Heart magnetic resonance image automatic segmentation method based on deep learning

Abstract

According to the heart magnetic resonance image automatic segmentation method based on deep learning provided by the invention, the automatic segmentation process can be completed within less than 1 minute, and the traditional manual segmentation method needs more than 30 minutes on average, so that the processing time is shortened by more than 30 times, and the processing time is greatly shortened. By adopting the technical scheme provided by the invention, in actual clinical application, a doctor can obtain a diagnosis result more quickly, and the diagnosis and treatment efficiency is improved. Through an automatic deep learning model, interference of human factors is reduced, and objectivity and consistency of segmentation results are improved. According to the method, the CMR image sequences of different pathological groups are included in the training process, so that the adaptability of the model to different pathological states is improved. According to the invention, good adaptability and stability can be shown for three pathological types of HCM, DCM and VA. By optimizing the network structure and the training strategy, the demand on high-performance computing equipment is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging diagnosis, and particularly to an automatic segmentation method for cardiac magnetic resonance images based on deep learning. Background Art

[0002] In the field of medical imaging diagnosis, Cardiac Magnetic Resonance Imaging (CMR) technology has become the gold standard for evaluating cardiac structure and function due to its excellent soft tissue contrast and non-radiation characteristics. CMR can provide detailed anatomical images and functional parameters of the heart, which are crucial for the diagnosis and treatment planning of heart diseases. The clinical application of CMR technology is very extensive, including the diagnosis of myocardial infarction, cardiomyopathy, cardiac tumors, heart valve diseases, and congenital heart diseases. It can evaluate the anatomical structure of the heart, functional parameters (such as ventricular volume, ejection fraction), and tissue characteristics (such as myocardial fibrosis). In addition, CMR can also be used for follow-up evaluation after cardiac surgery and interventional treatment.

[0003] The CMR system mainly consists of a magnet, a radiofrequency (RF) system, a gradient system, and a data acquisition system. The magnet generates a stable magnetic field, the RF system is used to excite and receive signals from hydrogen protons, the gradient system is used for spatial encoding, and the data acquisition system is responsible for signal acquisition and preliminary processing. Cardiac imaging usually uses a fast gradient echo sequence (cine-CMR) to capture the movement of the heart during systole and diastole. Image post-processing software is used to manually or automatically segment the myocardial boundary to calculate cardiac volume and functional parameters. The CMR imaging principle is based on the resonance phenomenon of hydrogen atomic nuclei (protons) in a magnetic field. When the human body is placed in a magnetic field, hydrogen protons will align with the magnetic field. By applying a specific RF pulse, the energy level of the protons can be changed. When the RF pulse stops, the protons return to their original energy level and release energy, and the signal released during this process is received and used to generate an image. The image reconstruction algorithm converts the received signal into a two-dimensional or three-dimensional image to provide diagnostic information for doctors.

[0004] Although CMR provides rich information on cardiac structure and function, there are some challenges and limitations in its post-processing and analysis process. Traditional CMR image analysis relies on manual segmentation of the myocardial boundary, which is not only time-consuming but also the results are easily affected by the operator's technical level and experience. Different doctors may have differences in the interpretation of images, resulting in limited consistency and reproducibility of the results. In addition, manual segmentation requires a high level of professional knowledge and long-term concentration, which is a heavy burden for clinicians.

[0005] Existing automated or semi-automated segmentation algorithms still lack accuracy and robustness when dealing with cardiac structures in different pathological states, especially in the ventricular base and apex regions. The anatomical structures in these areas are complex and vary greatly among different individuals and pathological conditions, posing challenges to automated segmentation. The motion and respiratory artifacts of the heart also affect the accuracy of automated segmentation. Most existing CMR analysis methods focus on the analysis of a single modality (such as structural information), while neglecting the integration of functional and tissue property information. The fusion of multi-modal data can provide a more comprehensive cardiac assessment, but requires more complex algorithms and computational resources.

[0006] Cardiac diseases exhibit high heterogeneity in morphology and function. Existing automated segmentation algorithms are often developed based on specific pathological models or datasets and are insufficiently adaptable to unseen pathological variations, limiting their widespread application in clinical practice. In addition, high-quality CMR image analysis typically requires a large amount of computational resources and time. Especially when performing three-dimensional reconstruction and advanced post-processing, high-performance computing devices and a long processing time are required. This is a limiting factor in the clinical environment where rapid diagnosis and treatment decisions are needed. Currently, there is a lack of a standardized process for the post-processing of CMR images, and different hospitals and clinics may use different software and methods, making it difficult to compare and integrate the results. The lack of standardization not only affects the consistency of the results but also limits the conduct of large-scale multi-center studies.

[0007] In summary, CMR technology plays a crucial role in the diagnosis and treatment of cardiac diseases, providing rich anatomical and functional information. However, the efficiency and accuracy of its post-processing and analysis process need to be improved urgently. The subjectivity of manual segmentation, the insufficient accuracy of automated algorithms, the complexity of multi-modal data fusion, the challenges of adapting to pathological variations, the demand for computational resources, and the lack of standardization in the post-processing process are all the main problems faced by current CMR technology. The existence of these problems limits the application of CMR technology in clinical practice and affects the potential of its use in the diagnosis and treatment of heart diseases. Therefore, developing new technologies to solve these problems is of great clinical significance for improving the efficiency and accuracy of CMR image analysis. Summary of the Invention

[0008] To achieve the above and other advantages of the present invention, the first object of the present invention is to provide a method for automatically segmenting cardiac magnetic resonance images based on deep learning, comprising the following steps:

[0009] Obtain cardiac magnetic resonance images;

[0010] Use a deep learning model to identify and segment the myocardium in the cardiac magnetic resonance images;

[0011] Output the myocardial segmentation result.

[0012] Furthermore, the deep learning model includes an encoder and a decoder; wherein,

[0013] The encoder is used to extract key features from the cardiac magnetic resonance images;

[0014] The decoder is used to map the high-level features extracted by the encoder back to the spatial resolution of the original image to achieve accurate left and right ventricular segmentation.

[0015] Furthermore, the deep learning model further includes a dense block, which is placed between the encoder and the decoder, and the dense block is used to achieve feature reuse and network performance improvement.

[0016] Furthermore, the encoder includes a convolutional block, a batch normalization module, a ReLU activation function, a max pooling module, and a skip connection module; wherein,

[0017] The convolutional block is used to capture the initial features of the image;

[0018] The batch normalization module reduces the internal covariate shift by normalizing the input of the layer;

[0019] The ReLU activation function introduces non-linearity by setting all negative values to zero and keeping positive values unchanged, enabling the model to learn more complex feature representations;

[0020] The max pooling module reduces the size of the feature map by selecting the maximum value in each local region;

[0021] The skip connection module is used to directly transfer the feature maps of different levels in the encoder to the corresponding layers of the decoder, enabling the model to utilize richer feature information during the upsampling process.

[0022] Furthermore, as the encoder deepens, the number of filters in the convolutional layers of the convolutional block is gradually increased to enable the model to capture features at different levels of abstraction;

[0023] The number of filters is doubled in each convolutional block to enable the model to effectively combine feature information at different resolutions;

[0024] During the training process of the encoder, an end-to-end training strategy is adopted to enable all layers of the model to be updated simultaneously in the same training process.

[0025] Furthermore, the decoder includes a transposed convolutional layer, a batch normalization module, and a ReLU activation function; wherein,

[0026] The transposed convolution layer is used to learn the mapping from the feature map to the pixel space, thereby recovering the spatial details of the image;

[0027] The batch normalization module reduces internal covariate shift by normalizing the input of the layer;

[0028] The ReLU activation function is used to introduce nonlinear characteristics, so that the model can learn more complex feature representations, thereby maintaining the nonlinear expression ability of the network.

[0029] Furthermore, the number of filters in the transposed convolutional layer is designed to be gradually reduced until it matches the number of target segmentation categories;

[0030] In the upsampling path, the output feature map of the transposed convolutional layer is combined with the skip connection feature map of the corresponding layer in the encoder;

[0031] During the training process of the decoder, the same loss function and optimizer as those of the encoder are adopted; wherein the loss function is a weighted cross entropy loss function, and the optimizer is an Adam optimizer.

[0032] Furthermore, the dense block includes a convolutional layer, a ReLU activation function, and a batch normalization module; wherein,

[0033] The convolutional layer is used to capture subtle features in the image;

[0034] The ReLU activation function introduces nonlinearity, enabling the model to learn more complex feature representations;

[0035] The batch normalization module reduces internal covariate shift by normalizing the input of the layer;

[0036] In the design of the dense block, the feature maps generated by each layer are not only concatenated with the feature maps from the downsampling path, but these feature maps are also passed to all subsequent layers in the upsampling path.

[0037] Furthermore, in the deep learning model training process, weighted cross entropy is used as the loss function to effectively deal with the class imbalance problem. The weighted cross entropy formula is as follows:

[0038]

[0039] Where N represents the number of samples, w i is the weight of the i-th sample, which is used to adjust the contribution of each category to the loss function, y i is the true label of the i-th sample, p i is the probability that the model predicts that the i-th sample is a positive class;

[0040] To optimize the model parameters, the Adam optimizer is adopted, and its update rule is as follows:

[0041]

[0042] Among them, θ represents the model parameters, α is the learning rate, which controls the step size of parameter update; β1 and β2 are the hyperparameters of Adam, which affect the exponential decay rates of the first-order and second-order moment estimates; m t and u t are the first-order and second-order moment estimates respectively, which are used to calculate the adaptive learning rate; ∈ is a tiny constant used to prevent division-by-zero errors;

[0043] Online data augmentation technology is adopted to improve the generalization ability and robustness of the model.

[0044] The second object of the present invention is to provide a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] The automatic segmentation process of the automatic cardiac magnetic resonance image segmentation method provided by the present invention can be completed in less than 1 minute, while the traditional manual segmentation method takes more than 30 minutes on average, and the processing time is shortened by more than 30 times, significantly shortening the processing time.

[0047] By adopting the technical solution provided by the present invention, in actual clinical applications, doctors can obtain diagnostic results faster, improving the diagnosis and treatment efficiency.

[0048] The present invention reduces the interference of human factors through an automated deep learning model, improving the objectivity and consistency of the segmentation results.

[0049] By including CMR image sequences of different pathological groups in the training process, the present invention improves the adaptability of the model to different pathological states.

[0050] The present invention shows good adaptability and stability for three pathological types: HCM, DCM, and VA.

[0051] By optimizing the network structure and training strategy, the present invention reduces the demand for high-performance computing devices.

[0052] The present invention has significant advantages in the field of automatic cardiac magnetic resonance image segmentation, and can provide a more accurate and efficient solution for the diagnosis and treatment of heart diseases.

[0053] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and to implement it according to the contents of the description, the following preferred embodiments of the present invention are described in detail with reference to the accompanying drawings. The specific implementation methods of the present invention are given in detail by the following embodiments and the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0055] Figure 1 This is a flowchart of the automatic segmentation method of cardiac magnetic resonance images based on deep learning;

[0056] Figure 2 Flowchart of the overall algorithm for cardiac magnetic resonance image recognition;

[0057] Figure 3 This is the target detection flow chart;

[0058] Figure 4 Schematic diagram of feature pyramid network;

[0059] Figure 5 is the myocardial segmentation result;

[0060] Figure 6 A schematic diagram of a computer device;

[0061] Figure 7 A schematic diagram of a computer-readable storage medium. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. It should be noted that, without conflict, the embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0063] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0064] The figure numbers in this application are only used to distinguish the various steps in the scheme and are not used to limit the execution order of the various steps. The specific execution order is subject to the description in the specification.

[0065] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs. The terms used in the description of this invention herein are for the purpose of describing specific embodiments only and are not intended to limit the invention.

[0066] In modern medical imaging technology, the automatic segmentation of cardiac magnetic resonance (CMR) images is of great significance for the diagnosis and treatment of heart diseases. CMR images can provide detailed information on the structure and function of the heart, which is crucial for evaluating heart diseases such as cardiomyopathy, cardiac tumors, and congenital heart diseases. However, traditional manual segmentation methods are not only time-consuming but also easily affected by the operator's experience, resulting in limited consistency and accuracy of the results. To overcome these challenges, this invention proposes a deep learning model based on convolutional neural network (CNN) for automatically identifying and segmenting the cardiac cavities in CMR images, with the aim of extracting key imaging features with minimal error.

[0067] As the basis for evaluating cardiac structure, the automatic segmentation of 2D-CMR images is considered the key to improving the diagnostic efficiency and accuracy. The deep learning model of this invention can accurately identify and segment the left ventricle (LV) and right ventricle (RV) from 2D-CMR images through an advanced network architecture, providing clinicians with a faster and more reliable diagnostic tool. In addition, through the reconstruction of three-dimensional CMR (3D-CMR) images, the model can provide more comprehensive cardiac structure information, which is of great value for the in-depth understanding and treatment planning of heart diseases. Although 3D-CMR provides more abundant information, in practical applications, it needs to be compared with 2D-CMR images to ensure the accuracy of the segmentation results and avoid diagnostic errors caused by image quality differences.

[0068] The deep learning model of this invention shows great potential in the field of automatic segmentation of cardiac CMR images. It can not only improve the efficiency and accuracy of segmentation but also provide deeper insights for the diagnosis and treatment of heart diseases. By adopting the solution provided by this invention, clinicians can more effectively plan treatment strategies, thereby providing more personalized medical services for patients.

[0069] Example 1

[0070] A method for automatically segmenting cardiac magnetic resonance images based on deep learning, as Figure 1 - Figure 2 shown, includes the following steps:

[0071] Obtain cardiac magnetic resonance images;

[0072] Use a deep learning model to identify and segment the myocardium in the cardiac magnetic resonance images;

[0073] Output the myocardial segmentation results, such as Figure 5 shown. Among them, the endocardium of the right ventricle is marked yellow in the image, the endocardium of the left ventricle is marked red in the image, and the epicardium of the left ventricle is marked green in the image.

[0074] In some embodiments, as Figure 3 - Figure 4 shown, the deep learning model includes an encoder and a decoder; among them,

[0075] The encoder is used to extract key features from the cardiac magnetic resonance (CMR) images;

[0076] The decoder is used to map the high-level features extracted by the encoder back to the spatial resolution of the original image to achieve accurate left and right ventricle segmentation.

[0077] Further, the encoder includes a convolutional block, a batch normalization module, a ReLU activation function, a max pooling module, and a skip connection module; among them,

[0078] The convolutional block is used to capture the initial features of the image;

[0079] In this embodiment, the starting layer of the encoder is a convolutional block with 48 filters, and this convolutional block is designed to capture the initial features of the image. Each convolutional block consists of two 3×3 convolutional layers. Such small-sized convolutional kernels can effectively capture local features and reduce the number of parameters, thereby improving the generalization ability of the model.

[0080] After each convolutional layer, the Batch Normalization technique is applied.

[0081] The batch normalization module normalizes the input of the layer to reduce the internal covariate shift, which helps to accelerate the training process and improve the stability of the model; immediately afterwards, the ReLU (Rectified Linear Unit) activation function is used;

[0082] The ReLU activation function introduces non-linearity by setting all negative values to zero and keeping positive values unchanged, enabling the model to learn more complex feature representations;

[0083] After the convolutional block, a 2×2 max pooling operation with a stride of 2 is adopted.

[0084] The max pooling module is a downsampling technique that reduces the size of the feature map by selecting the maximum value in each local region. This not only reduces the computational complexity but also helps to extract important features in the image. At the same time, max pooling also plays a certain role in regularization, reducing the sensitivity of the model to small position changes.

[0085] In the design of the encoder, the depth of the feature map is also considered. As the downsampling progresses, the depth of the feature map gradually increases, which helps the model to learn richer features while maintaining spatial information. This increase in depth is achieved by doubling the number of filters in each convolutional block, and such a design enables the model to effectively combine feature information at different resolutions.

[0086] In addition, the skip connections in the encoder are also a key feature of the Res-U-Net architecture. These connections directly transfer the feature maps at different levels in the encoder to the corresponding layers in the decoder, enabling the model to utilize richer feature information during the upsampling process. This design helps the model to achieve more accurate boundary localization in the segmentation task.

[0087] During the training process of the encoder, an end-to-end training strategy is adopted, which means that all layers of the model are updated simultaneously in the same training process. This strategy helps the model to learn the complete mapping from the original image to the final segmentation result. In this way, the model can automatically adjust its parameters to minimize the loss function, thereby improving the accuracy of segmentation.

[0088] Generally speaking, the design of the encoder is the key to the success of the deep learning model of the present invention. Through the carefully designed convolutional blocks, batch normalization, ReLU activation function, max pooling, and skip connections, the model can effectively extract the features crucial for the segmentation of the cardiac chambers from CMR images. These features include not only the geometry of the heart but also the dynamic information related to cardiac function, providing rich diagnostic information for clinicians. Through this innovative encoder design, the model of the present invention demonstrates excellent performance in the automatic segmentation task of cardiac magnetic resonance images.

[0089] At the same time, the decoder part plays a crucial role. It is responsible for mapping the high-level features extracted by the encoder back to the spatial resolution of the original image to achieve accurate left ventricle (LV) and right ventricle (RV) segmentation. The decoder is designed using transposed convolution, which is an upsampling technique that can learn the mapping from the feature map to the pixel space, thereby restoring the spatial details of the image.

[0090] The number of filters in the transposed convolutional layer is designed to gradually decrease until it matches the number of target segmentation classes. This design strategy allows the model to gradually adjust the depth of the feature map during the upsampling process to adapt to different levels of feature fusion. After each transposed convolutional layer, batch normalization and the ReLU activation function are applied. Batch normalization reduces the internal covariate shift by normalizing the input of the layer, which helps to accelerate the training process and improve the stability of the model. The ReLU activation function, on the other hand, introduces non-linearity, enabling the model to learn more complex feature representations and thus maintaining the non-linear expressive power of the network.

[0091] In the upsampling path, the output feature map of the transposed convolutional layer is combined with the skip connection feature map of the corresponding layer in the encoder. This skip connection is a key feature of the Res-U-Net architecture, which allows the model to utilize richer feature information during the upsampling process, thus achieving more accurate segmentation. In this way, the model can gradually restore the spatial details of the image while maintaining high-resolution features, which is crucial for the accurate segmentation of the cardiac chambers.

[0092] In addition, the design of the transposed convolutional layer also takes into account computational efficiency and model complexity. Through the carefully designed number of filters and kernel size, the model can reduce the consumption of computational resources while maintaining high performance. This design makes the model more efficient in practical applications, especially when dealing with high-resolution cardiac magnetic resonance (CMR) images.

[0093] During the training process of the decoder, the same loss function and optimizer as those in the encoder are adopted. The weighted cross-entropy (WCE) loss function helps to handle the class imbalance problem, while the Adam optimizer effectively updates the model parameters through an adaptive learning rate optimization algorithm. This training strategy ensures the performance of the model in the decoder part, enabling it to accurately recover the detailed structure of the cardiac chambers from the features transmitted by the encoder.

[0094] To further improve the efficiency and effectiveness of feature fusion, dense blocks are ingeniously introduced between the encoder and decoder of the Res-U-Net architecture. This innovative design is inspired by the architecture concept of DenseNet, and its core advantage lies in the reuse of features and the improvement of network performance.

[0095] Each dense block contains multiple 3×3 convolutional layers inside, which are the basis for constructing feature extraction. By using small-sized convolutional kernels, the model can capture subtle features in the image, which is crucial for the accurate segmentation of the left ventricle (LV) and right ventricle (RV) in cardiac magnetic resonance (CMR) images. After each convolutional layer, the ReLU activation function is applied, which introduces non-linearity and enables the model to learn more complex feature representations. At the same time, the introduction of batch normalization helps to stabilize the training process, reduce internal covariate shift, thereby accelerating the convergence speed and improving the generalization ability of the model.

[0096] In the design of the dense block, the feature maps generated by each layer are not only connected to the feature maps from the downsampling path, but these feature maps are also passed to all subsequent layers in the upsampling path. This feature transfer mechanism greatly improves the utilization rate of features, enabling the network to make more full use of the information learned in previous levels. Such a design reduces the number of learning parameters because each convolutional layer does not need to learn all features, but can perform further feature extraction based on the output of the previous layer. This not only improves the efficiency of the network, but also enhances the network's ability to learn complex features through feature reuse.

[0097] In the Res-U-Net architecture, the introduction of the dense block also helps to alleviate the vanishing gradient problem, which is one of the common problems in deep learning models. Through dense connections, the gradient can be directly transferred from the last layer to all previous layers, making the training of the network more stable. In addition, the design of the dense block helps to improve the accuracy of the model because it allows for more effective information exchange between different levels, which is particularly important for the accurate segmentation of cardiac cavities.

[0098] The introduction of the dense block enables the model of the present invention to better capture the complexity of the cardiac structure when processing cardiac CMR images.

[0099] During the training process of the deep learning model, weighted cross-entropy (WCE) is used as the loss function to effectively handle the class imbalance problem, especially in cardiac magnetic resonance (CMR) image segmentation, where the number of pixels in different classes may vary significantly. The formula for WCE is as follows:

[0100]

[0101] where N represents the number of samples, w i is the weight of the i-th sample, used to adjust the contribution of each class to the loss function, y i is the true label of the i-th sample, pi is the probability that the model predicts the i-th sample as the positive class; by assigning appropriate weights to each class, WCE helps the model to pay more attention to the less frequent classes during training, thus improving the model's recognition ability for all classes.

[0102] To optimize the model parameters, the Adam optimizer is adopted, and its update rule is as follows:

[0103]

[0104] where θ represents the model parameters, α is the learning rate, which controls the step size of parameter update; β1 and β2 are the hyperparameters of Adam, which affect the exponential decay rates of the first-order and second-order moment estimates; m t and u t are the first-order and second-order moment estimates respectively, which are used to calculate the adaptive learning rate; ∈ is a tiny constant used to prevent division-by-zero errors; the Adam optimizer effectively balances the stability and convergence speed of model training by combining the characteristics of momentum and adaptive learning rate.

[0105] To improve the generalization ability and robustness of the model, this embodiment also implements online data augmentation techniques. These techniques include gamma correction, blurring, and random rotation, which increase the data diversity by simulating different imaging conditions and observation angles. Gamma correction can adjust the contrast of the image, blurring simulates the image blurring caused by patient movement or other factors, and random rotation simulates different imaging angles. These data augmentation techniques not only enrich the training dataset but also help the model learn the consistent features of the cardiac structure under different conditions.

[0106] The present invention directly solves the problem of low efficiency in manually segmenting the myocardial boundary through a deep learning method, especially an improved model based on the Res-U-Net architecture. This method reduces the dependence on manual operations by professionals and significantly improves the speed and efficiency of the segmentation process.

[0107] Aiming at the problem of insufficient accuracy in segmenting the cardiac structure by existing automated segmentation algorithms under different pathological states, the present invention improves the Res-U-Net architecture by introducing dense blocks, enhancing the segmentation accuracy and robustness of the model for complex regions such as the ventricular base and apex.

[0108] By including CMR image sequences of different pathological groups in the training process, the present invention improves the adaptability of the model to different pathological states and solves the problem of insufficient adaptability of existing automated segmentation algorithms to pathological variations.

[0109] Embodiment 2

[0110] A computer device 100, such asFigure 6 As shown, it includes a memory 110, a processor 120, and a computer program 130 stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of an automatic segmentation method for cardiac magnetic resonance images based on deep learning. For a detailed description of the method, reference can be made to the corresponding description in the above method embodiments, which will not be elaborated here.

[0111] Embodiment 3

[0112] A computer-readable storage medium, as Figure 7 shown, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of an automatic segmentation method for cardiac magnetic resonance images based on deep learning. For a detailed description of the method, reference can be made to the corresponding description in the above method embodiments, which will not be elaborated here.

[0113] The number of devices and the scale of processing described here are used to simplify the description of the present invention. Applications, modifications, and variations of the present invention are obvious to those skilled in the art.

[0114] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to specific details and the illustrated and described examples here.

[0115] The device, computer device, non-volatile computer storage medium provided in the embodiments of this specification correspond to the method. Therefore, the device, computer device, and non-volatile computer storage medium also have beneficial technical effects similar to those of the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the corresponding device, computer device, and non-volatile computer storage medium will not be elaborated here.

[0116] Those skilled in the art also know that in addition to implementing the controller in the form of pure computer-readable program code, the method steps can be logically programmed to enable the controller to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers, etc. to achieve the same function. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be regarded as the structures within the hardware component. Or even, the devices for implementing various functions can be regarded as both software units for implementing the method and the structures within the hardware component.

[0117] The systems, devices or units illustrated in the above embodiments may specifically be implemented by computer chips or entities, or by products with certain functions. For the convenience of description, when describing the above devices, they are divided into various units according to their functions and described separately. Of course, when implementing one or more embodiments of this specification, the functions of each unit may be implemented in the same or multiple software and / or hardware.

[0118] Those skilled in the art should understand that the embodiments of this specification can be provided as a method, a system, or a computer program product. Therefore, the embodiments of this specification can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0119] This specification is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0120] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0122] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising said element.

[0123] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program units. Generally, program units include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The specification may also be practiced in a distributed computing environment where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program units may be located in local and remote computer storage media including storage devices.

[0124] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system embodiments, since they are basically similar to method embodiments, they are described relatively simply, and reference can be made to the corresponding parts of the method embodiments for the relevant content.

[0125] The above is only for the embodiments of this specification and is not intended to limit one or more embodiments of this specification. For those skilled in the art, various changes and modifications can be made to one or more embodiments of this specification. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of this specification shall be included within the scope of the claims of one or more embodiments of this specification.

Claims

1. An automatic segmentation method for cardiac magnetic resonance images based on deep learning, characterized in that, It includes the following steps: Obtain cardiac magnetic resonance images; Use a deep learning model to identify and segment the myocardium in the cardiac magnetic resonance images; Output the myocardial segmentation results.

2. The automatic segmentation method of cardiac magnetic resonance images based on deep learning according to claim 1, characterized in that: The deep learning model includes an encoder and a decoder; among them, The encoder is used to extract key features from the cardiac magnetic resonance images; The decoder is used to map the high-level features extracted by the encoder back to the spatial resolution of the original image to achieve accurate left and right ventricle segmentation.

3. The automatic segmentation method of cardiac magnetic resonance images based on deep learning according to claim 2, wherein: The deep learning model further includes a dense block, which is placed between the encoder and the decoder, and the dense block is used to achieve feature reuse and network performance improvement.

4. The automatic segmentation method of cardiac magnetic resonance images based on deep learning according to claim 2, characterized in that: The encoder includes a convolutional block, a batch normalization module, a ReLU activation function, a max pooling module, and a skip connection module; among them, The convolutional block is used to capture the initial features of the image; The batch normalization module reduces the internal covariate shift by normalizing the input of the layer; The ReLU activation function introduces non-linearity by setting all negative values to zero and keeping positive values unchanged, enabling the model to learn more complex feature representations; The max pooling module reduces the size of the feature map by selecting the maximum value in each local region; The skip connection module is used to directly transfer the feature maps of different levels in the encoder to the corresponding layers of the decoder, enabling the model to utilize richer feature information during the upsampling process.

5. The automatic segmentation method of cardiac magnetic resonance images based on deep learning according to claim 4, characterized in that: As the encoder deepens, gradually increase the number of filters in the convolutional layers of the convolutional block so that the model can capture features at different levels of abstraction; Double the number of filters in each convolutional block so that the model can effectively combine feature information at different resolutions; During the training process of the encoder, adopt an end-to-end training strategy so that all layers of the model are updated simultaneously in the same training process.

6. The automatic segmentation method of cardiac magnetic resonance images based on deep learning according to claim 2, wherein: The decoder includes a transposed convolutional layer, a batch normalization module, and a ReLU activation function; among them, The transposed convolutional layer is used to learn the mapping from the feature map to the pixel space, thereby restoring the spatial details of the image; The batch normalization module reduces the internal covariate shift by normalizing the input of the layer; The ReLU activation function is used to introduce non-linearity, enabling the model to learn more complex feature representations, thereby maintaining the non-linear expression ability of the network.

7. The automatic segmentation method of cardiac magnetic resonance images based on deep learning according to claim 6, characterized in that: The number of filters in the transposed convolutional layer is designed to gradually decrease until it matches the number of target segmentation categories; In the upsampling path, the output feature map of the transposed convolutional layer is combined with the skip connection feature map of the corresponding layer in the encoder; During the training process of the decoder, use the same loss function and optimizer as the encoder; among them, the loss function is a weighted cross-entropy loss function, and the optimizer is an Adam optimizer.

8. The automatic segmentation method of cardiac magnetic resonance images based on deep learning according to claim 3, characterized in that: The dense block includes a convolutional layer, a ReLU activation function, and a batch normalization module; among them, The convolutional layer is used to capture the subtle features in the image; The ReLU activation function enables the model to learn more complex feature representations by introducing non-linearity; The batch normalization module normalizes the input of the normalization layer to reduce internal covariate shift; In the design of the dense block, the feature maps generated by each layer are not only concatenated with the feature maps from the downsampling path, but these feature maps are also passed to all subsequent layers in the upsampling path.

9. The automatic segmentation method of cardiac magnetic resonance images based on deep learning according to claim 1, characterized in that: During the training process of the deep learning model, weighted cross-entropy is used as the loss function to effectively handle the class imbalance problem. The weighted cross-entropy formula is as follows: Among them, N represents the number of samples, and w i is the weight of the i-th sample, which is used to adjust the contribution degree of each category to the loss function. y i is the true label of the i-th sample, and p i is the probability that the model predicts the i-th sample as the positive class; To optimize the model parameters, the Adam optimizer is used, and its update rule is as follows: where θ represents the model parameters, α is the learning rate that controls the step size of parameter updates; β1 and β2 are hyperparameters of Adam that affect the exponential decay rates of the first and second moment estimates; m t and u t are the first and second moment estimates respectively, which are used to calculate the adaptive learning rate; ∈ is a small constant used to prevent division-by-zero errors; Online data augmentation technology is adopted to improve the generalization ability and robustness of the model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images