of 87.69%, a Parity-Average (PA) of 96.65%, and a Convexity of 91.17%. Comparative experiments show that the proposed method can improve DICE by 5.56%, IOU by 8.95%, PA by 4.66%, and Convexity by 2.43%.

[0129] The present invention proposes an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features are implemented.

[0130] The present invention proposes a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features are implemented.

[0131] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM bus RAM (DRRAM). It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0132] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated therein. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a high-density digital video disc (DVD)), or a semiconductor medium (eg, a solid state disc (SSD)).

[0133] During implementation, each step of the above method can be completed by an integrated logic circuit of the hardware in the processor or by instructions in the form of software. The steps of the method disclosed in conjunction with the embodiments of the present application can be directly embodied as being executed by a hardware processor, or can be executed by a combination of hardware and software modules in the processor. The software module can be located in a storage medium mature in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in conjunction with its hardware. To avoid repetition, it will not be described in detail here.

[0134] It should be noted that the processor in the embodiments of the present application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiment can be completed by an integrated logic circuit of the hardware in the processor or by instructions in the form of software. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. The various methods, steps, and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly embodied as being executed by a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0135] The above is a detailed introduction to the semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features proposed in the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A semi-supervised thyroid ultrasound image segmentation method based on discrete wavelet transform and edge features, characterized by: The method comprises the following steps: Step 1: ROI selection and preprocessing. The ROI size was set to 224*224. When selecting the ROI, the prominent part of the thyroid region with clear texture was selected, and the uniform part that could highlight the characteristics of the thyroid parenchyma was selected. The grayscale value of the ROI image was mapped to the range of 0-255. Step 2: Feature extraction based on convolution and wavelet transform; Step 3: Build a semi-supervised deep learning model based on Mean-Teacher; Step 4: Edge feature evaluation based on convexity; In step 4, convexity is the relative amount by which the image object differs from the convex hull. This measure of convexity can be obtained by taking the ratio of the perimeter of the object's convex hull to the perimeter of the object itself: The closer the convexity is to 1, the smoother the image boundary is, and the closer the convexity is to 0, the more burrs there are on the image boundary. Define the loss function Convexity Loss: Convexity Loss = | Mean Convexity - Convexity | = | Mean Convexity - | Mean Convexity is the average value of the convexity of all thyroid ultrasound image labels, which is approximately the standard value of the thyroid convexity; The step 4 is specifically as follows: (1) Calculate the coordinates of the thyroid center of mass; (2) Divide the annotated image into two parts: above the centroid and below the centroid; (3) Initialize an empty list BoundList; (4) For each column of pixels above the centroid, traverse from left to right to find the pixel point belonging to the thyroid boundary, calculate the distance between the point and the centroid, and add the distance to BoundList; (5) For each column of pixels below the centroid, traverse from right to left to find the pixel point belonging to the thyroid boundary, calculate the distance between the point and the centroid, and add the distance to BoundList; (6) Initialize the concavity counter to 0; (7) Traverse BoundList and find convex and concave points. If the left and right elements of an element are both smaller than the element, it is a convex point; if both are larger than the element, it is a concave point. Whenever the convexity and concaveness change, the counter is incremented by one. (8) Calculate convexity: divide the counter value by the length of BoundList; Step 5: Edge feature evaluation based on active contour; In step five, the two-dimensional active contour loss function is introduced, which is defined as follows: Loss AC = Length + λ · Region Length = Region = So, Length and Region They can be written pixel-wise as follows: Length = Region = 。 2. The method according to claim 1, wherein: In step 2, the encoder operation process is as follows: Step 2: Perform convolution operation on ROI to obtain spatial domain feature map X; Step 22: Perform two-dimensional discrete wavelet transform (2D-DWT) on X to obtain a low-frequency subgraph LL and high-frequency subgraphs LH, HL, and HH; Step 2: Perform convolution operation on the low-frequency sub-image LL to reduce the impact of high-frequency noise on feature extraction and obtain the feature image Y.

3. The method according to claim 2, wherein: In step 2, the decoder operation process is as follows: Step 24: Concatenate the feature map Z with the high-frequency sub-map obtained by the encoder at the same level, integrate the frequency domain information, and more comprehensively represent the texture information of the image to obtain the feature map T; Step 25: Perform inverse discrete wavelet transform IDWT on the feature map T to obtain the spatial domain feature map R; Step 26: Perform a convolution operation on the spatial feature map R, decode the information of the feature map, and obtain a fused image that contains the frequency characteristics and semantic information of the original image.

4. The method according to claim 1, wherein: The semi-supervised deep learning model includes a student model and a teacher model; the network parameters of the student model are obtained through learning and gradient descent; the network parameters of the teacher model are obtained through the moving average of the network parameters of the student model; and the network parameters of the teacher model are updated by the moving average calculation of the network parameters of the student model.

5. The method according to claim 4, characterized in that: The network parameter update of the teacher model is specifically as follows: Suppose there are a batch of training samples X1 and X2, where X1 is labeled data and the corresponding label is Z1, and X2 is unlabeled data; the specific process is as follows: (1) Take this batch of samples as the input of the student network, and then get the output labels: Ys1, Ys2 respectively; (2) Construct the loss function for labeled data X1, the labeled classification loss function L1(Z1, Ys1); (3) Take this batch of data as the input of the teacher model and obtain the output labels Yt1, Yt2; (4) Construct an unsupervised loss function L2 and use the mean square error loss function; (5) The total loss function L1+L2 gradient descent is used to update the network parameters of the student model, and the network parameters of the teacher model are updated by moving average.

6. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

7. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images