Ultrasonic left ventricular segmentation network model training method, system, equipment and medium

Abstract

The invention provides an ultrasonic left ventricular segmentation network model training method, system and device and a medium. The ultrasonic left ventricular segmentation network model training method comprises the steps of obtaining a medical image; screening and marking the medical images to obtain a medical image data set; training an initial deep learning model by using the medical image data set to obtain a segmentation network model; wherein the initial deep learning model is constructed by using a deep convolution module, an integrated coding module, an integrated decoding module and a bottleneck block. According to the ultrasonic left ventricle segmentation network model training method, advantages of various modules can be fused, the performance of the segmentation network model is improved, and the accuracy of image segmentation is improved.

Description

Technical Field

[0001] This application belongs to the field of image segmentation technology, and relates to a training method for an ultrasound left ventricular segmentation network model, and particularly to a training method, system, device and medium for an ultrasound left ventricular segmentation network model. Background Technology

[0002] Left ventricular segmentation is a crucial step in the diagnosis and assessment of cardiovascular diseases, providing key evidence for quantitative analysis of cardiac function and lesion detection. Echocardiography, due to its non-invasive, real-time, and repeatable characteristics, has become the primary imaging method for left ventricular segmentation. However, ultrasound images suffer from noise interference, low tissue contrast, and significant dynamic changes in left ventricular morphology during the cardiac cycle, making manual segmentation not only time-consuming and labor-intensive but also reliant on physician experience, leading to highly subjective and inconsistent results. The development of computer-aided diagnostic (CAD) systems has provided an effective solution for automated left ventricular segmentation, with medical image segmentation being a core step in CAD systems, directly impacting the accuracy of subsequent disease analysis and diagnosis. In recent years, deep convolutional neural networks (CNNs), with their powerful nonlinear feature learning capabilities, have been widely applied in left ventricular ultrasound image segmentation tasks. CNNs can automatically extract deep semantic features from images, effectively capturing the complex morphological features of the left ventricle, and have made significant progress in improving segmentation accuracy.

[0003] However, current segmentation network models are still prone to problems such as performance instability, inaccurate segmentation boundaries, or incomplete segmentation when processing left ventricular ultrasound images, which affects segmentation efficiency and accuracy and makes it difficult to meet the needs of clinical applications. Summary of the Invention

[0004] The purpose of this application is to provide a training method, system, device and medium for an ultrasound left ventricular segmentation network model, in order to solve the problem that existing segmentation network models are prone to low performance or insufficient segmentation.

[0005] In a first aspect, this application provides a method for training an ultrasound left ventricular segmentation network model, the method comprising: acquiring medical images; filtering and labeling the medical images to obtain a medical image dataset; and training an initial deep learning model using the medical image dataset to obtain a segmentation network model; wherein the initial deep learning model is constructed using a deep convolutional module, an ensemble encoding module, an ensemble decoding module, and a bottleneck block.

[0006] In this application, an initial deep learning model is constructed using a deep convolutional module, an ensemble coding module, an ensemble decoding module, and a bottleneck block. This initial deep learning model is then trained using a medical image dataset to obtain a segmentation network model. This training method for the ultrasound left ventricular segmentation network model integrates the advantages of multiple modules, improving the performance of the obtained segmentation network model and increasing the accuracy and precision of image segmentation for medical images.

[0007] In one implementation of the first aspect, filtering and labeling the medical images to obtain a medical image dataset includes: filtering the medical images to obtain filtered medical images; and performing lesion region labeling and key structure labeling processing on the filtered medical images to obtain the medical image dataset.

[0008] In one implementation of the first aspect, the deep convolution module includes a first deep convolution module and a second deep convolution module. Training an initial deep learning model using the medical image dataset includes: extracting features from the medical image dataset using the first deep convolution module to obtain a shallow feature map; processing the shallow feature map using the ensemble encoding module to obtain a deep feature map; processing the deep feature map using the bottleneck block to obtain a processed feature map; decoding the processed feature map using the ensemble decoding module to obtain a decoded map; and processing the decoded map using the second deep convolution module to obtain the processing result of the medical image dataset.

[0009] In one implementation of the first aspect, the integrated coding module includes a coding block and a Mamba module. Processing the shallow feature map using the integrated coding module to obtain a deep feature map includes: processing the shallow feature map using the coding block to obtain a processed shallow feature map; and processing the processed shallow feature map using the Mamba module to obtain the deep feature map.

[0010] In one implementation of the first aspect, processing the shallow feature map using the integrated coding module to obtain the deep feature map further includes: pooling the output of the Mamba module using a max pooling module to obtain the deep feature map.

[0011] In one implementation of the first aspect, the Mamba module includes a visual state space module, an adjustment factor, a normalization layer, and a residual visual module. Processing the processed shallow feature map using the Mamba module to obtain the deep feature map includes: processing the shallow feature map through the normalization layer and the visual state space module, and convolving it with the processing result from the adjustment factor and the residual visual module to obtain a convolution result; and processing the convolution result using another normalization layer to obtain the deep feature map.

[0012] In one implementation of the first aspect, the training method for the ultrasound left ventricular segmentation network model further includes: using a residual vision module to make skip connections between the depth convolution modules, between the integrated coding module and the integrated decoding module.

[0013] Secondly, this application provides an ultrasound left ventricular segmentation network model training system, which includes: a data acquisition module for acquiring medical images; a data processing module for filtering and labeling the medical images to obtain a medical image dataset; and a model training module for training an initial deep learning model using the medical image dataset to obtain a segmentation network model; wherein the initial deep learning model is constructed using a deep convolution module, an ensemble encoding module, an ensemble decoding module, and a bottleneck block.

[0014] Thirdly, this application provides an electronic device comprising: a memory for storing a computer program; and a processor for executing the computer program stored in the memory to cause the electronic device to perform the ultrasound left ventricular segmentation network model training method as described in any one of the first aspects.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the ultrasound left ventricular segmentation network model training method described in any one of the first aspects. Attached Figure Description

[0016] Figure 1A The diagram shown illustrates an application scenario of the ultrasound left ventricular segmentation network model training method described in this application.

[0017] Figure 1B This diagram illustrates the structure of the mid-cloud interaction scenario in these implementation methods.

[0018] Figure 2 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application embodiment.

[0019] Figure 3The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application embodiment.

[0020] Figure 4 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application embodiment.

[0021] Figure 5 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application embodiment.

[0022] Figure 6 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application embodiment.

[0023] Figure 7 The diagram shown is a schematic representation of the segmentation network model described in an embodiment of this application.

[0024] Figure 8 The diagram shown is a structural schematic of the ultrasound left ventricular segmentation network model training system described in this application embodiment.

[0025] Figure 9 The diagram shown is a structural schematic of the electronic device described in an embodiment of this application. Detailed Implementation

[0026] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0027] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0028] Methods based on deep convolutional neural networks (CNNs) often acquire a large receptive field and integrate global contextual information by stacking convolutional kernels, which has the following limitations: First, the deep stacking of convolutional operations increases network parameters, leading to low computational efficiency and difficulty in network optimization. Second, the local operation of convolutions can cause over- or under-segmentation when segmenting lesions in the left ventricle on echocardiography. Transformer-based segmentation models directly model long-range dependencies, compensating for the shortcomings of CNNs in handling long-range dependencies and thus improving network performance. However, these methods require significant computational resources for training, especially when dealing with large datasets and long sequences, which limits their application in resource-constrained devices or real-time applications. Furthermore, due to the sequential nature of the model, the training process is difficult to parallelize effectively, affecting training speed. In addition, the parameter efficiency of self-attention mechanisms is relatively low, requiring more computational resources and storage space as the length of the input sequence increases. Overall, although Transformer-based methods perform well in segmentation tasks, their limitations, such as high computational requirements, difficulty in parallelization, poor interpretability, hyperparameter sensitivity, and input length limitations, should be carefully considered in practical applications.

[0029] At least to address the above-mentioned problems, embodiments of this application provide a method for training an ultrasound left ventricular segmentation network model. The method includes: acquiring medical images; filtering and labeling the medical images to obtain a medical image dataset; and training an initial deep learning model using the medical image dataset to obtain a segmentation network model. The initial deep learning model is constructed using a deep convolutional module, an ensemble encoding module, an ensemble decoding module, and a bottleneck block.

[0030] In this embodiment, an initial deep learning model is constructed using a deep convolutional module, an ensemble coding module, an ensemble decoding module, and a bottleneck block. This initial deep learning model is then trained using a medical image dataset to obtain a segmentation network model. This method for training an ultrasound left ventricular segmentation network model integrates the advantages of multiple modules, improving the performance of the obtained segmentation network model and increasing the accuracy and precision of image segmentation for medical images.

[0031] Figure 1A This diagram illustrates an application scenario of the ultrasound left ventricular segmentation network model training method described in this application. The medical image segmentation device 1 can be used to implement the ultrasound left ventricular segmentation network model training method provided in this application embodiment, but the application scenarios of the ultrasound left ventricular segmentation network model training method provided in this application embodiment are not limited to... Figure 1A The medical image segmentation device 1 shown is as follows. Figure 1AAs shown, the medical image segmentation device 1 includes an image acquisition device 11, a local processor 12, and a display terminal 13. The ultrasound left ventricular segmentation network model training method provided in this embodiment can be applied to the local processor 12.

[0032] in, Figure 1A The local processor 12 can be a single local processor, a cluster of multiple local processors, or a cloud computing center, etc., and is not specifically limited here. Although Figure 1A Only one image acquisition device 11, one local processor 12, and one display terminal 13 are shown, but it should be understood that... Figure 1A The examples in this paper are only for understanding this solution. The specific number of local processors 12 and display terminals 13 should be flexibly determined based on the actual situation.

[0033] In some other implementations, the medical image segmentation device 1 may not include a display terminal 13, but only a local processor 12 with display function and an image acquisition device 11. The ultrasound left ventricular segmentation network model training method provided in this application embodiment can be applied to the local processor 12. The local processor 12 with display function may include tablet computers, PDAs, mobile phones, personal computers (PCs), and ultrasound detection equipment, etc., and is not limited here.

[0034] In some other implementations, the ultrasound left ventricular segmentation network model training method described in this application can be applied to edge-cloud interaction scenarios. Figure 1B This diagram illustrates the structure of the endpoint-cloud interaction scenario in these implementation methods. For example... Figure 1B As shown, the terminal-cloud interaction system 2 includes a terminal 20 and a cloud server 21. The terminal 20 and the cloud server 21 can communicate with each other, and the communication method is not limited to wired or wireless.

[0035] The terminal 20 can be mobile or fixed. For example, it can be a wireless terminal or a wired terminal. A wireless terminal can refer to a device with wireless transceiver capabilities, which can be deployed indoors, outdoors, and in industrial workshops. The terminal 20 can be an ultrasonic testing instrument, a mobile phone, a tablet computer, a laptop computer, etc., and is not limited thereto. The cloud server 21 can include one or more servers, or one or more processing nodes, or one or more virtual machines running on the server. The cloud server 21 can also be referred to as a server cluster, management platform, data processing center, etc., and is not limited thereto in this embodiment.

[0036] The technical solutions in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0037] The following embodiments of this application provide a method for training an ultrasound left ventricular segmentation network model, which, for example, can be achieved through… Figure 1A The local processor 12 shown or Figure 1B The cloud server 21 shown is used to implement this. Figure 2 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application. Figure 2 As shown, the training method for the ultrasound left ventricular segmentation network model includes steps S11 to S13.

[0038] Step S11: Acquire medical images. Optionally, the medical images are ultrasound medical images from different time periods and under different conditions.

[0039] Step S12 involves filtering and labeling the medical images to obtain a medical image dataset. Specifically, the medical images are filtered and segmented using masking to obtain the medical image dataset.

[0040] Step S13: Train the initial deep learning model using the medical image dataset to obtain a segmentation network model. The initial deep learning model is constructed using a deep convolutional module, an ensemble encoder module, an ensemble decoder module, and a bottleneck block.

[0041] In some possible implementations, the initial deep learning model's deep convolutional module includes a first deep convolutional module and a second deep convolutional module. After acquiring medical images, the medical images are filtered and labeled to obtain a medical image dataset. The medical image dataset is then sequentially passed through the first deep convolutional module, ensemble encoding module, bottleneck block, ensemble decoding module, and second deep convolutional module of the initial deep learning model to obtain the segmentation result of the medical image dataset.

[0042] In other possible implementations, the designed segmentation network model is systematically trained based on a constructed left ventricular ultrasound image annotation dataset. To improve model performance, a composite loss function combining binary cross-entropy (BCE), DICE coefficient, and Focal Loss is used. The formula for calculating the loss function is as follows: , , , in, Indicates the true label, This indicates the model output. This indicates the number of groups of objects predicted by the model. This represents the predicted probability of the model. It is an adjustable factor used to adjust the degree of influence of the modulation factor.

[0043] The overall loss function is expressed as: .

[0044] In this embodiment, an initial deep learning model is constructed using a deep convolutional module, an ensemble coding module, an ensemble decoding module, and a bottleneck block. This initial deep learning model is then trained using a medical image dataset to obtain a segmentation network model. This method for training an ultrasound left ventricular segmentation network model integrates the advantages of multiple modules, improving the performance of the obtained segmentation network model and increasing the accuracy and precision of image segmentation for medical images.

[0045] Figure 3 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application. Figure 3 As shown, step S12 includes steps S121 to S122.

[0046] Step S121: Filter the medical images to obtain the filtered medical images.

[0047] Step S122: Perform lesion region annotation and key structure annotation processing on the screened medical images to obtain the medical image dataset.

[0048] In some possible implementations, ultrasound medical images from different time periods and under different conditions are acquired, and the data undergoes preliminary screening to remove images of poor quality, incomplete information, or that do not meet research standards, thus obtaining the screened images. Each screened image is analyzed to determine lesion regions and key structures, and segmentation masks are added to each screened image using specialized annotation tools and software to obtain the medical image dataset. During the annotation process of the screened medical images, the annotation data is reviewed and verified to ensure the high quality and reliability of the annotation results.

[0049] Figure 4 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application. Figure 4 As shown, the depthwise convolution module includes a first depthwise convolution module and a second depthwise convolution module, and step S13 includes steps S131 to S135.

[0050] Step S131: Use the first deep convolution module to extract features from the medical image dataset to obtain a shallow feature map.

[0051] Step S132: The shallow feature map is processed using the integrated coding module to obtain a deep feature map.

[0052] Step S133: The deep feature map is processed using the bottleneck block to obtain the processed feature map.

[0053] Step S134: Use the integrated decoding module to decode the processed feature map to obtain a decoded map.

[0054] Step S135: The decoded image is processed using the second depth convolution module to obtain the processing result of the medical image dataset.

[0055] In some possible implementations, for any medical image in the input medical image dataset Where C, H, and W represent the number of channels, height, and width of the image, respectively. A first depthwise convolutional module is used to extract features from the medical image dataset to obtain a shallow feature map. The ensemble encoding module then processes the shallow feature map to obtain a deep feature map. The ensemble encoding module includes three encoding blocks; after each block, the number of channels in the shallow feature map is doubled while the resolution is halved. A bottleneck block is used to process the deep feature map to obtain a processed feature map. The bottleneck block is used to model long-distance spatial dependencies while maintaining the feature map size. The ensemble decoding module then decodes the processed feature map to obtain a decoded map. The ensemble decoding module includes three decoding blocks for feature decoding and image resolution restoration. After each decoding block, the number of channels in the processed feature map is halved while the resolution is doubled. The output of the last decoding block achieves the same resolution as the original image.

[0056] In other possible implementations, bottleneck blocks, as a commonly used neural network structure, aim to increase the network's expressive power while maintaining computational efficiency. This is achieved by reducing the dimensionality of intermediate layers to decrease computational cost, and then restoring the original dimensionality, thus enabling efficient feature extraction and propagation. Assuming the input is X, the computation process of the bottleneck block is as follows: , in, , and These represent the outputs of each layer in the middle of the bottleneck block, Indicates the number of intermediate channels.

[0057] The processed feature map is decoded using an integrated decoding module to restore the image resolution. Specifically, the feature map obtained from the skip connection MLP-Mixer... and the output of the previous block The decoding block performs feature fusion through addition. It utilizes a deep convolutional module, a skip connection MLP-Mixer module, and adjustments to the scale factor and ReLU activation function to decode the feature maps. The outputs of the three decoding blocks are as follows: , in, This indicates the output of the MLP-Mixer in the skip connection (the three MLP-Mixer outputs corresponding to the three decoded blocks). This indicates the output of the previous decoded block.

[0058] Output of the last decoded block Combined with the output of the first MLP-Mixer The input is then fed into the second depthwise convolutional module to obtain the final segmentation mask. , and They are represented as follows: , in, This represents the original input medical image. The decoding block ultimately uses interpolation to restore the prediction to the original resolution.

[0059] Figure 5 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application. Figure 5 As shown, the integrated encoding module includes an encoding block and a Mamba module, and step S132 includes steps S1321 to S1322.

[0060] Step S1321: Process the shallow feature map using the coding block to obtain the processed shallow feature map.

[0061] Step S1322: Use the Mamba module to process the processed shallow feature map to obtain the deep feature map.

[0062] Specifically, the Mamba module is a Mamba module. Methods based on the Mamba model, such as U-Mamba and VM-Unet, can maintain the advantage of linear computational complexity while modeling long-distance interactions by utilizing Mamba's state space model (SSM).

[0063] In some possible implementations, given a shallow feature map Input, where Indicates the first Each coding block initially flattens and transposes the feature map into a shape of ( ). The encoding block utilizes N consecutive RVMM (Resource and Virtual Memory Management) layers to capture global information, increasing the number of channels in the last RVMM layer. The encoding block reshapes and transposes the feature map to its desired shape, then performs max pooling to reduce the feature map's resolution. Finally, the... Output deep feature map of each coding block .

[0064] In one embodiment of this application, processing the shallow feature map using the integrated coding module to obtain the deep feature map further includes: pooling the output of the Mamba module using a max pooling module to obtain the deep feature map.

[0065] In some possible implementations, a max pooling module is used to perform max pooling operations to reduce the resolution of the feature map and obtain the deep feature map. .

[0066] Figure 6 The diagram shown is a flowchart illustrating the training method for the ultrasound left ventricular segmentation network model described in this application. Figure 6 As shown, the Mamba module includes a visual state space module, an adjustment factor, a normalization layer, and a residual visual module. Step S1322 includes steps S13221 to S13222.

[0067] Step S13221: The shallow feature map is processed by the normalization layer and the visual state space module, and then convolved with the processing result of the adjustment factor and the residual visual module to obtain the convolution result.

[0068] Step S13222: Process the convolution result using another normalization layer to obtain the deep feature map.

[0069] In some possible implementations, the shallow feature map output by the first depthwise convolutional module The integrated coding block is input. The shallow feature map is processed using the coding block to obtain a processed shallow feature map. The Mamba module uses the normalization layer and the Visual State Space Module (VSSM) to process the shallow feature map, capturing long-range spatial dependencies. Performance is improved in the residual connections using the adjustment factor and the residual vision module. The output of the Mamba module is: , in, Indicates intermediate output. Indicates the adjustment factor. This represents the shallow feature map output by the first depthwise convolutional module. This represents the visual state space module.

[0070] The Mamba module then uses another normalization layer to normalize the convolution result, and uses a projection layer to transform it into the deep feature map. The deep feature map is represented as follows: .

[0071] Visual state space module The module passes the input features to two branches. The first branch expands the channels using a linear layer and applies depthwise convolution, SiLU activation, SSM, and layer normalization. The second branch only expands the channels using a linear layer and applies SiLU activation. Finally, the outputs of the two branches are fused using a Hadamard product and mapped back to the original number of channels using a linear layer. The output of the first branch is as follows: , The output of the second branch is: , The final output is: , in, This indicates the input characteristics of the VSS module. The VSS is divided into two paths. This represents the output characteristics of the first branch. This represents the output characteristics of the second branch. This indicates the output characteristics of the layer after merging the two branches. This represents the dot product.

[0072] In one embodiment of this application, the training method for the ultrasound left ventricular segmentation network model further includes: using a residual vision module to make skip connections between the deep convolution modules, between the integrated coding module and the integrated decoding module.

[0073] Figure 7 The diagram shown is a schematic representation of the segmentation network model described in an embodiment of this application. Figure 7As shown, the medical image to be segmented is input into the segmentation network model. It is first processed by a first deep convolutional module, then passed through the first encoding block of the ensemble encoding module before being input into a second encoding block. The second encoding block includes a Mamba module and a max-pooling module. The shallow feature map is processed using the normalization layer and visual state space module of the Mamba module, and then convolved with the processing result of the adjustment factor and the residual visual module to obtain the convolution result. The convolution result is processed using the normalization layer and transformed into the deep feature map using a projection layer. The output of the Mamba module is pooled using the max-pooling module to obtain the deep feature map. The deep feature map is processed by the third encoding block of the ensemble encoding block and then input into a bottleneck block. The deep feature map is processed using the bottleneck block to obtain the processed feature map.

[0074] Visual state space module The module passes the input features to two branches. The first branch expands the channels using a linear layer and applies depthwise convolution, SiLU activation, SSM, and layer normalization. The second branch only uses a linear layer to expand the channels and applies SiLU activation. Finally, the outputs of the two branches are fused using a Hadamard product and mapped back to the original number of channels using a linear layer. The Mamba module then normalizes the convolution result using another normalization layer and transforms it into the deep feature map using a projection layer.

[0075] The processed feature map is decoded using three decoding blocks of the integrated decoding module to obtain a decoded image. After each decoding block, the number of channels in the processed feature map is halved while the resolution is doubled, with the output of the last decoding block achieving the same resolution as the original image. The input of each decoding block corresponds to the output of an MLP-Mixer. The three decoding blocks correspond to the three MLP-Mixers.

[0076] The decoded image is processed using a second depthwise convolutional module to obtain the processing result of the medical image dataset. The second depthwise convolutional module of the first depthwise convolutional module uses skip connections via an MLP-Mixer.

[0077] It should be noted that the above is only one possible implementation of the embodiments of this application, and this application is not limited thereto.

[0078] Figure 8 The diagram shown is a structural schematic of the ultrasound left ventricular segmentation network model training system described in this application embodiment. Figure 8As shown, the ultrasound left ventricular segmentation network model training system 100 includes a data acquisition module 110, a data processing module 120, and a model training module 130.

[0079] The data acquisition module 110 is used to acquire medical images.

[0080] The data processing module 120 is used to filter and label the medical images to obtain a medical image dataset.

[0081] The model training module 130 is used to train an initial deep learning model using the medical image dataset to obtain a segmentation network model. The initial deep learning model is constructed using a deep convolutional module, an ensemble encoder module, an ensemble decoder module, and a bottleneck block.

[0082] In some possible implementations, the deep convolutional module of the initial deep learning model includes a first deep convolutional module and a second deep convolutional module. After acquiring medical images, the images are filtered and labeled to obtain a medical image dataset. Ultrasound medical images from different time periods and under different conditions are acquired, and the data undergoes preliminary filtering to remove images of poor quality, incomplete information, or that do not meet research standards, thus obtaining the filtered images. Each filtered image is analyzed to determine lesion regions and key structures, and segmentation masks are added to each filtered image using professional annotation tools and software to obtain the medical image dataset. During the annotation process of the filtered medical images, the labeled data is reviewed and verified to ensure the high quality and reliability of the annotation results. The medical image dataset is then sequentially passed through the first deep convolutional module, ensemble encoding module, bottleneck block, ensemble decoding module, and second deep convolutional module of the initial deep learning model to obtain the segmentation results of the medical image dataset.

[0083] In this embodiment, an initial deep learning model is constructed using a deep convolution module, an ensemble encoding module, an ensemble decoding module, and a bottleneck block. The model training module 130 is used to train the initial deep learning model using a medical image dataset to obtain a segmentation network model. This ultrasound left ventricular segmentation network model training system 100 can integrate the advantages of multiple modules, improve the performance of the acquired segmentation network model, and increase the accuracy and precision of image segmentation for medical images.

[0084] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or modules or units may be electrical, mechanical, or other forms.

[0085] The modules / units described as separate components may or may not be physically separate. The components shown as modules / units may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected to achieve the objectives of the embodiments of this application, depending on actual needs. For example, the functional modules / units in the various embodiments of this application may be integrated into one processing module, or each module / unit may exist physically separately, or two or more modules / units may be integrated into one module / unit.

[0086] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0087] This application also provides an electronic device. Figure 9 The diagram shown is a structural schematic of the electronic device described in an embodiment of this application. Figure 9 As shown, in this embodiment, the electronic device 900 includes a memory 910 and a processor 920.

[0088] The memory 910 is used to store computer programs; preferably, the memory 910 includes various media that can store program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk.

[0089] Specifically, memory 910 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic device 900 may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 910 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of this application.

[0090] The processor 920 is connected to the memory 910 and is used to execute the computer program stored in the memory 910 so that the electronic device 900 executes the ultrasound left ventricular segmentation network model training method described in any embodiment of this application.

[0091] Optionally, the processor 920 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0092] Optionally, the electronic device 900 in this embodiment may further include a display 930. The display 930 is communicatively connected to the memory 910 and the processor 920, and is used to display the relevant graphical user interface (GUI) of the ultrasound left ventricular segmentation network model training method described in this application embodiment.

[0093] This application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the program implements the ultrasound left ventricular segmentation network model training method described in any embodiment of this application.

[0094] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0095] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for training an ultrasound left ventricular segmentation network model, characterized in that, include: Acquiring medical images; The medical images are filtered and labeled to obtain a medical image dataset; The initial deep learning model is trained using the medical image dataset to obtain a segmentation network model; The initial deep learning model is constructed using a deep convolutional module, an ensemble encoding module, an ensemble decoding module, and a bottleneck block.

2. The training method for the ultrasound left ventricular segmentation network model according to claim 1, characterized in that, The medical images are filtered and labeled to obtain a medical image dataset, including: The medical images are filtered to obtain the filtered medical images; The selected medical images are then annotated with lesion regions and key structures to obtain the medical image dataset.

3. The training method for the ultrasound left ventricular segmentation network model according to claim 1, characterized in that, The deep convolutional module includes a first deep convolutional module and a second deep convolutional module. Training the initial deep learning model using the medical image dataset includes: The first depthwise convolutional module is used to extract features from the medical image dataset to obtain a shallow feature map. The shallow feature map is processed using the integrated coding module to obtain a deep feature map; The bottleneck block is used to process the deep feature map to obtain a processed feature map; The processed feature map is decoded using an integrated decoding module to obtain a decoded map; The decoded image is processed using a second depthwise convolution module to obtain the processing results of the medical image dataset.

4. The training method for the ultrasound left ventricular segmentation network model according to claim 3, characterized in that, The integrated coding module includes coding blocks and a Mamba module. The integrated coding module is used to process the shallow feature map to obtain a deep feature map, including: The shallow feature map is processed using the coding block to obtain a processed shallow feature map; The Mamba module is used to process the shallow feature map to obtain the deep feature map.

5. The method for training an ultrasound left ventricular segmentation network model according to claim 4, characterized in that, The process of processing the shallow feature map using the integrated coding module to obtain the deep feature map further includes: The output of the Mamba module is pooled using a max pooling module to obtain the deep feature map.

6. The method for training an ultrasound left ventricular segmentation network model according to claim 4, characterized in that, The Mamba module includes a visual state space module, an adjustment factor, a normalization layer, and a residual visual module. The Mamba module is used to process the processed shallow feature map to obtain the deep feature map, including: The shallow feature map is processed by the normalization layer and the visual state space module, and then convolved with the processing result of the adjustment factor and the residual visual module to obtain the convolution result. The convolution result is processed using another normalization layer to obtain the deep feature map.

7. The method for training an ultrasound left ventricular segmentation network model according to claim 1, characterized in that, Also includes: The residual vision module is used to make skip connections between the depth convolution modules, the integrated encoding module, and the integrated decoding module.

8. A training system for an ultrasound left ventricular segmentation network model, characterized in that, include: The data acquisition module is used to acquire medical images; The data processing module is used to filter and label the medical images to obtain a medical image dataset; The model training module is used to train the initial deep learning model using the medical image dataset to obtain a segmentation network model. The initial deep learning model is constructed using a deep convolutional module, an ensemble encoding module, an ensemble decoding module, and a bottleneck block.

9. An electronic device, characterized in that, The electronic device includes: Memory, used to store computer programs; A processor for executing a computer program stored in the memory to cause the electronic device to perform the ultrasound left ventricular segmentation network model training method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the training method for the ultrasound left ventricular segmentation network model according to any one of claims 1 to 7.

Images