Lightweight cardiac ultrasound left ventricle segmentation method based on biological visual inspiration
Through the biovision-inspired lightweight cardiac ultrasound left ventricular segmentation method, the antagonistic module and direction selection module structure are used, combined with the Fusion-Up module, the problem of large calculations in the existing technology is solved, and efficient and accurate Cardiac ultrasound image segmentation is suitable for applications in resource-constrained environments.
Patent Information
- Application Number
- CN202411330281.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-24
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing automatic cardiac ultrasound image segmentation method relies on complex and computationally large deep learning models, limiting its application in resource-constrained environments.
Using a lightweight cardiac ultrasound left ventricular segmentation method based on biovision inspired, the model structure including antagonistic module and direction selection module is constructed, and the Fusion-Up module is combined for step-by-step upsampling, significantly reducing the amount of parameters and calculation needs, using the biovision system Extract cardiac ultrasound image features.
It realizes efficient and accurate cardiac ultrasound image segmentation, significantly reduces hardware costs, and improves the application applicability of the model in resource-constrained environments. The calculation volume is only 1.1 GFLOPS, the FPS reaches 120+, and it performs in accuracy and efficiency. Excellent.
Smart Images

Figure CN119068199A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of computers, and in particular relates to a lightweight cardiac ultrasound left ventricle segmentation method based on biological vision inspiration. Background Art
[0002] Cardiac ultrasound (echocardiography) is an important medical imaging technology. It provides doctors with detailed information about the structure and function of the heart and has been widely used in the diagnosis and treatment of various heart diseases. This imaging method has the advantages of being non-invasive, real-time, and dynamic. Clinicians need to choose and apply it according to specific needs.
[0003] Cardiac ultrasound image segmentation is an important technology that plays a key role in medical image analysis. By accurately segmenting the cardiac structure from ultrasound images, doctors can obtain clearer and more detailed cardiac information, which helps diagnose and treat some heart diseases. The accuracy of cardiac ultrasound segmentation of the corresponding area can not only provide accurate cardiac structure and function information, but also help doctors make correct clinical judgments and formulate effective treatment plans.
[0004] In cardiac ultrasound diagnosis, manual labeling of the left ventricular region is time-consuming and susceptible to observer bias. In order to overcome the limitations of traditional manual segmentation methods, automated segmentation techniques based on computer vision and deep learning have gradually emerged in recent years, which have significantly improved the efficiency and accuracy of cardiac ultrasound image segmentation. In the early days, researchers proposed some machine learning methods, such as Leclerc using structured random forests and Huang et al. developing dynamic appearance models to segment the left ventricle. However, these methods either rely on hand-crafted features or are not powerful enough. With the deepening of research, deep learning technology, especially convolutional neural networks (CNNs), has been widely used in cardiac image segmentation tasks. Smistad et al. developed a UNet-based model specifically for the segmentation of the left ventricle in two-dimensional cardiac ultrasound. In addition, Oktay et al. extended the model and used anatomical prior knowledge for location-based guidance, further improving the segmentation effect based on UNet. With the introduction of the Transformer architecture, Deng et al. applied this method with a global receptive field to the segmentation of the left ventricle in cardiac ultrasound and achieved good results. These techniques reduce the need for human intervention by automatically learning features and patterns in images, enabling accurate, efficient, and consistent segmentation of cardiac structures.
[0005] In the above studies, these advanced models usually have large parameters and complex computational requirements. Although this large number of parameters and computational complexity enable the model to learn image features more deeply and provide excellent segmentation performance, it also brings dependence on high-performance computing resources. In practical applications, this may limit the popularity and use of these advanced technologies, especially in resource-constrained environments. Therefore, although these models are technologically cutting-edge, they still need to consider cost-effectiveness and practical operability when deployed and operated. Summary of the invention
[0006] In view of the problem that existing automatic segmentation methods rely on complex and computationally intensive deep learning models, which limits their application in resource-constrained environments, the present invention proposes an efficient and ultra-lightweight cardiac ultrasound segmentation model method to improve the accuracy and efficiency of cardiac ultrasound segmentation, while significantly reducing hardware costs and expanding the scope of technical application.
[0007] In order to solve the above problems, the present invention adopts the following technical solutions: A lightweight cardiac ultrasound left ventricle segmentation method based on biological vision inspiration, the method comprising the following steps: Step 1) determining a cardiac ultrasound dataset; Step 2) Based on the biological visual system, a cardiac ultrasound left ventricle segmentation model structure is constructed, including an int (initialization) module, a Retina (retina) module, an LGN (lateral geniculate body) module, a V1 (primary visual cortex) module, and a Fusion-Up module, and a maximum pooling layer is connected after the Retina module and the LGN module, the Retina module and the LGN module are antagonistic modules, and the V1 module is a direction selection module; the int module receives the initial cardiac ultrasound image, adjusts the number of channels of the image through the convolution module, and inputs the initialized ultrasound image into the antagonistic module, and the antagonistic module uses peripheral-center antagonistic and center-peripheral antagonistic convolution to extract the feature information of the inner cavity and the wall of the left ventricle, and then extracts the left ventricle contour boundary information through the direction selection module; then, two Fusion-Up modules are used for step-by-step upsampling to output the left ventricle segmentation result of the cardiac ultrasound image; Step 3) Based on the constructed cardiac ultrasound data set and in combination with the loss function, the constructed cardiac ultrasound left ventricle segmentation model structure is trained to obtain a trained model, namely, a cardiac ultrasound left ventricle segmentation model; Step 4) According to the target cardiac ultrasound image, a cardiac ultrasound left ventricle segmentation model is applied to obtain a target cardiac ultrasound left ventricle segmentation image.
[0008] Furthermore, the CAMUS (Cardiac Acquisitions for Multi-structure Ultrasound Segmentation) cardiac ultrasound dataset is used as the cardiac ultrasound dataset in step 1).
[0009] Furthermore, the Retina module adopts a 3×3 convolution kernel, subtracts the center weight value from the peripheral weight value of the convolution kernel and adjusts the center weight value to zero, uses the adjusted 3×3 convolution kernel to perform peripheral-center antagonistic convolution, and after InstanceNorm normalization and ReLU activation function operation, subtracts the weight value of the 3×3 convolution kernel from the sum of the peripheral weight values, and implements re-parameterization of the convolution kernel to become a 1×1 convolution kernel, performs center-peripheral antagonistic convolution, and finally reduces the height and width dimensions of the image to 1 / 2 of the original through a maximum pooling layer pooling to extract the characteristic information of the left ventricular cavity and wall.
[0010] Furthermore, due to the hierarchical mechanism of biological vision, LGN has a larger receptive field than Retina, so the expansion coefficient of the 3×3 peripheral-central antagonistic convolution of the LGN module is set to 2. The image information is processed by the Retina module and transmitted to the LGN module, and the same antagonistic convolution operation as in the Retina module is performed. Finally, the height and width dimensions of the image are reduced to 1 / 4 of the original through pooling in the maximum pooling layer, and the characteristic information of the left ventricular cavity and wall is further extracted.
[0011] Furthermore, the peripheral-center antagonistic convolution is set as a grouped convolution, and the number of groups is equal to the number of channels to reduce the number of parameters.
[0012] Furthermore, the V1 module, i.e., the direction selection module, constructs a convolution kernel and sets the value of the convolution kernel to zero in vertical, horizontal, left oblique, and right oblique directions, respectively, to perform convolution to extract the left ventricle contour boundary information.
[0013] Furthermore, in order to avoid the loss of extracted image information due to direct upsampling, two Fusion-Up modules are used for step-by-step upsampling. The Fusion-Up module adjusts the number of channels of low-resolution feature input and high-resolution feature input through convolution operation, and then fuses and outputs the information of the two input channels after InstanceNorm normalization and ReLU activation function operation respectively.
[0014] Furthermore, a 1×1 convolution operation with only one channel is performed on the image information output by the second Fusion-Up module, and then an activation operation of the Sigmoid activation function is performed to output a left ventricle segmentation image with the same size as the initial image.
[0015] Furthermore, the cardiac ultrasound left ventricle segmentation model structure was constructed based on the Python 3.9 environment and the PyTorch framework, and the Adam optimizer with an initial learning rate of 0.0001 was used during training, the regularization step size was set to 1, and the gamma was set to 0.75; the number of training rounds was set to 500, and the batch size was 16.
[0016] Furthermore, the BCEWithLogitsLoss loss function is used during training, which combines the calculation of the sigmoid activation function and the binary cross-entropy loss (BCE) function, which can improve numerical stability and reduce the number of calculation steps. The calculation formula of BCEWithLogitsLoss is as follows:
[0017] Among them, N is the batch size, y is the true value, and z is the predicted value. is the sigmoid activation function.
[0018] Further, based on the determined cardiac ultrasound data set, the constructed cardiac ultrasound left ventricle segmentation model structure is trained to obtain a trained model, namely, a cardiac ultrasound left ventricle segmentation model.
[0019] Furthermore, according to the target cardiac ultrasound image, the cardiac ultrasound left ventricle segmentation model is applied to obtain the target cardiac ultrasound left ventricle segmentation image, and the Dice coefficient and mIoU (mean Intersection over Union) indicators are used to measure the accuracy and consistency of the results. At the same time, the parameter quantity and FPS (Frames Per Second) are introduced to comprehensively evaluate the performance of the model from different angles.
[0020] Compared with the prior art, the present invention has the following beneficial effects: First, our method is inspired by the high efficiency of biological visual systems. By simulating the feature extraction mechanism of GC (ganglion cells), LGN (lateral geniculate body) and V1 (primary visual cortex) in the visual system, the antagonistic module and direction selection module developed can effectively extract features from cardiac ultrasound images, and greatly reduce the number of model parameters and computing requirements while maintaining segmentation accuracy, significantly reducing the dependence on high-performance computing resources; experimental verification results show that our model only requires 1.1 GFLOPS of computing when processing a 224x224 pixel image with 0.1M parameters, and the FPS on the RTX 3090 graphics card is 120+; Secondly, in order to better restore the segmentation details, the present invention adopts a step-by-step upsampling method to achieve a more refined segmentation result. Compared with other models, the model constructed by the present invention performs equally well or even better in the accurate segmentation of the left ventricle. The segmentation method provided by the present invention not only breaks through the limitations of the existing model technically, but also provides a more lightweight and efficient solution for the application of cardiac ultrasound image segmentation technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a structural schematic diagram of the segmentation model of the present invention; Figure 2 Schematic diagram of antagonistic convolution; Figure 3 Schematic diagram of the process of extracting image information for the Retina and LGN modules; Figure 4 This is a schematic diagram of the V1 module, i.e. the direction selection module; Figure 5 This is a schematic diagram of the Fusion-Up module; Figure 6 This is the implementation effect diagram. DETAILED DESCRIPTION
[0022] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0023] A lightweight cardiac ultrasound left ventricle segmentation method based on biological vision inspiration, the method comprising the following steps: Step 1) determining a cardiac ultrasound dataset; Step 2) Based on the biological vision system, construct a cardiac ultrasound left ventricle segmentation model structure, such as Figure 1 As shown, the model structure includes an int (initialization) module, a Retina (retina) module, an LGN (lateral geniculate body) module, a V1 (primary visual cortex) module, and a Fusion-up module, and the Retina module and the LGN module are both connected to a maximum pooling layer, the Retina module and the LGN module are antagonistic modules, and the V1 module is a direction selection module; the int module receives a cardiac ultrasound image with an initial size of H×W×1, and adjusts the number of channels of the image to C through a convolution module, and the initialized ultrasound image is input to the antagonistic module, and the antagonistic module uses peripheral-center antagonistic and center-peripheral antagonistic convolutions to extract the left ventricular contour information, and then the left ventricular boundary information is extracted through the direction selection module; then, two Fusion-up modules are used for stepwise upsampling, and finally, after a 1×1 convolution operation with a channel of 1 and an activation operation of a Sigmoid activation function, the cardiac ultrasound left ventricle segmentation result is output; Step 3) Based on the constructed data set and in combination with the loss function, the constructed cardiac ultrasound left ventricle segmentation model structure is trained to obtain a trained model, namely, a cardiac ultrasound left ventricle segmentation model; Step 4) According to the target cardiac ultrasound image, a cardiac ultrasound left ventricle segmentation model is applied to obtain a target cardiac ultrasound left ventricle segmentation image.
[0024] Furthermore, the CAMUS (Cardiac Acquisitions for Multi-structure Ultrasound Segmentation) cardiac ultrasound dataset is used as the cardiac ultrasound dataset for step 1); the CAMUS dataset is a public, professionally annotated cardiac ultrasound dataset designed for automatic segmentation of cardiac structures; the dataset includes two-dimensional cardiac ultrasound videos from 500 patients, each video contains multiple frames at different stages of the cardiac cycle; in order to evaluate the performance of the cardiac segmentation model, key cardiac structures such as the left ventricle, right ventricle, left atrium, and right atrium in each frame are carefully annotated by experienced clinical experts.
[0025] Furthermore, usually in cardiac ultrasound images, the cavities of the ventricles appear as relatively dark areas because they are cavities filled with blood, while the walls of the heart appear as brighter areas. The brightness comes from the high density of the myocardial tissue, which can reflect ultrasound waves more, so that there is a clear contrast between the cavities of the ventricles and the walls of the heart. In biological vision, center-surround antagonism and surround-center antagonism are two important information processing modes in the visual nervous system. These two mechanisms are extremely important for enhancing visual contrast and are the basic mechanisms for the visual system to parse environmental details. Based on this, the present invention designs antagonistic modules that use surround-center and center-surround antagonistic convolutions, namely the Retina module and the LGN module, to extract characteristic information of the left ventricular cavity and heart wall. Figure 2 and 3 As shown in the figure, the Retina module uses a 3×3 convolution kernel, subtracts the center weight value from the peripheral weight value of the convolution kernel and adjusts the center weight value to zero, uses the adjusted 3×3 convolution kernel to perform peripheral-center antagonistic convolution, and the peripheral-center antagonistic convolution is set to grouped convolution, the number of groups is equal to the number of channels, to reduce the number of parameters, after InstanceNorm normalization and ReLU activation function operation, the center weight value of the 3×3 convolution kernel is subtracted from the sum of the peripheral weight values, and the convolution kernel is re-parameterized to become a 1×1 convolution kernel, and center-peripheral antagonistic convolution is performed. Finally, the height and width dimensions of the image are reduced to 1 / 2 of the original through a maximum pooling layer, and the number of channels is 2C.
[0026] Furthermore, due to the hierarchical mechanism of biological vision, LGN has a larger receptive field than Retina, so the expansion coefficient of the 3×3 peripheral-center antagonistic convolution of the LGN module is set to 2, and the image information is processed by the Retina module and transmitted to the LGN module, and the same antagonistic convolution operation as in the Retina module is performed. The height and width dimensions of the image are reduced to 1 / 4 of the original through pooling in the maximum pooling layer, and the number of channels is 4C.
[0027] Further, if Figure 4 As shown, the V1 module, i.e., the direction selection module, performs convolution by constructing a convolution kernel and setting the value of the convolution kernel to zero in the vertical, horizontal, left oblique, and right oblique directions, so as to better extract the left ventricular contour boundary information.
[0028] Further, if Figure 5 As shown in the figure, in order to avoid the loss of extracted image information caused by direct upsampling, two Fusion-Up modules are used for step-by-step upsampling. The Fusion-Up module adjusts the number of channels of low-resolution feature input and high-resolution feature input through 5×5 convolution operation, and then fuses and outputs the information of the two input channels after InstanceNorm normalization and ReLU activation function operation respectively.
[0029] Furthermore, a 1×1 convolution operation with only one channel is performed on the image information output by the second Fusion-Up module, and then an activation operation of the Sigmoid activation function is performed to output a left ventricle segmentation image with the same size as the initial image.
[0030] Furthermore, the cardiac ultrasound left ventricle segmentation model structure was constructed based on the Python 3.9 environment and the PyTorch framework, and the Adam optimizer with an initial learning rate of 0.0001 was used during training, the regularization step size was set to 1, and the gamma was set to 0.75; the number of training rounds was set to 500, and the batch size was 16.
[0031] Furthermore, the BCEWithLogitsLoss loss function is used to train the constructed cardiac ultrasound left ventricle segmentation model structure. It combines the calculation of the sigmoid activation function and the binary cross-entropy loss function (BCE), which can improve numerical stability and reduce calculation steps. The calculation formula of BCEWithLogitsLoss is as follows:
[0032] Among them, N is the batch size, y is the true value, and z is the predicted value. is the sigmoid activation function.
[0033] Furthermore, based on the determined cardiac ultrasound dataset, the constructed heart model structure was trained and tested on a computer with an Intel CPU i7 12700 processor and an NVIDIA RTX 2080ti graphics card to obtain a cardiac ultrasound left ventricle segmentation model.
[0034] Furthermore, according to the target cardiac ultrasound image, the cardiac ultrasound left ventricle segmentation model is applied to obtain the target cardiac ultrasound left ventricle segmentation image. The Dice coefficient and mloU (mean Intersection over Union) are used to measure the accuracy and consistency of the results. At the same time, the parameter quantity and FPS (Frames Per Second) are introduced to comprehensively evaluate the performance of the model from different angles: Dice coefficient The Dice coefficient is an indicator for evaluating the similarity between two samples and is widely used in the field of medical image segmentation. It mainly measures the degree of overlap between the segmented area predicted by the model and the true annotation. The formula is as follows:
[0035] Where P represents the predicted segmentation area, and G represents the real segmentation area; the value of the Dice coefficient is between 0 and 1, and the closer the value is to 1, the higher the consistency between the predicted result and the real data; mIoU(mean Intersection over Union): mIoU is another core indicator used to quantify the performance of multi-class image segmentation tasks. It calculates the average value of the intersection over union ratio between the predicted segmentation and the true segmentation. It is applicable to situations involving multiple categories. The calculation formula is:
[0036] Where N is the number of categories, P i and G i They represent the predicted and true segmented areas of the i-th category respectively. The value range of mIoU is also 0 to 1. The higher the value, the more accurate the prediction of the model. c) Parameter: Parameter count refers to the total number of parameters that make up the model, which is a direct indicator of model size and complexity. The larger the parameter count, the more complex the model structure is, which may require more computing resources and storage space. In resource-limited application scenarios, it is very important to optimize the parameter count to achieve a smaller model size. d) FPS (Frames Per Second): FPS indicates the number of frames that the model can process per second and is a key indicator for measuring the efficiency of the model. The higher the FPS value, the better the real-time performance of the model. Example
[0037] The left ventricle of the target cardiac ultrasound image was segmented using our model and the UNet and AttSWinUnet models, and the segmentation image performance was evaluated using the evaluation indicators.
[0038] Table 1
[0039] As can be seen from Table 1, when the input is 224×224×3, compared with UNet, AttSWinUnet and our model, our model has obvious advantages in terms of parameter quantity, computing resource requirements and frame rate. Specifically, our model has only 0.11M parameters, which is 98.6% less than UNet's 7.76M and 99.7% less than AttSWinUnet's 41.34M, greatly reducing the model's storage and computational burden; in terms of computational workload, our model only requires 1.12GFLOPS, which is 94.7% and 93.7% less than UNet's 21.04 GFLOPS and AttSWinUnet's 17.73 GFLOPS, respectively, which means that our model has lower energy consumption and faster processing speed; on the high-performance NVIDIA GeForce RTX 3090 GPU, our model achieves a processing speed of 123 FPS, which is 50% higher than UNet's 82 FPS and 105% higher than AttSWinUnet's 60 FPS, demonstrating its advantage in high-speed processing. These comparisons and analyses show that our model significantly improves operating efficiency and reduces resource consumption while maintaining high-precision image segmentation effects, making it an ideal choice for real-time and resource-constrained applications.
[0040] As can be seen from Table 2, when analyzing the performance of our model in the segmentation of left ventricular end-diastole (LV ED) and left ventricular end-systole (LV ES) of cardiac function, the segmentation effect of our model is comparable to that of UNet and better than AttSWinUnet. At the end of diastole, the Dice coefficient of our model is 0.938, which is very close to UNet's 0.945 and better than AttSWinUnet's 0.927; in the mIoU index, the mIoU index of our model is 0.884, which is close to UNet's 0.895 and better than AttSWinUnet's 0.864; at the end of systole, our model maintains a similar performance trend, with a Dice coefficient of 0.916, second only to UNet's 0.917, and significantly better than AttSWinUnet's 0.849; mIoU is the same as UNet, far exceeding AttSWinUnet's 0.739; Figure 6 The results of our model, UNet and AttSWinUnet models on the same patient in ED and ES are shown.
[0041] Table 2
[0042] The lightweight cardiac ultrasound left ventricle segmentation method based on biological vision inspiration provided by the present invention has extremely low parameter volume (0.11M) and computing resource requirements, making it very suitable for deployment on resource-constrained devices; in terms of performance, although the parameters and calculation amount are greatly reduced, the image segmentation accuracy is still comparable to larger models such as UNet, especially in terms of processing speed, which can reach 123 FPS on high-end GPUs, far exceeding similar models, and can maintain 45FPS on low-end graphics cards, ensuring high efficiency in application scenarios that require real-time processing; the method provided by the present invention is an ideal choice for application environments that require fast, accurate and low computational costs.
Claims
1. A lightweight cardiac ultrasound left ventricle segmentation method based on biological vision inspiration, It is characterized in that The method comprises the following steps: Step 1) determining a cardiac ultrasound dataset; Step 2) Based on the biological vision system, a cardiac ultrasound left ventricle segmentation model structure is constructed, including an int module, a Retina module, a LGN module, a V1 module, and a Fusion-Up module, and a maximum pooling layer is connected after the Retina module and the LGN module, the Retina module and the LGN module are antagonistic modules, and the V1 module is a direction selection module; the int module receives the initial cardiac ultrasound image, adjusts the number of channels of the image through the convolution module, and inputs the initialized ultrasound image into the antagonistic module, and the antagonistic module uses peripheral-center antagonistic and center-peripheral antagonistic convolution to extract the feature information of the inner cavity and the wall of the left ventricle, and then extracts the left ventricle boundary information through the direction selection module; then, two Fusion-Up modules are used for stepwise upsampling to output the left ventricle segmentation result of the cardiac ultrasound image; Step 3) Based on the constructed cardiac ultrasound data set and in combination with the loss function, the constructed cardiac ultrasound left ventricle segmentation model structure is trained to obtain a trained model, namely, a cardiac ultrasound left ventricle segmentation model; Step 4) According to the target cardiac ultrasound image, a cardiac ultrasound left ventricle segmentation model is applied to obtain a target cardiac ultrasound left ventricle segmentation image.
2. A lightweight cardiac ultrasound left ventricle based on biological vision inspiration according to claim 1 Segmentation method, It is characterized in that The CAMUS cardiac ultrasound dataset is used as the cardiac ultrasound dataset in step 1).
3. A lightweight cardiac ultrasound left ventricle based on biological vision inspiration according to claim 2 Segmentation method, It is characterized in that The Retina module uses a 3×3 convolution kernel, subtracts the center weight value from the peripheral weight value of the convolution kernel and adjusts the center weight value to zero, uses the adjusted 3×3 convolution kernel to perform peripheral-center antagonistic convolution, and after InstanceNorm normalization and ReLU activation function operation, subtracts the weight value of the 3×3 convolution kernel from the sum of the peripheral weight values, implements re-parameterization of the convolution kernel to become a 1×1 convolution kernel, performs center-peripheral antagonistic convolution, and finally reduces the height and width dimensions of the image to 1 / 2 of the original through a maximum pooling layer pooling to extract the characteristic information of the left ventricular cavity and wall.
4. A lightweight cardiac ultrasound left ventricle based on biological vision inspiration according to claim 3 Segmentation method, It is characterized in that The expansion coefficient of the 3×3 peripheral-central antagonistic convolution of the LGN module is set to 2. The image information is processed by the Retina module and transmitted to the LGN module, and the same antagonistic convolution operation as in the Retina module is performed. Finally, the height and width dimensions of the image are reduced to 1 / 4 of the original through a maximum pooling layer, and the characteristic information of the left ventricular cavity and the wall is further extracted.
5. A lightweight cardiac ultrasound left ventricle based on biological vision inspiration according to claim 4 Segmentation method, It is characterized in that The peripheral-center antagonistic convolution is set as a grouped convolution, and the number of groups is equal to the number of channels.
6. A lightweight cardiac ultrasound left ventricle based on biological vision inspiration according to claim 5 Segmentation method, It is characterized in that The V1 module constructs a convolution kernel and sets the value of the convolution kernel to zero in vertical, horizontal, left oblique and right oblique directions to perform convolution and extract the left ventricle contour boundary information.
7. A lightweight cardiac ultrasound left ventricle based on biological vision inspiration according to claim 6 Segmentation method, It is characterized in that Two Fusion-Up modules are used for step-by-step upsampling. The Fusion-Up modules adjust the number of channels of low-resolution feature input and high-resolution feature input through convolution operations, and then fuse and output the information of the two input channels after InstanceNorm normalization and ReLU activation function operations respectively. A 1×1 convolution operation with only one channel is performed on the image information output by the second Fusion-Up module, and then an activation operation of the Sigmoid activation function is performed to output a left ventricle segmentation image with the same size as the initial image.
8. A lightweight cardiac ultrasound ventricular separation device based on biological vision inspiration according to claim 7 Cutting method, It is characterized in that The cardiac ultrasound left ventricle segmentation model structure was constructed based on the Python 3.9 environment and the PyTorch framework. The Adam optimizer with an initial learning rate of 0.0001 was used for training, the regularization step size was set to 1, and the gamma was set to 0.75; the number of training rounds was set to 500, and the batch size was 16; the BCEWithLogitsLoss loss function was used, and its calculation formula was as follows: Among them, N is the batch size, y is the true value, and z is the predicted value. is the sigmoid activation function.
9. The method for lightweight cardiac ultrasound left ventricle segmentation based on biological vision inspiration according to claim 8, It is characterized in that Based on the determined cardiac ultrasound data set, the constructed cardiac ultrasound left ventricle segmentation model structure is trained to obtain a trained model, namely, the cardiac ultrasound left ventricle segmentation model; according to the target cardiac ultrasound image, the cardiac ultrasound left ventricle segmentation model is applied to obtain a target cardiac ultrasound left ventricle segmentation image, and the Dice coefficient, mIoU indicator, parameter quantity and FPS are used to evaluate the effect of the target cardiac ultrasound left ventricle segmentation image and the performance of the cardiac ultrasound left ventricle segmentation model.
10. A lightweight cardiac ultrasound left ventricle based on biological vision inspiration according to claim 9 Segmentation method, It is characterized in that The cardiac ultrasound left ventricle segmentation model has a parameter volume of only 0.1M and a computational complexity of only 1.1GFLOPS when processing an image with a pixel size of 224×224, and the FPS on the RTX 3090 graphics card is 120+.
Citation Information
Patent Citations
Visual pathway multi-receptive-field directional association-based profile detection method
CN106127740A
Two-dimensional echocardiography left ventricular ejection fraction evaluation algorithm based on deep learning
CN118537317A
Cited By
Zero-sample anti-interference target detection method and related device
CN122049627A
A zero-shot anti-interference target detection method and related device
CN122049627B
