Boundary enhanced ultrasonic image segmentation method based on double-decoder structure

Abstract

The invention belongs to the technical field of medical image processing and deep learning crossing, and particularly provides a boundary-enhanced ultrasonic image segmentation method based on a double-decoder structure, which is used for solving the problems of unsmooth boundary and insufficient precision of ultrasonic image segmentation in the prior art. According to the method, an ultrasonic image segmentation network comprising a main decoder (Seg-decoder) and a boundary decoder (Bor-decoder) is constructed, the Bor-decoder is utilized to focus on processing boundary pixel segmentation, boundary feature information is provided for the Seg-decoder through a feature fusion mechanism, the overall segmentation capability of the model is improved, and meanwhile, a multi-scale structure is introduced to enhance the adaptability to a complex target; according to the invention, by introducing a boundary feature enhancement mechanism and a multi-scale structure design, ultrasonic image segmentation with higher precision is realized; the segmentation performance and the boundary segmentation quality are improved, the reliability of a diagnosis result is remarkably improved, and the method has important research value and application prospects in the field of medical image automatic segmentation.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary technical field of medical image processing and deep learning, and specifically provides a boundary-enhanced ultrasound image segmentation method based on a dual decoder structure, which is used to improve the accuracy of ultrasound image segmentation and the ability to capture boundary information. Background Art

[0002] In modern medical imaging technology, ultrasound imaging is widely used in clinical scenarios such as obstetric examinations, cardiovascular disease assessment, and organ structure diagnosis due to its low cost, no radiation, high portability, and real-time imaging, becoming an important tool for medical diagnosis. However, due to the limitations of the imaging principle, ultrasound images usually have problems such as low resolution, blurred boundaries, and high noise interference, which significantly increases the difficulty for doctors to interpret images, especially when it is necessary to accurately locate lesions and segment target areas. Compared with other imaging methods (such as MRI and CT), ultrasound images are more dependent on the experience of doctors, especially in areas with limited medical resources. The lack of experienced doctors further exacerbates the difficulty of diagnosis.

[0003] In recent years, with the development of deep learning technology, the automation of medical image segmentation has been greatly promoted. Many studies have applied convolutional neural networks (CNNs) to accurately segment regions of interest in medical images. In this field, U-Net and its variants have become mainstream segmentation models due to their adaptability to small sample scenarios and superior segmentation performance. However, although these methods perform well in overall segmentation tasks, they usually have weak processing capabilities for boundary pixels. The reason is that mainstream segmentation models use the same processing strategy for all pixels and do not fully explore the rich information contained in the boundary area. In addition, due to the particularity of ultrasound images (such as weakened boundaries, artifact interference, etc.), existing models often perform poorly in boundary prediction, resulting in uneven and inaccurate boundaries in segmentation results, and may even lead to misinterpretation in clinical diagnosis.

[0004] Traditional segmentation methods also face the problem of serious imbalance in the ratio of target area and background area. Taking ultrasound images as an example, the target area often occupies a very small part of the image, which makes it difficult for the pixel-level average loss function (such as cross entropy loss or Dice loss) to effectively balance the weights between categories during training, thereby limiting the model's sensitivity to the target area, especially the boundary area; at the same time, the high cost of data annotation in the field of medical images and the scarcity of annotated data are also one of the important bottlenecks currently faced. The existence of a large amount of unlabeled data provides opportunities for the development of self-supervised or semi-supervised learning methods, but how to improve model performance in small sample or unlabeled data scenarios is still a difficult problem that needs to be solved in current research.

[0005] In response to the above problems, some studies in recent years have proposed multi-task learning frameworks, hoping to improve boundary segmentation capabilities through shared feature representations. However, these methods still have problems such as high model complexity, high hardware requirements, and lack of universality, and have failed to fully solve the core difficulties of boundary segmentation. Therefore, how to design an efficient boundary enhancement mechanism to improve the performance of ultrasound image segmentation, especially to improve the classification ability of boundary pixels, has become a key research direction and an important challenge in the field of medical image segmentation. Summary of the invention

[0006] The purpose of the present invention is to address the problems of uneven boundaries and insufficient precision in ultrasound image segmentation in the prior art, and propose a boundary-enhanced ultrasound image segmentation method based on a dual decoder structure to improve segmentation performance and diagnostic credibility. The present invention achieves higher-precision ultrasound image segmentation by introducing a boundary feature enhancement mechanism and a multi-scale structural design; in particular, by adopting an edge label enhancement strategy and feature fusion technology, the smoothness of the segmentation result is significantly improved, making the segmentation boundary closer to the smooth boundary manually marked by the doctor, thereby improving the clinical significance and visual credibility of the segmentation result.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A boundary enhanced ultrasound image segmentation method based on a dual decoder structure, characterized in that it comprises the following steps:

[0009] Step 1: construct an ultrasound image dataset and preprocess all ultrasound images separately;

[0010] Step 2: Use edge detection method to extract the target area boundaries of all ultrasound images, generate a binary boundary label matrix, and complete smoothing processing in combination with pooling operation;

[0011] Step 3, construct an ultrasound image segmentation network, including: an encoder and a main decoder (Seg-decoder), a boundary decoder (Bor-decoder), wherein the input image passes through the encoder, the encoder sequentially outputs multi-scale coding features, the multi-scale coding features are respectively input to the main decoder and the boundary decoder through jump connections, and the main decoder and the boundary decoder form an MS-Unet structure with the encoder respectively; after the boundary decoder is decoded step by step, the output layer outputs the boundary area segmentation result, and at the same time, the upsampled feature map obtained by each level of decoding in the boundary decoder is jumped to the corresponding decoding layer of the main decoder, and the main decoder performs feature splicing with the upsampled feature map obtained by its own decoding and then performs the next level of decoding, and the main decoder outputs the target area segmentation result by the output layer after decoding step by step;

[0012] Step 4: define a mixed loss function, complete the training of the ultrasound image segmentation network based on the training set, and obtain an ultrasound image segmentation model;

[0013] Step 5: After the ultrasonic image to be processed undergoes the same preprocessing, it is input into the ultrasonic image segmentation model, and the ultrasonic image segmentation model outputs the target area segmentation result and the boundary area segmentation result.

[0014] Furthermore, in step 1, the specific process of preprocessing is:

[0015] Step 1.1: grayscale the ultrasound image, converting the original RGB three-channel color image into a single-channel grayscale image;

[0016] Step 1.2: Scaling the grayscale processed ultrasound image to unify the resolution;

[0017] Step 1.3: Convert the ultrasound image after the scaling operation into a binary file.

[0018] Furthermore, the specific process of step 2 is:

[0019] Step 2.1: Use edge detection algorithm to extract the boundary of the target area and generate a binary boundary label matrix for each ultrasound image, where the boundary pixel value is 1 and the background pixel value is 0;

[0020] Step 2.2: Smooth the boundary labels: Perform maximum pooling and average pooling operations on the boundary labels in sequence.

[0021] Furthermore, in step 3, each level of the decoding module of the main decoder includes two layers of depth convolution blocks, a ReLU activation function layer, and an upsampling layer connected in sequence.

[0022] Furthermore, in step 3, the structure of each level decoding module of the boundary encoder is similar, and also includes two layers of depth convolution blocks, a ReLU activation function layer, and an upsampling layer connected in sequence.

[0023] Furthermore, in step 4, the hybrid loss function is specifically expressed as:

[0024] L=A×L t (G S ,P S )+B×L t (G BS ,P BS ),

[0025] Among them, L represents the mixed loss function, L t represents Tversky loss, G S With P S Represent the true label and predicted value of the target area, GBS With P BS They represent the true label and predicted value of the boundary area respectively, and A and B represent the loss weights of the target area and the boundary area respectively.

[0026] Furthermore, both A and B have the value 1.

[0027] Furthermore, in step 4, the training process is specifically as follows:

[0028] Perform online data augmentation on training data, including random cropping and flipping;

[0029] Based on the training data after online data enhancement, the Adamax optimizer is used to complete the training of the ultrasound image segmentation network, and the ultrasound image segmentation model is obtained to complete the training.

[0030] Based on the above technical solution, the beneficial effects of the present invention are:

[0031] The present invention provides a boundary-enhanced ultrasound image segmentation method based on a dual decoder structure, which can significantly improve the segmentation performance; the present invention introduces a dual decoder structure and a boundary feature enhancement mechanism, which significantly improves the overall performance of ultrasound image segmentation, especially in the classification ability of boundary pixels, and the improvement proposed by the present invention can better capture boundary details and avoid the segmentation error caused by boundary blur in traditional methods. Specifically:

[0032] 1) The present invention improves the reliability of diagnosis: by optimizing the quality of boundary segmentation, the segmentation results generated by the present invention are smoother and close to the best standard of manual annotation. Such high-quality segmentation results help doctors to more accurately judge the lesion boundaries and improve the accuracy and clinical reliability of diagnosis;

[0033] 2) The present invention can make more effective use of annotated data: the boundary generation method of the present invention can automatically extract boundary information in combination with a large number of annotated ultrasound images, thereby reducing the reliance on high-quality annotated data and providing a solution for medical image segmentation in data-scarce scenarios;

[0034] 3) The present invention has strong scalability and can flexibly adapt to different data sets and model structures;

[0035] 4) The present invention improves the superiority in small sample scenarios: through the optimization of the hybrid loss function, the present invention shows excellent segmentation ability in small sample scenarios, and can still achieve high segmentation accuracy even when the data annotation is limited, which provides important technical support for resource-constrained medical environments;

[0036] 5) The present invention has wide applicability: the present invention is not only applicable to ultrasound image segmentation, but can also be extended to other medical image segmentation tasks, such as boundary enhancement segmentation in CT and MRI, thereby providing technical support for a wider range of clinical application scenarios.

[0037] In summary, the present invention not only improves the segmentation performance and boundary segmentation quality, but also significantly improves the reliability of the diagnosis results. It has important research value and application prospects in the field of automatic medical image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the principle of Canny edge detection method.

[0039] Figure 2 It is a structural schematic diagram of the ultrasound image segmentation network based on the dual decoder structure in the present invention.

[0040] Figure 3 Schematic diagram of the principle of the feature fusion strategy of the ultrasound image segmentation network based on the dual decoder structure in the present invention.

[0041] Figure 4 This is a diagram showing the results of the boundary enhancement ultrasound image segmentation method based on the dual decoder structure in an embodiment of the present invention.

[0042] Figure 5 This is a visualization result diagram of some test samples of the boundary enhancement ultrasound image segmentation method based on the dual decoder structure in an embodiment of the present invention.

[0043] Figure 6 Graphs showing segmentation effects of a boundary-enhanced ultrasound image segmentation method based on a dual-decoder structure in multiple scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the objectives, technical solutions and beneficial effects of the present invention more clearly understood, preferred embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0045] This embodiment provides a boundary-enhanced ultrasound image segmentation method based on a dual decoder structure, which is used to improve the accuracy of ultrasound image segmentation and the ability to capture boundary information. The method specifically includes the following steps:

[0046] Step 1: Preprocess all ultrasound images and divide them into training set, validation set and test set, unify the format and resolution of the images to adapt to the subsequent processing requirements, and convert the data into an efficient storage format to improve reading efficiency;

[0047] The following preprocessing operations are performed on all images in the training set, validation set, and test set:

[0048] Step 1.1: Grayscale all images in the training set, validation set, and test set, converting ultrasound images from original RGB three-channel color images to single-channel grayscale images to reduce computational complexity and focus on the structural features of the image;

[0049] Step 1.2: Scaling the image processed in step 1.1 to unify the resolution of the ultrasound image to 256×256 pixels to adapt to the size requirements of the model input while ensuring resolution consistency;

[0050] Step 1.3: Convert the image processed in step 1.2 into a binary file in .npz format to improve the efficiency of data storage and loading and accelerate the subsequent training and inference process;

[0051] Step 2: Generate additional boundary labels using the training data preprocessed in step 1, automatically extract the boundary of the target area through edge detection technology, and optimize the continuity and authenticity of the boundary labels in combination with pooling operations, so as to provide better boundary annotations for model training and obtain boundary labels; specifically:

[0052] Step 2.1: Use edge detection algorithm to extract the boundary of the target area: Use Canny edge detection algorithm, set the threshold range (for example: 50-150), and control the edge detection sensitivity; generate a binary boundary label matrix for each training image, where the boundary pixel value is 1 and the background pixel value is 0, and save the generated boundary label file;

[0053] Step 2.2: Smooth the boundary label matrix generated in step 2.1: use the maximum pooling operation to even out the width of the boundary labels and improve the smoothness of the boundaries; then use the average pooling operation to generate softer labels to reduce the impact of boundary noise on model training;

[0054] Step 3: Construct an ultrasound image segmentation network (MS-Unet+Bor-decoder) including a main decoder (Seg-decoder) and a boundary decoder (Bor-decoder). Use Bor-decoder to focus on the segmentation of boundary pixels, and provide boundary feature information to Seg-decoder through a feature fusion mechanism to improve the overall segmentation ability of the model. At the same time, introduce a multi-scale structure to enhance the adaptability to complex targets.

[0055] The ultrasound image segmentation network includes: an encoder and a main decoder (Seg-decoder), and a boundary decoder (Bor-decoder), wherein the input image passes through the encoder, and the encoder sequentially outputs multi-scale coding features, which are respectively input to the main decoder and the boundary decoder through jump connections, and the main decoder and the boundary decoder form an MS-Unet structure with the encoder respectively; after the boundary decoder is decoded level by level, the output layer outputs the boundary area segmentation result, and at the same time, the upsampled feature map obtained by each level of decoding in the boundary decoder is jump-connected to the corresponding decoding layer of the main decoder, and the main decoder performs feature splicing with the upsampled feature map obtained by its own decoding and then performs the next level of decoding, and after the main decoder is decoded level by level, the output layer outputs the target area segmentation result;

[0056] Furthermore, in the main decoder, each level of decoding module includes 2 layers of deep convolution blocks and upsampling operations, and finally outputs the segmentation result by gradually restoring the spatial resolution; in the boundary decoder, each level of decoding module also includes 2 layers of deep convolution blocks and upsampling operations, and finally outputs a single-channel boundary segmentation map;

[0057] Step 4: Define a hybrid loss function, train the ultrasound image segmentation network based on the training set, and obtain an ultrasound image segmentation model; during the training phase, guide the model through the hybrid loss function and optimize the segmentation performance of the target area and boundary at the same time; specifically:

[0058] Step 4.1: Define a hybrid loss function based on Tversky loss, and combine boundary loss to optimize the overall segmentation performance. The hybrid loss function is specifically expressed as:

[0059] L=A×L t (G S ,P S )+B×L t (G BS ,P BS ),

[0060] Among them, L represents the mixed loss function, L t represents Tversky loss, G S With P S Represent the true label and predicted value of the target area, G BS With P BS Represent the true label and predicted value of the boundary area respectively, A and B represent the loss weights of the target area and the boundary area respectively;

[0061] Tversky loss is specifically expressed as:

[0062]

[0063] Among them, G represents the true label, P represents the predicted value, and α and β are both weight parameters;

[0064] During the training process, the same Tversky loss is used for the target area (the overall target segmentation task) and the boundary area, which can effectively guide the Bor-decoder of the model to obtain the ability to extract boundary features;

[0065] Step 4.2: Perform online data augmentation on the training data, including random cropping and flipping, to improve the generalization ability of the model;

[0066] Step 4.3: Based on the training data after online data enhancement, the Adamax optimizer (dynamically adjusting the learning rate) is used to complete the training of the ultrasound image segmentation network, and the ultrasound image segmentation model is obtained to complete the training;

[0067] Step 4.4: Evaluate the model on the validation set and test set, measure the target area segmentation performance with the Dice coefficient, and verify the boundary segmentation accuracy with the boundary indicator;

[0068] Step 5: After the ultrasound image to be processed undergoes the same preprocessing as in step 1, it is input into the ultrasound image segmentation model, and the ultrasound image segmentation model outputs the target area segmentation result and the boundary area segmentation result.

[0069] The beneficial effects of the present invention are described in detail below in conjunction with simulation tests.

[0070] 1. Simulation test conditions;

[0071] This embodiment is simulated in a hardware environment where the central processor is an Intel(R) Xeon(R) Silver 4210R CPU@2.40GHz, the graphics card is an NVIDIA GeForce RTX 3090 24GB video memory, and the operating system is Ubuntu 20.04.1LTS. The simulation platform uses PyTorch as the backend computing support tool;

[0072] The datasets used in this embodiment are the public Tn3k and DDTI ultrasound image datasets, where the Tn3k dataset was proposed by Gong et al. in the document "Gong H, Chen G, Wang R, et al. Multi-task learning for thyroid nodule segmentation with thyroid region prior [C] / / 2021 IEEE 18th international symposium on biomedical imaging (ISBI). IEEE, 2021: 257-261.", and contains 3493 ultrasound images from 2421 patients and their segmentation labels; the DDTI dataset comes from the MICCAI 2020 TN-SCUI Challenge, and contains 637 thyroid nodule ultrasound images and their segmentation labels; all simulation tests are evaluated using a standard dataset partitioning strategy to verify the effectiveness and versatility of the present invention.

[0073] 2. Simulation test content;

[0074] The Tn3k dataset used in this embodiment contains 3493 ultrasound images from 2421 patients, including at least one annotated area of ​​thyroid nodules; the dataset is divided into a training set (2303 images), a validation set (576 images) and a test set (641 images) according to official recommendations. The DDTI dataset has a total of 637 images, which are divided into a training set and a test set at a ratio of 80% training and 20% testing after cleaning. During the training process, in order to optimize the model's ability to process target boundaries, the present invention proposes a boundary label generation method, such as Figure 1 As shown in the figure, the Canny edge detection method is used to extract the initial boundary information, and the boundary area is smoothed through the pooling operation to generate boundary labels with consistent thickness and accurate details. This process not only improves the training accuracy of the boundary, but also provides high-quality supervision information for the subsequent boundary decoder (Bor-decoder).

[0075] like Figure 2 As shown in FIG. 1 , the present invention adopts a dual decoder structure, in which the Seg-decoder is responsible for the overall segmentation task of the target area, while the Bor-decoder focuses on the segmentation of the target boundary; and the two realize information sharing and complementation through feature fusion strategy, such as Figure 3As shown, each layer of the Seg-decoder not only receives the skip connection features from the encoder, but also fuses the up-sampling features of the Bor-decoder, thereby enhancing the comprehensive understanding of the target area and boundary information; In addition, it should be noted that: the present invention adopts the existing MS-Unet model as comparative example 1, and at the same time, only for the feature fusion strategy, the present invention also provides comparative examples 2 and 3. The only difference between the comparative examples and the embodiment is that the feature fusion strategy is different; the feature fusion strategy of this embodiment is as follows Figure 3 As shown in (a), the feature fusion strategies of Comparative Example 2 and Comparative Example 3 are as follows: Figure 3 (b) and Figure 3 As shown in (c), a comparative test is performed between this embodiment and comparative examples 1 to 3. The test results are shown in Table 1. On the Tn3k dataset, the performance of this embodiment is the best, and the Dice score reaches 87.46%. Compared with the complex fusion methods of comparative examples 2 and 3, the present invention can retain important information more efficiently while avoiding excessive parameter overhead.

[0076] Table 1

[0077]

[0078] The results of the boundary enhancement ultrasound image segmentation method based on the dual decoder structure in this embodiment are as follows: Figure 4 As shown, from left to right are the input, the true label and the prediction result. It can be seen from the figure that after the Bor-decoder is introduced into the present invention, the segmentation result is closer to the manually marked boundary.

[0079] At the same time, the present invention also compares the embodiment with other existing models. The results are shown in Table 2. It can be seen from the table that whether it is the basic U-Net model or the improved MS-Unet model, the introduction of the boundary enhancement module can significantly improve the segmentation performance; specifically, on the Tn3k dataset, the Dice score of the basic U-Net is 84.71%, which is increased to 85.95% after adding Bor-decoder; the Dice score of MS-Unet is increased from 86.10% to 87.46%.

[0080] Table 2

[0081]

[0082] In addition, the present invention optimizes the design of the weight parameters (A, B) in the hybrid loss function and tests the model performance under different weight combinations. The results are shown in Table 3. When A and B are both set to 1, the model performs best, indicating that maintaining a balanced optimization strategy when segmenting the target area and the boundary area is the key to improving the overall segmentation performance.

[0083] Table 3

[0084]

[0085] In order to intuitively demonstrate the segmentation performance, the present invention performs a visual analysis on some test samples, such as Figure 5 As shown in Figure 2, in the feature maps generated by Seg-decoder and Bor-decoder respectively, the boundary enhancement module can capture more detailed boundary features and effectively improve the contour clarity of the target area. In addition, the present invention also selects complex scene samples containing large targets, small targets and multiple targets for testing, such as Figure 6 The figure shows the segmentation effects of each model in these scenes. It can be seen that the MS-Unet+Bor-decoder model proposed in the present invention performs better than other methods in boundary preservation and target shape capture, can accurately distinguish adjacent targets, and effectively avoid over-segmentation or missed segmentation.

[0086] In summary, the effectiveness of the present invention has been verified in multiple dimensions. It not only performs well in quantitative indicators (such as Dice score), but also shows obvious advantages in visual effects.

[0087] The above description is only a specific implementation mode of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other alternative features that are equivalent or have similar purposes; all the disclosed features, or all the steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A boundary-enhanced ultrasound image segmentation method based on a dual decoder structure, characterized in that: The following steps are involved: Step 1: construct an ultrasound image dataset and preprocess all ultrasound images separately; Step 2: Use edge detection method to extract the target area boundaries of all ultrasound images, generate a binary boundary label matrix, and complete smoothing processing in combination with pooling operation; Step 3, construct an ultrasound image segmentation network, including: an encoder and a main decoder (Seg-decoder), a boundary decoder (Bor-decoder), wherein the input image passes through the encoder, the encoder sequentially outputs multi-scale coding features, the multi-scale coding features are respectively input to the main decoder and the boundary decoder through jump connections, and the main decoder and the boundary decoder form an MS-Unet structure with the encoder respectively; after the boundary decoder is decoded step by step, the output layer outputs the boundary area segmentation result, and at the same time, the upsampled feature map obtained by each level of decoding in the boundary decoder is jumped to the corresponding decoding layer of the main decoder, and the main decoder performs feature splicing with the upsampled feature map obtained by its own decoding and then performs the next level of decoding, and the main decoder outputs the target area segmentation result by the output layer after decoding step by step; Step 4: define a mixed loss function, complete the training of the ultrasound image segmentation network based on the training set, and obtain an ultrasound image segmentation model; Step 5: After the ultrasonic image to be processed undergoes the same preprocessing, it is input into the ultrasonic image segmentation model, and the ultrasonic image segmentation model outputs the target area segmentation result and the boundary area segmentation result.

2. The boundary-enhanced ultrasound image segmentation method based on a dual decoder structure according to claim 1, characterized in that: In step 1, the specific process of preprocessing is: Step 1.1: grayscale the ultrasound image, converting the original RGB three-channel color image into a single-channel grayscale image; Step 1.2: Scaling the grayscale processed ultrasound image to unify the resolution; Step 1.3: Convert the ultrasound image after the scaling operation into a binary file.

3. The boundary-enhanced ultrasound image segmentation method based on a dual decoder structure according to claim 1, characterized in that: The specific process of step 2 is: Step 2.1: Use edge detection algorithm to extract the boundary of the target area and generate a binary boundary label matrix for each ultrasound image, where the boundary pixel value is 1 and the background pixel value is 0; Step 2.2: Smooth the boundary labels: Perform maximum pooling and average pooling operations on the boundary labels in sequence.

4. The boundary-enhanced ultrasound image segmentation method based on a dual decoder structure according to claim 1, characterized in that: In step 3, each level of the decoding module of the main decoder includes a 2-layer deep convolution block (DConv Block-2-layer conv), a ReLU activation function layer, and an upsampling layer (Transposed Convolution) connected in sequence.

5. The boundary-enhanced ultrasound image segmentation method based on a dual decoder structure according to claim 1, characterized in that: In step 3, the decoding module structure of each level of the boundary encoder includes a 2-layer deep convolution block (DConvBlock-2-layer conv), a ReLU activation function layer, and an upsampling layer (Transposed Convolution) connected in sequence.

6. The boundary-enhanced ultrasound image segmentation method based on a dual decoder structure according to claim 1, characterized in that: In step 4, the hybrid loss function is specifically expressed as: L=A×L t (G S ,P S )+B×L t (G BS ,P BS ), Among them, L represents the mixed loss function, L t represents Tversky loss, G S With P S Represent the true label and predicted value of the target area, G BS With P BS They represent the true label and predicted value of the boundary area respectively, and A and B represent the loss weights of the target area and the boundary area respectively.

7. The boundary-enhanced ultrasound image segmentation method based on a dual decoder structure according to claim 6, characterized in that: The values ​​of A and B are both 1.

8. The boundary-enhanced ultrasound image segmentation method based on a dual decoder structure according to claim 1, characterized in that: In step 4, the training process is as follows: Perform online data augmentation on training data, including random cropping and flipping; Based on the training data after online data enhancement, the Adamax optimizer is used to complete the training of the ultrasound image segmentation network, and the ultrasound image segmentation model is obtained to complete the training.

Images