Small-scale ultrasonic image segmentation model based on shape prior constraint

Abstract

A small-scale ultrasound image segmentation model based on shape prior constraints is provided. The method for generating the small-scale ultrasonic image segmentation model based on the shape prior constraint comprises the following steps: S1, collecting ultrasonic data, and determining a segmentation target in the ultrasonic data; s2, image preprocessing is carried out for the problems of high noise, low contrast and edge blurring of the ultrasonic image, and data enhancement is carried out in a deep learning framework; s3, constructing an ultrasonic segmentation evaluation index; s4, a basic U-Net segmentation model is constructed, and the basic U-Net segmentation model is matched with the ultrasonic segmentation data size and the segmentation target category number and serves as a baseline model to carry out follow-up step work; s5, designing a shape prior constraint module, and introducing a variational auto-encoder VAE constraint model to learn shape features of the target structure; s6, model lightweight improvement, wherein the model efficiency is improved while a depth separable convolution and coordinate attention mechanism lightweight model is adopted; and S7, constructing a loss function and performing training to obtain the small-scale ultrasonic image segmentation model based on the shape prior constraint.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, specifically a neural network-based medical ultrasound image structure segmentation model. The method is particularly well-suited for small-scale ultrasound data segmentation tasks and involves ultrasound image segmentation techniques that incorporate shape prior constraints and lightweight network design. Background Art

[0002] Ultrasound imaging plays a key role in disease diagnosis and robotic surgical navigation due to its advantages such as being radiation-free and highly real-time. Accurate segmentation of organ tissues can effectively assist in surgical path planning, reduce operational risks and improve diagnostic consistency.

[0003] However, the current field of medical ultrasound image segmentation still faces many practical challenges, and limited data size is a core issue that needs to be addressed urgently. Due to medical ethics, data sharing restrictions and equipment differences, it is difficult to accumulate effective ultrasound image data, and publicly available high-quality annotated datasets have been scarce for a long time. Restricted by limited annotated data, existing complex networks are prone to overfitting data noise and irrelevant features due to the imbalance between sample size and model capacity, resulting in excellent performance of the model in the training set but insufficient generalization ability. Faced with differences in data distribution between the validation set and the test set, the segmentation performance degrades significantly, resulting in problems such as missed targets and broken boundaries. In addition, the inherent speckle noise, low contrast and blurred edges of ultrasound images further exacerbate the difficulty of segmentation: these defects reduce the visual distinction between the target structure and the background, making it difficult for traditional preprocessing methods to suppress noise while retaining target edge details.

[0004] Therefore, in view of the segmentation difficulties caused by the scarcity of clinical liver ultrasound image annotation data, the easy overfitting of the model and the characteristics of ultrasound images, designing an efficient segmentation method adapted to small-scale data (small-scale refers to training samples with a total number not exceeding 150) has important clinical significance for improving recognition accuracy and surgical navigation safety. The present invention first addresses the problems of ultrasound image noise and low contrast, and adopts special preprocessing techniques (such as edge sharpening and speckle noise suppression) to enhance the visual recognition of the target structure, providing high-quality input for subsequent segmentation; then, by utilizing the stable shape features of structures such as liver blood vessels and breast nodules, a shape prior constraint module is introduced to force the model to focus on the target geometric contour and spatial distribution law, effectively alleviating the problem of insufficient feature learning caused by insufficient samples under small-scale data, and reducing the risk of overfitting to noise and irrelevant details; at the same time, through lightweight network architecture design (such as depthwise separable convolution and coordinate attention mechanism), the number of model parameters and computational complexity are greatly reduced while ensuring the feature extraction capability, so that the model scale matches the small-scale data, and the inference speed is significantly improved to meet the real-time feedback requirements during surgery. This method systematically resolves the contradiction between data, image characteristics and model performance, providing key support for the clinical practical application of ultrasound-assisted recognition technology. Summary of the Invention

[0005] In order to solve the problems that small-scale ultrasound image segmentation models are prone to overfitting training data, have poor segmentation performance and generalization ability, and have inherent defects of ultrasound images themselves, the present invention designs a small-scale ultrasound image segmentation model based on shape prior constraints.

[0006] The specific contents of the present invention are as follows:

[0007] A small-scale ultrasound image segmentation model based on shape prior constraints includes the following steps:

[0008] The first step is to collect ultrasound data, including three types of liver ultrasound data and two types of other parts of the ultrasound data, and determine the segmentation target. The dataset includes three types of liver ultrasound data (BK-500, KL, Philips) and two types of other parts of the ultrasound data (BUSI breast ultrasound [1] , FASCICLE calf muscle ultrasound [2] The segmentation targets of liver ultrasound are the main blood vessels and cysts in the section, the segmentation targets of the BUSI dataset are breast nodules, and the segmentation targets of the FASCICLE dataset are the calf aponeurosis.

[0009] The second step is to preprocess the ultrasound images to address the issues of high noise, low contrast, and blurred edges, and perform data enhancement within the deep learning framework. The preprocessing methods include Sobel sharpening and SRAD (superseded gradient descent). [3] Denoising. This method suppresses noise while preserving edge information. Data enhancement methods include geometric transformations, brightness and contrast adjustments, etc.

[0010] The third step is to construct ultrasound segmentation evaluation indicators, including pixel-level segmentation evaluation indicators and geometric feature evaluation indicators. The former includes classification Dice coefficient (mDice) and intersection-over-union (IOU), and the latter includes average surface distance

[0011] (ASD) and Hausdorff distance (HD).

[0012] The fourth step is to build a basic U-Net segmentation model. This model uses a classic encoder-decoder structure. The encoder activates image features using the ReLU function and performs max-pooling downsampling. The decoder fuses multi-scale features through upsampling and skip connections. Finally, a 1×1 convolution maps the features to the target categories, adapting to the size of ultrasound segmentation data and the number of target categories.

[0013] In the fifth step, a variational autoencoder (VAE) is introduced to compress the pixel-level segmentation and the true mask into a low-dimensional latent space through the encoder, and the Euclidean distance of the latent variables is calculated to construct a low-dimensional manifold loss to constrain the model to learn the shape characteristics of the target structure; the reconstruction loss is constructed through the mean square error loss (MSE) and binary cross entropy loss (BCE) to enhance the VAE shape prior constraint module’s ability to capture and reconstruct shape features.

[0014] In the sixth step, depth-wise separable convolution is used to decompose the conventional convolution into depth-wise convolution and point-wise convolution, reducing the amount of computation and parameters. The coordinate attention module (CoordAtt) is introduced to decompose the global pooling of the spatial dimension into independent 1D operations in the horizontal and vertical directions, explicitly encoding the coordinate information.

[0015] The seventh step is to verify the effect of the method on five small-scale datasets and compare it with Att U-Net. [4] 、U-Net++ [5] 、SEGNet [6] 、DeepLabv3+ [7] 、MSNet [8] 、PolySeg [9] CE-Net

[10] A comparative experiment on segmentation effect and computational efficiency was conducted on seven current mainstream medical image segmentation models.

[0016] Compared with the prior art, the present invention has the following innovations:

[0017] 1. Propose a preprocessing scheme (sharpening, noise reduction, and enhancement) adapted to the characteristics of ultrasound images, effectively improve the low contrast and noise problems, and provide a standardized data processing reference for similar tasks.

[0018] 2. Design a VAE shape prior constraint module that only works during the training phase. This forces the model to learn target shape features to cope with overfitting on small-scale data. This module does not impose any additional computational burden on inference, balancing accuracy and lightweightness.

[0019] 3. Integrate depthwise separable convolution and coordinate attention to enhance spatial positioning capabilities while reducing computational complexity, achieving a balance between efficiency and accuracy under small-scale data.

[0020] The method execution process provided in the embodiments of the present application runs on devices such as personal computers, servers, embedded computing devices, and cloud computing platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present application 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 only some embodiments recorded in this application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0022] Figure 1 This is a comparison chart of the effects before and after ultrasound data preprocessing. The upper picture is before preprocessing, and the lower picture is after preprocessing.

[0023] Figure 2 Figure 2 is the structural diagram of the basic U-Net segmentation model constructed.

[0024] Figure 3 This is the overall structure diagram of the model after introducing the VAE shape prior constraint module.

[0025] Figure 4 This is the depth-wise separable convolution structure diagram.

[0026] Figure 5 This is the structure diagram of the coordinate attention mechanism.

[0027] Figure 6-10 The models are BK-500, KL, Philips, BUSI [1] FASCICLE [2] Segmentation results in five categories of small-scale datasets, where mask represents the true mask position of the segmentation structure, and the remaining eight methods, including our method (Ours), show the visual segmentation effects.

[0028] Figure 11 The figure compares this method with other methods in terms of parameter number (M), computational complexity (GFLOPs), and inference speed (frames / second).

[0029] Explanation of the attached table

[0030] Table 1 shows the methods and related parameters used in the data enhancement operation.

[0031] Table 2 shows the experimental environment configuration.

[0032] Table 3 shows the comparison of the segmentation effects of the eight segmentation methods in three types of liver ultrasound images (BK-500, KL, and Philips), following the evaluation indicators proposed in step three.

[0033] Table 4 The performance of eight segmentation methods on two types of public ultrasound images (BUSI [1] FASCICLE [2] ) in the comparison of segmentation effects. DETAILED DESCRIPTION

[0034] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0035] 1. Collect ultrasound data, including three types of liver ultrasound data and two types of other parts of the ultrasound data, and determine the segmentation target. The dataset includes three types of liver ultrasound data (BK-500, KL, Philips) and two types of other parts of the ultrasound data (BUSI breast ultrasound [1] , FASCICLE calf muscle ultrasound [2] After removing information such as the scanning software UI from the original data, the image and the mask containing the pixel-level annotations of the real structure position are saved as a 512*512 pixel size, 24 pixel bit depth, and the image format is PNG.

[0036] 2. Use the Sobel operator to sharpen the ultrasound image. The specific implementation method is to use a 3×3 horizontal detection kernel G X With the vertical detection kernel G Y By comparing with the original image I origianl The convolution operation (*)) obtains the horizontal gradient amplitude G of the corresponding pixel point (x, y) x (x,y) and vertical gradient magnitude G y (x,y), calculated as:

[0037]

[0038] Then calculate the gradient amplitude of the corresponding pixel And then used to sharpen the image:

[0039] I sharpened =I origianl +λ·G

[0040] I sharpened represents the sharpened image, λ is the sharpening strength coefficient, and this method has been verified to take a value of 2.0.

[0041] By SRAD [3] The method performs denoising on the noise introduced by sharpening and the original noise of the ultrasound image:

[0042]

[0043] in:

[0044]

[0045]

[0046] q is a normalized gradient based on speckle noise, which is used to characterize the significance of local features. A small constant ∈ is added to ensure that the denominator is not zero. q0 is adaptively adjusted with the number of iterations t to control the balance between smoothness and edge preservation. Var[I(t)] and Mean[I(t)] 2 Is the variance and mean of the current image. In the initial stage, input ultrasound image input image I 0 , set the iteration step size Δt, the total number of iterations N and the initial q0, and calculate the gradient of the current image respectively Laplace operator Normalized gradient q and diffusion coefficient c(q), where Δt and N are adjustable parameters, are updated iteratively by setting appropriate parameter values:

[0047]

[0048] Where n represents the number of iterations and div represents the divergence calculation. In addition, the data enhancement methods are shown in Table 1. The overall model structure is as follows: by replacing the original conventional 3×3 convolution with depthwise separable convolution and introducing the coordinate attention mechanism to achieve lightweight improvement, the variational autoencoder (VAE) shape prior constraint module is introduced into the output part of the lightweight improved U-Net model, as shown in Figure 3 As shown. The next process is to first build and adjust the basic U-Net model according to the method described in step 3, and then design and introduce the VAE shape prior constraint module according to the method in step 4 to form the overall model architecture of the small-scale ultrasound image segmentation model based on shape prior constraint. During the training process of the overall model, the pixel-level segmentation prediction and the real mask output by the model need to be input into the VAE shape prior constraint module, and the low-dimensional manifold loss of the two is constructed. Used to update the parameters of the VAE shape prior constraint module encoder and U-Net model to construct image reconstruction loss and It is used to optimize the VAE shape prior constraint module's ability to extract segmentation target shape information. At the same time, the structure of the basic U-Net model is lightweighted through step 5. The specific improvement method is to replace the original ordinary convolution of the basic U-Net model with Figure 4The depthwise separable convolutional network shown in the figure embeds the coordinate attention mechanism into the feature map obtained after the convolution operation, so that the lightweight improved U-Net model matches the sample size and improves the computational efficiency and inference speed. The overall model needs to introduce the VAE shape prior constraint module and calculate the related loss during the training phase. With the help of the VAE shape prior constraint module to calculate the loss and back-propagate the updated parameters, a more accurate prediction of the image structure is made. Therefore, the output of the lightweight improved U-Net model is still only pixel-level segmentation. In the verification phase, only the lightweight improved U-Net model ( Figure 2 Structure+ Figure 4 、 Figure 5 During the inference phase, the output of the overall model is the pixel-level segmentation output of the lightweight and improved U-Net model. The VAE shape prior constraint module does not need to participate in the inference process. This feature saves parameters and computation for the overall model, which is consistent with the original design intention of this method, which is lightweight.

[0049] 3. Build a basic U-Net model with the following structure: Figure 2 Regularization techniques, such as batch normalization and dropout, were added, with dropout probabilities set to p = 0.2 for the spatial dimension and p = 0.15 for the channel dimension, respectively, to enhance the robustness of the basic U-Net model in medical segmentation tasks. LeakyReLU activation function and AdamW optimizer were used to improve training stability and convergence speed.

[0050] 4. Design and integrate the VAE shape prior constraint module at the output of the basic U-Net segmentation model. Its structure is as follows Figure 3 As shown. The VAE shape prior constraint module is composed of an encoder that maps the input to the parameters of the latent variable distribution, obtains the latent variables through reparameterized sampling, and then restores it to the output through the decoder. The specific calculation method is: the input of the VAE shape prior constraint module is a pixel-level segmentation map and the corresponding unique hot label (512×512 size), which is compressed to 16×16×512 features through the encoder's four-level convolution downsampling (channel 64→128→256→512, step size 2). After flattening, a 128-dimensional mean and variance parameters are generated through a fully connected layer. The differentiable latent variables are obtained by reparameterization with normal distribution noise to achieve a low-dimensional manifold representation of high-dimensional shape information. The encoder obtains the mean μ and variance σ compressed to 128 dimensions through two fully connected layers. 2 , and then calculate the randomly sampled latent variables representing the shape features:

[0051] z=μ+ε⊙σ

[0052] The two differentiable parameters μ and σ output by the encoder can be reparameterized by introducing noise ε. The noise ε conforms to the multidimensional normal distribution N(0,I). While introducing randomness, it makes the latent variable z differentiable, thereby enabling gradient descent. After obtaining the latent variable z, the VAE shape prior constraint module will gradually reduce the number of channels and restore the spatial dimension through deconvolution upsampling in the decoding part. This method denotes the low-dimensional randomly sampled latent variables obtained by pixel-level segmentation and the true mask calculation as Figure 3 The z shown a With z b , used for distinguishing expressions in the study, and the same applies to other related parameters. The module is optimized through dual loss constraints: one is to calculate the Euclidean distance (low-dimensional manifold loss) between the segmentation map and the true mask latent variable, guiding the overall model to focus on shape differences:

[0053]

[0054] The second is to optimize the parameters of the VAE shape prior constraint module through reconstruction loss (MSE for segmentation images and BCE for masks, both containing KL divergence terms) to improve shape capture and reconstruction capabilities:

[0055]

[0056] where x i represents the pixel values of the input encoder image (whether segmented or masked), Represents the corresponding pixel value output by the image reconstruction decoder. Both types of losses contain the same KL divergence loss term It is used to represent the difference between the latent variable distribution and the preset multidimensional normal distribution N(0,I), preventing the encoder from generating meaningless "collapsed" distributions, ensuring the continuity and smoothness of the latent space structure, and maintaining good shape representation capabilities:

[0057]

[0058] D represents the dimension of the generated latent variable (128 dimensions). The β coefficient added before the KL divergence loss is used to balance the prior distribution and reconstruction quality. This method has been verified to take β = 1. The reconstruction loss updates the encoder and decoder parameters of the VAE shape prior constraint module during training, so that the pre-trained module can further adapt to the specific task requirements and achieve directional optimization of the module parameters towards the task goal. Different from the low-dimensional manifold loss This optimization process is independent of the U-Net segmentation model. To avoid conflicts in optimization objectives caused by multi-task gradient propagation, the pixel-level segmentation and reconstruction loss needs to be truncated. The gradient path of the U-Net segmentation model ensures that the VAE shape prior constraint module can correctly represent the shape distribution information of the image structure.

[0059] 5. Use the depth-wise separable convolution method to improve the original conventional 3×3 convolution operation. Figure 4 As shown in the figure, n×n channel-wise convolution is performed independently on the channel of each input feature map, so that each channel feature map only interacts with the corresponding convolution kernel, avoiding redundant calculations between channels and effectively suppressing the spread of noise between channels. Subsequently, the independent channel features are spliced after pointwise convolution with a 1×1 convolution kernel to achieve inter-channel feature fusion and dimensionality adjustment. In addition, the feature map obtained after the depthwise separable convolution is embedded in the coordinate attention module, whose structure is as follows: Figure 5 As shown: For a feature map x with a size of H (height) × W (width) c , in the horizontal and vertical directions, global average pooling with sizes of H×1 and 1×W is used to obtain horizontal compression features Compress features perpendicular to the direction The feature maps of the two directions are concatenated and processed by 1×1 convolution And nonlinear activation function σ (such as Sigmoid) generates intermediate features F. The intermediate features are then decomposed into horizontal vectors F H and perpendicular vector F W After adjusting the number of channels and activating them through 1×1 convolution, the horizontal and vertical attention maps are generated respectively. att , weight the original feature map to generate a weighted feature map. The specific calculation method is:

[0060]

[0061] 6. Construct the loss function and train the whole model. The model training process is as follows: the lightweight improved U-Net model predicts the output pixel-level segmentation, and the difference between the pixel-level segmentation and the true mask is calculated to obtain the segmentation loss. The specific calculation method included is:

[0062] First, the basic multi-classification Dice loss constraint segmentation task is adopted, which is expressed as:

[0063]

[0064] Where K represents the number of segmentation categories, and Dice represents the Dice loss of the segmentation category, which is calculated as follows:

[0065]

[0066] Where A and B represent the pixel-level segmentation and the corresponding area of the real mask respectively, and |·| represents the number of elements in the set (the number of pixels in image segmentation). In addition, the boundary loss based on distance transformation is introduced.

[0067] Based on Loss). The description of the feature shape by the loss function and the constraints provided by the VAE shape prior constraint module are implemented as explicit-implicit joint constraints to ensure the overall rationality of the segmented shape and improve the accuracy of edge details:

[0068]

[0069] where p i Represents the predicted value corresponding to a certain point in pixel-level segmentation, d i Represents the closest distance from the point to the boundary of the true mask segmentation. The pixel-level segmentation and the true mask are further input into the VAE shape prior constraint module, and the corresponding low-dimensional random sampling latent variable z is generated through the module encoder part. a 、z a Thus, the low-dimensional manifold loss that can represent the difference in the segmented shapes of the two is calculated Used to update the parameters of the VAE shape prior constraint module encoder and the lightweight improved U-Net model. In addition, in order to ensure the VAE shape prior constraint module's ability to extract shape features, it is necessary to continuously optimize the encoder-decoder parameters of the VAE shape prior constraint module, and obtain the loss by the difference between pixel-level segmentation and the real mask before and after reconstruction by the VAE shape prior constraint module. and The above three losses constitute the loss of the VAE shape prior constraint module Among them L LDM During the calculation process, the gradient path of the lightweight improved U-Net model is retained to update the U-Net parameters and the VAE shape prior constraint module encoder parameters, while truncating With the gradient path of the lightweight improved U-Net model, the reconstruction loss is only used to adjust the parameters of the pre-trained VAE shape prior constraint module itself. In summary, the total loss of the overall model is Expressed as:

[0070]

[0071] 7. Configure the experimental environment according to the conditions shown in Table 2, use AdamW as the overall model optimizer, and use weight decay to alleviate overfitting. Set the initial learning rate to 10 -6 , the weight decay coefficient is 10 -4. The optimizer momentum coefficient β1 = 0.85, β2 = 0.999, warm_up_ratio = 0.1. Set batch_size = 4. The other hyperparameters follow the settings when the overall model is built. The segmentation effect of the method is jointly evaluated by the classification Dice coefficient (mDice) and intersection over union (IOU), the latter including the average surface distance (ASD) and Hausdorff distance (HD), and compared with Att U-Net. [4] 、U-Net++ [5] 、SEGNet [6] 、DeepLabv3+ [7] 、MSNet [8] 、PolySeg [9] CE-Net

[10] Seven mainstream medical image segmentation models were compared experimentally. Quantitative comparisons of the methods' performance are shown in Tables 3 and 4. The liver ultrasound data segmentation results shown in Table 3 demonstrate that our method achieves the best performance on the KL ultrasound dataset in terms of the mDice and mIOU metrics, which focus on pixel-wise segmentation accuracy, outperforming the runner-up model by 0.41% and 2.06%, respectively. Although our method performs lower than the MSNet and CE-Net methods on the BK-500 dataset and lower than the PolySeg method on the Philips dataset, the combined results across all three datasets still place us at the forefront of all methods. In terms of the HD95 and ASD metrics, which focus on geometric segmentation contours, our method demonstrates significant advantages over other methods across all three datasets. The HD95 metric achieves the best performance among all methods on the KL and Philips datasets, demonstrating our method's clear advantage in small-scale liver ultrasound image segmentation. The public ultrasound data segmentation results shown in Table 4 demonstrate that our method also achieves good results on public data segmentation tasks. Among them, the mDice indicator ranked third in the BUSI dataset and second in the FASCICLE dataset; the mIOU indicator ranked second in both datasets. Although the HD95 and ASD indicators also reached the cutting-edge level among all methods in the two datasets, they did not show a more obvious advantage like the liver ultrasound dataset. This may be because the distribution of segmentation targets in the public dataset is more diverse and random, which has a certain impact on the judgment of boundary contours and makes it more dependent on the segmentation results of specific pixels. The above experimental results prove that the method in this paper has the value of being promoted to other types of small-scale ultrasound data. Refer to the corresponding visualization effect of the experiment Figure 6-10 The mask below the image represents the real mask. The segmentation effect diagrams of eight methods including this method are arranged in order. The comparison of this method with other methods in terms of parameter quantity, computational complexity and inference speed is shown in the table below. Figure 11 . The results show that the method in this paper performs well in terms of lightweightness and has significant advantages over other methods. The number of parameters is the lowest among all methods, only 17.9% of the number of parameters of the second-ranked CE-Net; the amount of computation and inference speed are second only to DeepLabv3+, of which the computational amount is equivalent to 70.0% of the third-ranked CE-Net, while the inference speed is 72.0% higher than that. It is worth noting that although MSNet has better segmentation effects on multiple categories of data sets than this method, its number of parameters and computational costs are 6.97 times and 6.68 times that of this method, respectively. Overall, this method has achieved a good balance between segmentation accuracy and computational efficiency.

[0072] Although preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of this application. Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if such changes and modifications fall within the scope of the claims of this application and their equivalents, then this application is intended to include such changes and modifications.

[0073] Appendix

[0074] Table 1 Data enhancement methods and parameters

[0075]

[0076] Table 2 Experimental environment configuration

[0077]

[0078] Table 3 Experimental results of liver dataset

[0079]

[0080] Table 4 Experimental results of public dataset

[0081]

[0082] References

[0083] [1]Al-Dhabyani W,Gomaa M,Khaled H,et al.Dataset of breast ultrasound images[J].Data in brief,2020,28:104863.

[0084] [2]Michard H,Luvison B,Pham Q C,et al.AW-Net:Automatic musclestructure analysis on B-mode ultrasound images for injury prevention[C] / / Proceedings of the 12th ACM International Conference on Bioinformatics,

[0085] Computational Biology,and Health Informatics.2021:1-9.

[0086] [3]Michailovich O V,Tannenbaum A.Despeckling of medical ultrasoundimages[J].ieee transactions on ultrasonics,ferroelectrics,and frequencycontrol,2006,53(1):64-78.

[0087] [4]Oktay O,Schlemper J,Folgoc L L,et al.Attention u-net:Learningwhere to look for the pancreas[J].arXiv preprint arXiv:1804.03999,2018.

[0088] [5]Chen L C,Zhu Y,Papandreou G,et al.Encoder-decoder with atrousseparable convolution for semanticimage segmentation[C] / / Proceedings of theEuropean conference on computer vision(ECCV).2018:801-818.

[0089] [6]Badrinarayanan V,KendallA,Cipolla R.Segnet:A deepconvolutionalencoder-decoder architecture forimage segmentation[J].IEEEtransactions on pattern analysis and machine intelligence,2017,39(12):2481-2495. [7]Chen L C,Zhu Y,Papandreou G,et al.Encoder-decoder with atrousseparable convolution for semanticimage segmentation[C] / / Proceedings of theEuropean conference on computer vision(ECCV).2018:801-818.

[0090] [8]Zhao X,Zhang L,Lu H.Automatic polyp segmentation via multi-scalesubtraction network[C] / / Medical Image Computing and Computer-AssistedIntervention(MICCAI).2021:120-130.

[0091] [9]Zhong J,Wang W,Wu H,et al.PolypSeg:An efficient context-awarenetwork for polyp segmentation from colonoscopy videos[C] / / Medical ImageComputing and Computer Assisted Intervention-MICCAI 2020.Springer,2020:285-294.

[0092] Gu Z,Cheng J,Fu H,et al.CE-Net:Context encoder network for 2d medicalimage segmentation[J].IEEE Transactions on Medical Imaging,2019,38(10):2281-2292.

Claims

1. A method for generating a small-scale ultrasound image segmentation model based on shape prior constraints, characterized in that: The following steps are involved: S1: Collect ultrasound data, including three types of liver ultrasound data, breast ultrasound data, and calf muscle ultrasound data, and determine the segmentation target in the ultrasound data; S2: Image preprocessing and data enhancement within a deep learning framework to address high noise, low contrast, and blurred edges in ultrasound images; S3: Constructing ultrasound segmentation evaluation indicators, including pixel-level segmentation evaluation indicators and geometric feature evaluation indicators; S4: Build a basic U-Net segmentation model, adapt it to the size of ultrasound segmentation data and the number of segmentation target categories, and use it as a baseline model to carry out subsequent steps; S5: Design a shape prior constraint module and introduce a variational autoencoder (VAE) constraint model to learn the shape characteristics of the target structure; S6: Model lightweighting improvement, using depthwise separable convolution and coordinate attention mechanism to lightweight the model while improving model efficiency; S7: Construct a loss function and perform training to obtain the small-scale ultrasound image segmentation model based on shape prior constraints.

2. The method according to claim 1, characterized in that The ultrasound data described in step S1 include three types of liver ultrasound data: BK-500, KL, and Philips, as well as BUSI breast ultrasound data and FASCICLE calf muscle ultrasound data. The segmentation targets of the liver ultrasound data are the main blood vessels and cysts in the cross section, the segmentation targets of the BUSI breast ultrasound data set are breast nodules, and the segmentation targets of the FASCICLE calf muscle data set are the calf aponeurosis.

3. The method according to claim 1, characterized in that The ultrasound image preprocessing method described in step S2 includes using the Sobel operator for sharpening to improve the clarity of low-contrast edges in order to address the high noise, low contrast, and blurred edge problems of ultrasound images; then using the speckle noise suppression anisotropic diffusion (SRAD) algorithm for noise reduction to suppress noise while retaining edge information; data enhancement methods include geometric transformation, brightness adjustment, contrast adjustment, and elastic deformation to enrich the diversity of training samples and reduce the risk of overfitting.

4. The method according to claim 1, wherein The segmentation evaluation indicators described in step S3 include classification Dice coefficient mDice, intersection-over-union ratio IOU, average surface distance ASD and Hausdorff distance HD.

5. The method according to claim 1, wherein The basic U-Net segmentation model described in step S4 adopts a classic encoder-decoder structure. The encoder activates image features through the ReLU function and performs maximum pooling downsampling. The decoder fuses multi-scale features through upsampling and skip connections. Finally, the features are mapped to the target category through 1×1 convolution, adapting to the size of ultrasound segmentation data and the number of segmentation target categories.

6. The method according to claim 1, characterized in that The shape prior constraint module described in step S5 introduces a variational autoencoder (VAE) to compress the pixel-level segmentation and the true mask into a low-dimensional latent space through the encoder, calculates the Euclidean distance of the latent variables to construct a low-dimensional manifold loss, and constrains the model to learn the shape features of the target structure; the reconstruction loss is constructed by the mean square error loss (MSE) and the binary cross entropy loss (BCE), thereby improving the variational autoencoder (VAE) module's ability to capture and reconstruct shape features. The variational autoencoder (VAE) module only participates in network optimization during the training phase and does not affect the model's lightweightness during the inference phase.

7. The method according to claim 1, characterized in that In step S6, depthwise separable convolution is used to decompose the conventional 3×3 convolution into depthwise convolution and point-by-point convolution, reducing the amount of computation and parameters. The coordinate attention module CoordAtt is introduced to decompose the global pooling of the spatial dimension into independent 1D operations in the horizontal and vertical directions, explicitly encoding coordinate information, enhancing the model's ability to focus on the features of the segmented target, and improving model efficiency while making it lightweight.

8. The method according to claim 1, characterized in that Step S7 consists of multi-classification Dice loss, boundary loss based on distance transformation, and VAE module loss.

9. The method according to claim 1, characterized in that After step S7, the method further includes: Removing the variational autoencoder (VAE) constraint model from the small-scale ultrasound image segmentation model based on shape prior constraints; During the inference process using the small-scale ultrasound image segmentation model based on shape prior constraints, the ultrasound data to be identified is provided to the small-scale ultrasound image segmentation model based on shape prior constraints with the variational autoencoder VAE constraint model removed to obtain a pixel-level segmentation map of the ultrasound data to be identified.

10. An information processing device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images