Medical image segmentation model construction method, system, device and storage medium

Abstract

The application is suitable for the technical field of image processing, and provides a medical image segmentation model construction method, system, device and storage medium, the method comprises the following steps: acquiring an image data set for model training; training a medical image segmentation model according to the image data set; observing the medical image segmentation model to perform grid optimization to determine optimal parameters until the medical image segmentation model reaches the best fitting to complete the construction of the medical image segmentation model, the embodiment of the application can obtain more high-resolution information when performing upsampling, and then more perfectly restore the details in the original image, the medical image segmentation model can retain more high-resolution information contained in the high-level feature map, thereby improving the image segmentation precision.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and particularly relates to a method, system, device and storage medium for constructing a medical image segmentation model. Background Technology

[0002] In recent years, deep learning has been widely applied to tasks such as classification, segmentation, tracking, and recognition of natural images. Thanks to the excellent transfer learning capabilities of convolutional neural networks, deep learning research in the field of medical imaging is also increasing. In medical imaging, it is necessary to segment different organs or lesion regions within an image. This task involves automatically or semi-automatically classifying pixels in medical images, thereby segmenting the image into different meaningful regions.

[0003] The segmentation accuracy of an image directly affects a physician's judgment of the lesion. Therefore, this application provides a method, system, device, and storage medium for constructing a medical image segmentation model with high segmentation accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, device, and storage medium for constructing a medical image segmentation model, aiming to solve the technical problems existing in the prior art as identified in the background art.

[0005] The present invention is implemented as follows: a method for constructing a medical image segmentation model, the method comprising the following steps:

[0006] Obtain the image dataset for model training;

[0007] The medical image segmentation model is trained based on the image dataset. The medical image segmentation model is an image semantic segmentation network with an improved ECA module. The improved ECA module is composed of an ECA module and a CBAM attention module. In the medical image segmentation model, the image dataset is processed by the improved ECA module and used as the input for the next layer of downsampling. Feature maps that have not been processed by the improved ECA module are used for skip connections.

[0008] The medical image segmentation model is observed to perform grid optimization to determine the optimal parameters until the medical image segmentation model reaches the best fit, thus completing the construction of the medical image segmentation model.

[0009] Another objective of this invention is to provide a system for constructing a medical image segmentation model, the system comprising:

[0010] The image dataset acquisition module is used to acquire the image dataset for model training.

[0011] The model training module is used to train the medical image segmentation model based on the image dataset. The medical image segmentation model is an image semantic segmentation network with an improved ECA module. The improved ECA module is composed of an ECA module and a CBAM attention module. In the medical image segmentation model, the image dataset is processed by the improved ECA module and used as the input for the next layer of downsampling. Feature maps that have not been processed by the improved ECA module are used for skip connections.

[0012] The optimization and reconstruction module is used to observe the medical image segmentation model to perform grid optimization to determine the optimal parameters until the medical image segmentation model reaches the best fit, thus completing the construction of the medical image segmentation model.

[0013] Another objective of this invention is to provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method for constructing the medical image segmentation model.

[0014] Another objective of this invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for constructing the medical image segmentation model.

[0015] The beneficial effects of this invention are as follows: Since the feature map obtained after each downsampling of the image is combined with the upsampled feature map at the corresponding position, i.e., the skip connection, more high-resolution information can be obtained during upsampling, thereby more completely restoring the details in the original image. Through the fusion of low-level features and high-level features, the medical image segmentation model can retain more high-resolution information contained in the high-level feature map, thereby improving the image segmentation accuracy. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the method for constructing a medical image segmentation model according to an embodiment of the present invention;

[0017] Figure 2 This is a schematic diagram of the structure of the medical image segmentation model provided in an embodiment of the present invention;

[0018] Figure 3 A flowchart for obtaining an image dataset for model training provided in an embodiment of the present invention;

[0019] Figure 4 This is a flowchart of mesh optimization for a medical image segmentation model provided in an embodiment of the present invention;

[0020] Figure 5A structural block diagram of the medical image segmentation model construction system provided in this embodiment of the invention;

[0021] Figure 6 This is a structural block diagram of the image dataset acquisition module provided in an embodiment of the present invention;

[0022] Figure 7 This is a structural block diagram of the optimization and reconstruction module provided in an embodiment of the present invention;

[0023] Figure 8 This is a block diagram of the internal structure of a computer device in one embodiment;

[0024] Figure 9 The images show the segmentation results of each model algorithm for hippocampal slices;

[0025] Figure 10 The images show the segmentation results of each model algorithm for the cell image;

[0026] Figure 11 The images show the segmentation results of each model algorithm for the retina. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] like Figure 1 As shown, in one embodiment, a method for constructing a medical image segmentation model is proposed, the method comprising the following steps:

[0029] Step S100: Obtain the image dataset for model training.

[0030] In this embodiment of the invention, images containing the hippocampus are selected from the selected slices to form an image dataset. These images should be clear in detail, consistent in size, and of excellent quality. The image dataset can be divided into a training set and a test set. The training set is used to train the model, while the test set is used to test the model's accuracy, etc. Preferably, the ratio of the number of images in the training set to the number of images in the test set is 9:1. It is important to note that, in this embodiment of the invention, for the image dataset of the medical image segmentation model, if the subsequent medical image segmentation model is used to segment and recognize hippocampal images, then its training images should be hippocampal-related images.

[0031] Step S200: Train a medical image segmentation model based on the image dataset; the medical image segmentation model is an image semantic segmentation network with an improved ECA module. The improved ECA module is a combination of an ECA module and a CBAM attention module. In the medical image segmentation model, the image dataset is processed by the improved ECA module and used as the input for the next layer of downsampling. Feature maps that have not been processed by the improved ECA module are used for skip connections.

[0032] In this embodiment of the invention, the image semantic segmentation network is the UNet network model, which is built upon the foundation of the UNet network model, but with the addition of an improved ECA module. The medical image segmentation model in this embodiment includes downsampling and upsampling components, or what can be called an encoder-decoder structure. In practical applications, the image to be processed is input into the medical image segmentation model, undergoing downsampling and upsampling processes sequentially, finally yielding the segmentation result.

[0033] In this embodiment of the invention, the feature map obtained after each downsampling layer is concatenated with the upsampled feature map at the corresponding position, i.e., the skip connection. At the same time, the feature map is used as the input for the next downsampling layer after passing through the improved ECA module. After channel-level global average pooling without reducing the dimension, the improved ECA module considers each channel and its k neighbors to capture local cross-channel interaction information. By using one-dimensional convolution to achieve local cross-channel interaction without reducing the dimension, and combining spatial attention, the positional information on the feature map can be better obtained.

[0034] Regarding the improved ECA module in this embodiment of the invention, after the global average pooling layer of the ECA module, the MLP module (FC->ReLU>FC->Sigmoid) is transformed into a one-dimensional convolution. The channel attention module in the CBAM module is basically the same as that in SENet, except that the single average pooling is replaced by the simultaneous use of max pooling and average pooling. This embodiment of the invention performs max pooling and average pooling on the image dataset along the channel dimension, adds the two pooled feature maps along the channel dimension, and then uses a 1x1 convolutional layer. This avoids dimensionality reduction and can effectively capture cross-channel interactions. Simultaneously, the spatial attention module uses a stack of max pooling and average pooling. The improved ECA module of this application combines the advantages of the ECA module in changing the fully connected layer to a one-dimensional convolution to achieve appropriate cross-channel interactions with the advantages of the CBAM module incorporating spatial attention after channel attention and using the addition or stacking of max pooling and average pooling.

[0035] Step S300: The medical image segmentation model is observed to perform grid optimization to determine the optimal parameters until the medical image segmentation model reaches the best fit, thus completing the construction of the medical image segmentation model.

[0036] In this embodiment of the invention, since the medical image segmentation model may have different effects on a specific dataset, grid optimization is required to obtain the optimal parameters for the specific dataset.

[0037] In one embodiment, such as Figure 3 As shown, step S100 may specifically include the following steps:

[0038] Step S101: Obtain the initial image set.

[0039] Step S102: Preprocess the initial image set. The preprocessing includes at least filtering, type conversion, and size revision of the images in the initial image set.

[0040] In this embodiment of the invention, for the obtained initial image set, it is necessary to select images with clear details, consistent size, and excellent quality, and remove unclear images containing blur or artifacts. Moreover, the above selection process can be performed one or more times, and this embodiment of the invention does not impose specific limitations on it.

[0041] Step S103: Perform data augmentation on the preprocessed initial image set to obtain an image dataset for model training.

[0042] In this embodiment of the invention, data augmentation of the preprocessed initial image set can greatly increase the number of samples to ensure training accuracy. In practical applications, data augmentation methods can include affine transformations such as rotation and shifting of the images, or mirror transformations.

[0043] Preferably, this embodiment of the invention provides a method for data augmentation of a preprocessed initial image set, specifically: augmenting the preprocessed initial image set using an elastic transformation method. After augmenting the sample by performing an elastic transformation on the initial image set, the final recognition performance is significantly improved.

[0044] In one embodiment, such as Figure 2 As shown, the medical image segmentation model includes a downsampling feature extraction unit and an upsampling feature extraction unit. The downsampling feature extraction unit consists of two 3×3 convolutional layers and a 2×2 pooling layer. The upsampling feature extraction unit consists of upsampling and feature concatenation performed by bilinear interpolation, as well as convolutional layers.

[0045] To further illustrate the relevant parameters of the model, the parameters of the network model are listed in detail. The network model is divided into a downsampling feature extraction unit and an upsampling feature extraction unit. The parameters of the medical image segmentation model are shown in Table 1.

[0046] Table 1. Parameters of the Medical Image Segmentation Model

[0047]

[0048] In one embodiment, such as Figure 4 As shown, step S300 may specifically include the following steps:

[0049] Step S301: Observe the training process of the medical image segmentation model and gradually determine the parameter range of the model.

[0050] In practical applications, this invention utilizes the PyTorch framework (an open-source Python machine learning library) to create a medical image segmentation model, and uses Tensorboard (a visualization tool for symbolic mathematics systems) to observe the details of the entire model training process, thereby gradually determining the range of model parameters.

[0051] Step S302: Determine the optimal parameter values ​​of the medical image segmentation model step by step according to the parameter range.

[0052] In this embodiment of the invention, since the medical image segmentation model may have different effects on a specific dataset, grid optimization is required. Through the above steps, the range of optimal values ​​of the medical image segmentation model parameters has been determined. Within this range, the optimal values ​​of the model parameters are gradually determined, which can achieve the determination of the optimal parameter values ​​for a specific network model and dataset, thereby producing the optimal effect.

[0053] Step S303: Retrain the medical image segmentation model based on the optimal values ​​of the parameters.

[0054] Step S304: Determine whether the medical image segmentation model has reached the best fit. When the medical image segmentation model reaches the best fit, the construction of the medical image segmentation model is completed.

[0055] In this embodiment of the invention, when retraining the medical image segmentation model, the training details of the model are observed again through Tensorboard to determine whether the medical image segmentation model has reached the best fit. When the best fit is reached, the construction of the medical image segmentation model is completed; when the best fit is not reached, the model parameters are readjusted.

[0056] like Figure 5As shown, in one embodiment, a system for constructing a medical image segmentation model is provided, which may specifically include an image dataset acquisition module 100, a model training module 200, and an optimization and reconstruction module 300. Wherein:

[0057] The image dataset acquisition module 100 is used to acquire an image dataset for model training.

[0058] The model training module 200 is used to train a medical image segmentation model based on the image dataset. The medical image segmentation model is an image semantic segmentation network with an improved ECA module. The improved ECA module is composed of an ECA module and a CBAM attention module. In the medical image segmentation model, the image dataset is processed by the improved ECA module and used as the input for the next layer of downsampling. Feature maps that have not been processed by the improved ECA module are used for skip connections.

[0059] The optimization and reconstruction module 300 is used to observe the medical image segmentation model to perform grid optimization to determine the optimal parameters, and to complete the construction of the medical image segmentation model after the medical image segmentation model reaches the best fit.

[0060] In this embodiment of the invention, since the feature map obtained after each downsampling of the image is spliced ​​with the upsampled feature map at the corresponding position, i.e., the skip connection, more high-resolution information can be obtained during upsampling, thereby more completely restoring the details in the original image. Through the fusion of low-level features and high-level features, the medical image segmentation model can retain more high-resolution information contained in the high-level feature map, thereby improving the image segmentation accuracy.

[0061] like Figure 6 As shown, in one embodiment, the image dataset acquisition module 100 includes an initial image dataset acquisition unit 101, a preprocessing unit 102, and a data augmentation unit 103. Wherein,

[0062] The initial image set acquisition unit 101 is used to acquire an initial image set;

[0063] The preprocessing unit 102 is used to preprocess the initial image set, and the preprocessing includes at least filtering, type conversion and size revision of the images in the initial image set;

[0064] The data augmentation unit 103 is used to augment the preprocessed initial image set to obtain an image dataset for model training.

[0065] In this embodiment of the invention, for the obtained initial image set, it is necessary to select images with clear details, consistent size, and excellent quality, and remove images containing blurriness, artifacts, or other unclear features. This selection process can be performed one or more times, and this embodiment of the invention does not impose specific limitations. Data augmentation of the preprocessed initial image set can greatly increase the number of samples to ensure training accuracy. In practical applications, data augmentation methods can include affine transformations such as rotation and shifting of images, or mirror transformations. In this embodiment of the invention, data augmentation is performed on the preprocessed initial image set using an elastic transformation method.

[0066] like Figure 7 As shown, in one embodiment, the optimization and reconstruction module 300 includes an observation unit 301, an optimal value determination unit 302, a retraining unit 303, and a fitting determination unit 304. Wherein:

[0067] The observation unit 301 is used to observe the training process of the medical image segmentation model and gradually determine the parameter range of the model.

[0068] The optimal value determination unit 302 is used to gradually determine the optimal values ​​of the parameters of the medical image segmentation model according to the parameter range;

[0069] The retraining unit 303 is used to retrain the medical image segmentation model based on the optimal values ​​of the parameters;

[0070] The fitting determination unit 304 is used to determine whether the medical image segmentation model has reached the best fit. When the medical image segmentation model reaches the best fit, the construction of the medical image segmentation model is completed.

[0071] In this embodiment of the invention, the medical image segmentation model is created using the PyTorch framework (an open-source Python machine learning library). Tensorboard (a visualization tool for symbolic mathematics systems) is used to observe the details of the entire model training process, thereby gradually determining the range of model parameters. Within this range, the optimal values ​​of the model parameters are gradually determined. Based on these optimal parameter values, the medical image segmentation model is retrained. During retraining, Tensorboard is used again to observe the training details of the model to determine whether the medical image segmentation model has reached the best fit. When the best fit is reached, the construction of the medical image segmentation model is completed; otherwise, the model parameters are readjusted.

[0072] Figure 8 An internal structural diagram of a computer device in one embodiment is shown. Figure 8As shown, the computer device includes a processor, memory, network interface, input device, and display screen connected via a system bus. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and may also store computer programs. When executed by the processor, these programs enable the processor to implement a method for constructing a medical image segmentation model. The internal memory may also store computer programs, which, when executed by the processor, enable the processor to execute the method for constructing the medical image segmentation model. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0073] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0074] In one embodiment, the medical image segmentation model construction system provided in this application can be implemented as a computer program, which can be implemented in, for example... Figure 8 The computer device shown runs on this system. The computer device's memory can store the various program modules that make up the system for constructing this medical image segmentation model, for example... Figure 5 The image dataset acquisition module 100, model training module 200, and optimization and reconstruction module 300 are shown. The computer program comprised of these modules causes the processor to execute the steps in the methods for constructing medical image segmentation models according to the various embodiments of this application described in this specification.

[0075] For example, Figure 8 The computer equipment shown can be used as follows Figure 5 The image dataset acquisition module 100 in the medical image segmentation model construction system shown executes step S100. The computer device can execute step S200 through the model training module 200. The computer device can execute step S300 through the optimization and reconstruction module 300.

[0076] In one embodiment, a computer device is provided, the computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0077] Step S100: Obtain the image dataset for model training.

[0078] Step S200: Train a medical image segmentation model based on the image dataset; the medical image segmentation model is an image semantic segmentation network with an improved ECA module. The improved ECA module is a combination of an ECA module and a CBAM attention module. In the medical image segmentation model, the image dataset is processed by the improved ECA module and used as the input for the next layer of downsampling. Feature maps that have not been processed by the improved ECA module are used for skip connections.

[0079] Step S300: The medical image segmentation model is observed to perform grid optimization to determine the optimal parameters until the medical image segmentation model reaches the best fit, thus completing the construction of the medical image segmentation model.

[0080] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, causes the processor to perform the following steps:

[0081] Step S100: Obtain the image dataset for model training.

[0082] Step S200: Train a medical image segmentation model based on the image dataset; the medical image segmentation model is an image semantic segmentation network with an improved ECA module. The improved ECA module is a combination of an ECA module and a CBAM attention module. In the medical image segmentation model, the image dataset is processed by the improved ECA module and used as the input for the next layer of downsampling. Feature maps that have not been processed by the improved ECA module are used for skip connections.

[0083] Step S300: The medical image segmentation model is observed to perform grid optimization to determine the optimal parameters until the medical image segmentation model reaches the best fit, thus completing the construction of the medical image segmentation model.

[0084] To evaluate the performance of the medical image segmentation model in this embodiment, three evaluation metrics are selected: Dice coefficient (similarity coefficient), mIOU (mean crossover ratio), and Specificity.

[0085] First, based on the relationship between the predicted image value (B) and the true value (A), the samples can be divided into four categories:

[0086] True Value (TP): Predicted value is 1, actual value is 1; A∩B

[0087] True Negative Value (TN): The predicted value is 0, and the actual value is 0; (A∪B)

[0088] False Positive (FP): The predicted value is 1, and the actual value is 0; B - (A ∩ B)

[0089] False Negative (FN): The predicted value is 0, but the actual value is 1; A - (A ∩ B)

[0090] The similarity coefficient (Dice) measures the overlap between two samples, ranging from 0 to 1, where 1 represents complete overlap and 0 represents no overlap.

[0091]

[0092] The mIOU (mean crossover ratio) is calculated as the ratio of the intersection to the union of two regions in the predicted region and the actual region, i.e.:

[0093]

[0094] This ratio can also be transformed into TP (intersection) divided by the sum of TP, FP, and FN (union), that is:

[0095]

[0096] The specificity value represents the proportion of pixels that are not actually in the target region but are correctly identified as not being in the target region. It signifies the ability to correctly identify pixels that are not in the target region. A specificity value closer to 1 indicates better segmentation. The presence of non-target regions in the segmentation result leads to lower specificity. The formula for calculating specificity is:

[0097]

[0098] Experimental environment

[0099] Specifically, the experimental environment here is shown in Table 2:

[0100] Table 2. Detailed parameters of the online medical image segmentation model.

[0101]

[0102] Experimental results

[0103] To illustrate the effectiveness of the medical image segmentation model in this embodiment, several mainstream methods were used as comparison algorithms. These comparison algorithms include medical image segmentation algorithms based on UNet networks, UNet++ networks, Attention UNet networks, DANet networks, UNet networks with added CBAM modules, and UNet networks with added ECA modules. To visually demonstrate the experimental results, the 75th slice of a patient was selected for display, and the results are as follows: Figure 9 As shown.

[0104] contrast Figure 9 The experimental results show that the regions segmented by the more traditional UNet and UNet++ algorithms differ significantly from the ground truth. They exhibit poor segmentation accuracy at the edges and bends of the hippocampus, failing to accurately recreate the original shape of the hippocampus. Even after adding different attention mechanisms, the segmentation results remain unsatisfactory. The results indicate that Attention UNet pays more attention to bends but does not accurately segment the details at the tail. DANet performs even worse than UNet, resulting in its performance index falling below the baseline. UNet with CBAM and UNet with ECA focus more on the overall shape, leading to redundancy at bends and details. The method proposed in this invention is the most effective among these approaches, successfully segmenting the overall contour while also recreating the details of the hippocampus.

[0105] Experimental results demonstrate that the method proposed in this invention achieves better segmentation performance than the original U-net network, thus meeting the expected optimization goals. Figure 9 As can be seen, the prediction results obtained by segmentation using the method proposed in this embodiment of the invention are very close to the results manually annotated by doctors in the dataset in terms of overall shape and position. Moreover, the boundaries are smoother than those of manual annotation, and the local image information is also displayed in more detail and clearly. The lines are relatively realistic and accurate, and can basically achieve the same effect as manual annotation by doctors, which can play a certain auxiliary role in medical diagnosis.

[0106] The images clearly show that the image generated by the method proposed in this embodiment of the invention produces better results. Next, we will compare the experimental effects of several algorithms from a numerical perspective, using the Dice coefficient and IOU as evaluation metrics (all results are accurate to 0.01% for easy observation), as shown in Tables 2 and 3 respectively:

[0107] Table 2. Dice coefficients of hippocampal segmentation results.

[0108]

[0109]

[0110] Table 3. IOU Table of Hippocampal Segmentation Results

[0111]

[0112] The DSB2018 cell dataset was then segmented using the method described above, again employing the method of this invention and several mainstream methods for comparison. The comparison algorithms mainly include medical image segmentation algorithms based on UNet networks, medical image segmentation algorithms based on UNet++ networks, medical image segmentation algorithms based on Attention UNet networks, medical image segmentation algorithms based on UNet networks with added CBAM modules, and medical image segmentation algorithms based on UNet networks with added ECA modules. The segmentation results of the above methods are shown below. Figure 10 As shown.

[0113] from Figure 10 As can be seen, the segmentation results of UNet and UNet++ algorithms differ significantly from the ground truth in dark regions. Attention UNet, UNet with CBAM, and UNet with ECA have larger segmentation errors in regions where multiple cells overlap. The method of the present invention is the best among these, as it can effectively segment cells in poorly lit areas and handle cell clustering well.

[0114] from Figure 10 As can be clearly seen, the image generated by the method of this embodiment is better. Next, we will compare the results using two evaluation metrics: Dice coefficient and IOU (all results are accurate to 0.01% for easy observation), as shown in Tables 4 and 5 respectively.

[0115] Table 4. Dice coefficients for cell dataset segmentation results.

[0116]

[0117] Table 5. IOU Table of Cell Dataset Segmentation Results

[0118]

[0119]

[0120] Taking the segmentation results on the DRIVE eye retina dataset as an example again, the method of this invention and several mainstream methods are used as comparison algorithms, which are the same as above. To intuitively demonstrate the experimental results, an image of a patient is selected for display. The segmentation results of the above methods are as follows: Figure 11 As shown.

[0121] from Figure 11 In the first part, we can notice that the segmentation results of UNet and UNet++ algorithms in the lower left corner are broken. For the other three algorithms that have added attention mechanism, the broken parts are smaller, but they still exist. Looking at the area where multiple branches intersect in the middle, several comparative algorithms either fail to segment or missegment. The method of the present invention is the best in this case, which can not only segment the overall outline well, but also restore the ends of branches and intersections well.

[0122] from Figure 11 As can be clearly seen, the image generated by the method of this embodiment is better. The results are compared using two evaluation metrics: Dice coefficient and IOU (all results are accurate to 0.01% for easy observation), as shown in Tables 6 and 7 respectively.

[0123] Table 6. Dice coefficients for segmentation results of the ocular retinal dataset.

[0124] UNet UNet++ AttntionGated ECA CBAM This application method 1 72.00％ 70.25％ 72.96％ 71.67％ 66.87％ 77.11％ 2 67.10％ 64.67％ 67.99％ 69.12％ 62.60％ 74.71％ 3 72.68％ 72.73％ 73.45％ 74.02％ 69.71％ 78.70％ 4 73.31％ 74.64％ 72.47％ 72.99％ 73.71％ 79.70％ 5 72.78％ 74.23％ 73.37％ 73.27％ 70.24％ 77.86％ mean 71.57％ 71.30％ 72.05％ 72.22％ 68.63％ 77.62％ variance 2.54％ 4.09％ 2.30％ 1.93％ 4.15％ 1.89％

[0125] Table 4. IOU table of segmentation results for the ocular retinal dataset.

[0126] UNet UNet++ AttntionGated ECA CBAM This application method 1 56.25％ 54.14％ 57.43％ 55.85％ 50.23％ 62.74％ 2 50.49％ 47.78％ 51.51％ 52.82％ 45.56％ 59.63％ 3 57.09％ 57.14％ 58.04％ 58.76％ 53.50％ 64.88％ 4 57.86％ 59.54％ 56.82％ 57.47％ 58.36％ 66.25％ 5 57.20％ 59.01％ 57.94％ 57.82％ 54.13％ 63.75％ mean 55.78％ 55.52％ 56.35％ 56.54％ 52.36％ 63.45％ variance 3.01％ 4.82％ 2.75％ 2.33％ 4.77％ 2.51％

[0127] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0128] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0129] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0131] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of constructing a medical image segmentation model, characterized by, The method comprises the following steps: acquiring an image dataset for model training; training a medical image segmentation model according to the image dataset; the medical image segmentation model is an image semantic segmentation network mixed with an improved ECA module, the improved ECA module is coupled by an ECA module and a CBAM attention module, wherein in the medical image segmentation model, the image dataset is taken as the input of the next layer after being processed by the improved ECA module, and the feature map not processed by the improved ECA module is connected by jumping; observing the training process of the medical image segmentation model to gradually determine the parameter range of the model; gradually determining the optimal value of the parameters of the medical image segmentation model according to the parameter range; retraining the medical image segmentation model according to the optimal value of the parameters; determining whether the medical image segmentation model reaches the best fitting, and when the medical image segmentation model reaches the best fitting, the construction of the medical image segmentation model is completed.

2. The method of claim 1, wherein, The step of acquiring the image dataset for model training specifically comprises: acquiring an initial image set; preprocessing the initial image set, the preprocessing at least comprising screening, type conversion and size revision of the images in the initial image set; performing data augmentation on the preprocessed initial image set to obtain the image dataset for model training.

3. The method of claim 1, wherein, The medical image segmentation model comprises a down-sampling feature extraction part and an up-sampling feature extraction part, wherein the down-sampling feature extraction part is composed of two 3*3 convolution layers and one 2*2 pooling layer; and the up-sampling feature extraction part is composed of up-sampling by bilinear interpolation, feature splicing and convolution layers.

4. The method of claim 2, wherein, The data augmentation method for the preprocessed initial image set is to perform data augmentation on the preprocessed initial image set by an elastic transformation method.

5. A system for constructing a medical image segmentation model, characterized by, The system comprises: an image dataset acquisition module for acquiring an image dataset for model training; a model training module for training a medical image segmentation model according to the image dataset; the medical image segmentation model is an image semantic segmentation network mixed with an improved ECA module, the improved ECA module is coupled by an ECA module and a CBAM attention module, wherein in the medical image segmentation model, the image dataset is taken as the input of the next layer after being processed by the improved ECA module, and the feature map not processed by the improved ECA module is connected by jumping; an optimization and reconstruction module for observing the training process of the medical image segmentation model to gradually determine the parameter range of the model; gradually determining the optimal value of the parameters of the medical image segmentation model according to the parameter range; retraining the medical image segmentation model according to the optimal value of the parameters; determining whether the medical image segmentation model reaches the best fitting, and when the medical image segmentation model reaches the best fitting, the construction of the medical image segmentation model is completed. 6.The system of claim 5, wherein, The image dataset acquisition module comprises: an initial image set acquisition unit for acquiring an initial image set; A preprocessing unit is configured to preprocess the initial image set, and the preprocessing at least includes screening, type conversion and size revision of images in the initial image set; A data augmentation unit is configured to perform data augmentation on the preprocessed initial image set to obtain an image data set for model training. 7.The system of claim 5, wherein, The optimization reconstruction module comprises: An observation unit is configured to observe the training process of the medical image segmentation model, and gradually determine the parameter range of the model; An optimal value determination unit is configured to gradually determine the optimal value of the parameters of the medical image segmentation model according to the parameter range; A retraining unit is configured to retrain the medical image segmentation model according to the optimal value of the parameters; A fitting determination unit is configured to determine whether the medical image segmentation model reaches the best fitting, and when the medical image segmentation model reaches the best fitting, the construction of the medical image segmentation model is completed.

8. A computer device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the construction method of the medical image segmentation model in any one of claims 1 to 4.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the construction method of the medical image segmentation model in any one of claims 1 to 4.

Images