A liver tumor CT image segmentation device, system and storage medium

Abstract

The application discloses a liver tumor CT image segmentation device and system and a storage medium. The application reads a liver tumor CT image, pre-processes the liver tumor CT image to obtain a target liver tumor feature image, and segments the target liver tumor feature image through a preset liver tumor segmentation model based on compressed attention to obtain a liver tumor segmentation image. The liver tumor segmentation model based on compressed attention is embedded with a gradient centering optimizer, and the optimizer is used to standardize a weight space and an output feature space and improve the generalization performance of the model. Compared with the existing liver tumor segmentation image segmentation mode, the liver tumor segmentation model based on compressed attention embedded with the gradient centering optimizer can learn more representative features in a semantic segmentation task and obtain a liver tumor segmentation image with higher segmentation accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image segmentation, in particular to a liver tumor CT image segmentation device, system and storage medium. BACKGROUND

[0002] The liver is an organ mainly responsible for metabolic function in the human body, but it is also the second largest tumor lesion organ after the lung. The 5-year survival rate of early liver cancer patients can reach 40%~70%, and the survival time of advanced liver cancer patients is only about 6 months. Early diagnosis and treatment can effectively reduce the severity and mortality of liver cancer, therefore, CT image segmentation of the liver and liver tumor is a key step in the diagnosis and treatment of liver cancer. Traditional liver segmentation methods include region growing method, threshold segmentation method, etc. The basic idea of region growing method is to aggregate pixels or sub-regions into larger regions according to similar properties. Image threshold segmentation method is a traditional image segmentation method, which is also a region-based image segmentation technology. It has become the most basic and widely used segmentation technology in image segmentation because of its simple implementation, small amount of calculation, fast segmentation speed and stable performance.

[0003] At present, deep learning has made great progress in the field of medical image segmentation. Convolutional neural network is a relatively powerful deep neural network, which has achieved better segmentation results than traditional algorithms in liver and liver tumor segmentation, retinal blood vessel segmentation, brain tumor segmentation, cell image segmentation, etc. Deep learning method does not need to extract features in the image manually. It can automatically extract features from a large amount of data, obtain low-level features of input data, and then perform combination transformation operation to obtain higher-level features, which can effectively reduce the interference of human factors. Although the above methods have achieved good segmentation results in the field of medical image segmentation, there are still some problems in liver tumor segmentation. First, the boundary between liver tumor and surrounding tissue is blurred, and the edge information of the lesion also affects the accuracy of segmentation. Secondly, only CNN is used to segment the target area in most networks, which may weaken the image to some extent, thereby affecting the extraction of subsequent boundary contour information. The above two factors interact with each other, resulting in low accuracy of liver tumor segmentation image.

[0004] The above content is only used to assist in understanding the technical solutions of the present application, and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The main purpose of the present application is to provide a liver tumor CT image segmentation device, system and storage medium, which aims to solve the technical problem of low accuracy of liver tumor segmentation image in the prior art.

[0006] To achieve the above object, the application provides a liver tumor CT image segmentation device, which comprises a memory, a processor and a liver tumor CT image segmentation program stored in the memory and executable on the processor, and is configured to implement the following steps:

[0007] reading a liver tumor CT image, preprocessing the liver tumor CT image, and obtaining a target liver tumor feature image;

[0008] segmenting the target liver tumor feature image through a preset liver tumor segmentation model based on compressed attention, and obtaining a liver tumor segmentation image, wherein the liver tumor segmentation model based on compressed attention is embedded with a gradient centering optimizer, and the optimizer is used to standardize the weight space and the output feature space, and improve the generalization performance of the model.

[0009] Optionally, the preset liver tumor segmentation model based on compressed attention comprises an encoding module, a decoding module and a first convolution module, and the liver tumor CT image segmentation program is configured to implement the following steps:

[0010] extracting features of the target liver tumor feature image through the encoding module and performing down-sampling operation to obtain first to fourth initial feature images and first to fifth contraction path feature images;

[0011] performing feature fusion on the first to fifth contraction path feature images through the decoding module and performing up-sampling operation to obtain fifth to eighth initial feature images and first to fourth expansion path feature images;

[0012] mapping the fourth expansion path feature image to a preset category tree through the first convolution module to obtain a liver tumor segmentation image with foreground and background regions.

[0013] Optionally, the encoding module comprises a first preset convolution submodule, a pooling submodule and a compressed attention submodule, and the liver tumor CT image segmentation program is configured to implement the following steps:

[0014] performing convolution operation and non-linear correction operation on the target liver tumor feature image through the first preset convolution submodule to obtain a first contraction path feature image;

[0015] performing down-sampling operation on the first contraction path feature image through the pooling submodule to obtain a first initial feature image;

[0016] performing convolution operation and non-linear correction operation on the first initial feature image through the first preset convolution submodule to obtain a second contraction path feature image;

[0017] performing down-sampling operation on the second contracted path feature image through the pooling submodule to obtain a second initial feature image;

[0018] performing convolution operation and non-linear correction operation on the second initial feature image through the first preset convolution submodule to obtain a first target contracted path feature image, and performing feature extraction on the first target contracted path feature image through the compression attention submodule by using a compression attention mechanism to obtain a third contracted path feature image;

[0019] performing down-sampling operation on the third contracted path feature image through the pooling submodule to obtain a third initial feature image;

[0020] performing convolution operation and non-linear correction operation on the third initial feature image through the first preset convolution submodule to obtain a fourth contracted path feature image;

[0021] performing down-sampling operation on the fourth contracted path feature image through the pooling submodule to obtain a fourth initial feature image;

[0022] performing convolution operation and non-linear correction operation on the fourth initial feature image through the first preset convolution submodule to obtain a second target contracted path feature image, and performing feature extraction on the second target contracted path feature image through the compression attention submodule by using a compression attention mechanism to obtain a fifth contracted path feature image.

[0023] Optionally, the decoding module comprises a second preset convolution submodule, a skip concatenation submodule and a bilinear interpolation submodule, and the liver tumor CT image segmentation program is configured to implement the following steps:

[0024] performing up-sampling operation on the fifth contracted path feature image through the bilinear interpolation submodule to obtain a fifth initial feature image, performing feature fusion on the fifth initial feature image and the fourth contracted path feature image through the skip concatenation submodule to obtain a first initial expansion path feature image, and performing convolution operation and non-linear correction operation on the first initial expansion path feature image through the second preset convolution submodule to obtain a first expansion path feature image;

[0025] performing up-sampling operation on the first expansion path feature image through the bilinear interpolation submodule to obtain a sixth initial feature image, performing feature fusion on the sixth initial feature image and the third contracted path feature image through the skip concatenation submodule to obtain a second initial expansion path feature image, and performing convolution operation and non-linear correction operation on the second initial expansion path feature image through the second preset convolution submodule to obtain a second expansion path feature image.

[0026] The second expansion path feature image is up-sampled by the bilinear interpolation submodule to obtain a seventh initial feature image, and the seventh initial feature image is fused with a second contraction path feature image by the jump concatenation submodule to obtain a third initial expansion path feature image, and the third initial expansion path feature image is subjected to convolution operation and non-linear correction operation by the second preset convolution submodule to obtain a third expansion path feature image.

[0027] The third expansion path feature image is up-sampled by the bilinear interpolation submodule to obtain an eighth initial feature image, and the eighth initial feature image is fused with a first contraction path feature image by the jump concatenation submodule to obtain a fourth initial expansion path feature image, and the fourth initial expansion path feature image is subjected to convolution operation and non-linear correction operation by the second preset convolution submodule to obtain a fourth expansion path feature image.

[0028] Optionally, the liver tumor CT image segmentation program is configured to implement the following steps:

[0029] A plurality of initial liver tumor segmentation models based on compressed attention are constructed.

[0030] The plurality of initial liver tumor segmentation models based on compressed attention are trained respectively according to a preset liver tumor image segmentation training sample set to obtain a plurality of liver tumor segmentation models based on compressed attention to be verified.

[0031] The plurality of liver tumor segmentation models based on compressed attention to be verified are verified by a preset liver tumor image segmentation test set to obtain a verification result.

[0032] A preset liver tumor segmentation model based on compressed attention is selected from the plurality of liver tumor segmentation models based on compressed attention to be verified according to the verification result.

[0033] Optionally, the liver tumor CT image segmentation program is configured to implement the following steps:

[0034] An initial liver tumor image segmentation training sample set and an initial liver tumor image segmentation test set are constructed.

[0035] The initial liver tumor image segmentation training sample set and the initial liver tumor image segmentation test set are preprocessed respectively to obtain a preset liver tumor image segmentation training sample set and a preset liver tumor image segmentation test set, and the preprocessing includes channel screening, CT value interception, data augmentation, and data normalization.

[0036] Optionally, the liver tumor CT image segmentation program is configured to implement the following steps:

[0037] According to the verification result, a target compression attention-based liver tumor segmentation model is selected from the plurality of compression attention-based liver tumor segmentation models to be verified;

[0038] According to the Dice coefficient, the volume overlap error, the precision, and the recall, the image segmentation performance of the target compression attention-based liver tumor segmentation model is tested.

[0039] When the test result meets the preset condition, the target compression attention-based liver tumor segmentation model is used as a preset compression attention-based liver tumor segmentation model.

[0040] Optionally, the liver tumor CT image segmentation program is configured to implement the following steps:

[0041] The binary cross-entropy function is used as the loss function of the preset compression attention-based liver tumor segmentation model.

[0042] In addition, to achieve the above-mentioned purpose, the application further provides a liver tumor CT image segmentation system, which comprises an image reading module and an image segmentation module.

[0043] The image reading module is used to read a liver tumor CT image.

[0044] The image segmentation module is used to segment the liver tumor CT image by using a preset liver tumor segmentation model, so as to obtain a liver tumor segmentation image.

[0045] In addition, to achieve the above-mentioned purpose, the application further provides a storage medium, which stores a liver tumor CT image segmentation program.

[0046] The liver tumor CT image is read, and the liver tumor CT image is preprocessed to obtain a target liver tumor feature image.

[0047] The target liver tumor feature image is segmented by using a preset compression attention-based liver tumor segmentation model to obtain a liver tumor segmentation image, wherein the compression attention-based liver tumor segmentation model is embedded with a gradient centering optimizer, and the optimizer is used to standardize the weight space and the output feature space, thereby improving the generalization performance of the model.

[0048] The application reads a liver tumor CT image, pre-processes the liver tumor CT image to obtain a target liver tumor feature image, and segments the target liver tumor feature image through a preset liver tumor segmentation model based on compressed attention to obtain a liver tumor segmentation image, wherein the liver tumor segmentation model based on compressed attention is embedded with a gradient centering optimizer, and the optimizer is used to standardize a weight space and an output feature space and improve the generalization performance of the model. Compared with the existing liver tumor segmentation image segmentation mode, the liver tumor segmentation model based on compressed attention embedded with the gradient centering optimizer can learn more representative features in the semantic segmentation task, and the liver tumor segmentation image with higher segmentation accuracy can be obtained. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A liver tumor CT image segmentation device structure schematic diagram of a hardware running environment related to an embodiment scheme of the application.

[0050] Figure 2 A flowchart of a first embodiment of a liver tumor CT image segmentation device of the application.

[0051] Figure 3 A flowchart of a second embodiment of a liver tumor CT image segmentation device of the application.

[0052] Figure 4 A network structure schematic diagram of a compressed attention mechanism model of the second embodiment of the liver tumor CT image segmentation device of the application.

[0053] Figure 5 A flowchart of a third embodiment of a liver tumor CT image segmentation device of the application.

[0054] Figure 6 A structure block diagram of a first embodiment of a liver tumor CT image segmentation system of the application.

[0055] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0056] It should be understood that the specific embodiments described herein are intended to explain the application and are not intended to limit the application.

[0057] REFERENCE Figure 1 , Figure 1 A liver tumor CT image segmentation device structure schematic diagram of a hardware running environment related to an embodiment scheme of the application.

[0058] As Figure 1As shown, the liver tumor CT image segmentation device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk storage device. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0059] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the liver tumor CT image segmentation device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0060] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a liver tumor CT image segmentation program.

[0061] exist Figure 1 In the streaming data writing device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the liver tumor CT image segmentation device of the present invention can be set in the liver tumor CT image segmentation device, and the liver tumor CT image segmentation device calls the streaming data writing program stored in the memory 1005 through the processor 1001.

[0062] This invention provides a liver tumor CT image segmentation device, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the liver tumor CT image segmentation device of the present invention.

[0063] In this embodiment, the liver tumor CT image segmentation device comprises a memory, a processor and a liver tumor CT image segmentation program stored on the memory and executable on the processor, the liver tumor CT image segmentation program is configured to implement the following steps:

[0064] Step S10: reading a liver tumor CT image, preprocessing the liver tumor CT image, and obtaining a target liver tumor feature image.

[0065] It should be noted that the liver tumor CT image is usually in the medical image format nii, which is a three-dimensional data, and can be read by Python SimpleITK. Then the data obtained by reading can be preprocessed by channel filtering, CT value interception, data augmentation and normalization. Under the premise of ensuring the semantic segmentation effect, the medical image format nii file is processed into a 512 pixel*512 pixel image, i.e. the above-mentioned target liver tumor feature image.

[0066] Step S20: segmenting the target liver tumor feature image by a preset liver tumor segmentation model based on compressed attention, obtaining a liver tumor segmentation image, and embedding a gradient centering optimizer in the liver tumor segmentation model based on compressed attention, the optimizer is used to standardize the weight space and output feature space, and improve the generalization performance of the model.

[0067] It should be noted that the preset liver tumor segmentation model based on compressed attention is a liver tumor segmentation model based on compressed attention constructed based on U-net, taking U-Net network as the backbone network, retaining the symmetric structure of the network, mainly including two parts of encoding network and decoding network. The reweighting mechanism considering local and global factors is used to learn more representative features in the semantic segmentation task. In order to improve the problem of inaccurate segmentation of liver tumor boundary, a compressed attention mechanism is introduced in the encoding network, and an embedded gradient centering optimizer is combined to fully utilize the improved hierarchical features.

[0068] It should be understood that the U-net is a new segmentation architecture based on FCN (Fully Convolutional Network), unlike FCN, U-Net is commonly used for medical image segmentation tasks, and adopts a symmetric encoder-decoder structure, also known as a contraction path and an expansion path. The contraction path of the U-Net follows a typical CNN (Convolutional Neural Network) architecture, which increases the number of feature channels by one during downsampling, and the resolution of the feature map is also reduced. The expansion path includes the upsampling of the feature map, which gradually maps the feature map to the original size and reduces the feature channel, restores the resolution, and fuses the features with the corresponding cropped feature maps in the contraction path, so that the deep features and shallow features are combined to capture local and global feature information, and enhance the segmentation effect. In the last layer, a 1x1 convolution is used to map each feature vector to the required class tree. The output segmentation map has two channels for foreground regions (segmented target regions) and background regions.

[0069] It should be understood that the compression attention mechanism can utilize the context to determine which parts of the feature map are activated at the global image level, and through the context features, it can be known which classes can appear in the image at the same time. Through the attention mechanism, the global context features are used to encode these regions as a whole, rather than learning independent reweighting for each part of the image. The present application introduces a compression attention mechanism in the encoding layer of the U-net, which learns more representative features in the semantic segmentation task by considering the reweighting mechanism from local and global aspects. Different from the channel attention mechanism (SE), the attention channel of the compression attention mechanism uses average pooling to downsample the feature map, but unlike the channel attention mechanism, the spatial information is not completely compressed.

[0070] It should be noted that the formula of the gradient centering optimizer of the present application is: , represents the gradient, and the subscript i represents the column vector of the ith column in the gradient matrix, The formula of is: The formula indicates that for the ith column vector of each gradient matrix, the average value of each element in the column vector is calculated Finally, the average value is removed from the column vector, and the above centering operation is performed on each column vector to obtain the centering gradient matrix.

[0071] The embodiment reads a liver tumor CT image, pre-processes the liver tumor CT image to obtain a target liver tumor feature image, and segments the target liver tumor feature image through a preset liver tumor segmentation model based on compressed attention, to obtain a liver tumor segmentation image. The liver tumor segmentation model based on compressed attention embedded with a gradient centering optimizer is used to standardize a weight space and an output feature space, and improve the generalization performance of the model. Compared with the existing liver tumor segmentation image segmentation mode, the liver tumor segmentation model based on compressed attention embedded with the gradient centering optimizer is used to segment the target tumor feature image, so that more representative features in a semantic segmentation task can be learned, and the liver tumor segmentation image with higher segmentation accuracy can be obtained.

[0072] Reference Figure 3 , Figure 3 It is a flowchart of a second embodiment of the liver tumor CT image segmentation device.

[0073] Based on the first embodiment, in the embodiment, the step S20 can include:

[0074] Step S21: feature extraction and down-sampling operation are performed on the target liver tumor feature image through the encoding module, to obtain first to fourth initial feature images and first to fifth shrinkage path feature images.

[0075] It should be noted that the feature extraction and down-sampling operation on the liver tumor feature image are performed to obtain a feature image useful for medical diagnosis. The feature extraction method adopted in the embodiment is a convolution operation and a non-linear correction operation, and the down-sampling is realized through a pooling operation.

[0076] It can be understood that the first to fifth shrinkage path feature images are feature images after the first, second, third, fourth and fifth complete feature extraction operations, respectively, and the first to fourth initial feature images are feature images obtained by performing down-sampling operation on the first to fourth shrinkage path feature images, respectively.

[0077] Further, in order to obtain more accurate first to fourth initial feature images and first to fifth shrinkage path feature images, the step S21 can include:

[0078] Step S211: a convolution operation and a non-linear correction operation are performed on the target liver tumor feature image through the first preset convolution submodule, to obtain a first shrinkage path feature image;

[0079] Step S212: down-sampling operation is performed on the first shrinkage path feature image through the pooling submodule, to obtain a first initial feature image;

[0080] Step S213: performing convolution operation and non-linear correction operation on the first initial feature image by the first preset convolution submodule to obtain a second shrinking path feature image;

[0081] Step S214: performing down-sampling operation on the second shrinking path feature image by the pooling submodule to obtain a second initial feature image;

[0082] Step S215: performing convolution operation and non-linear correction operation on the second initial feature image by the first preset convolution submodule to obtain a first target shrinking path feature image, and performing feature extraction on the first target shrinking path feature image by the compression attention submodule using a compression attention mechanism to obtain a third shrinking path feature image;

[0083] Step S216: performing down-sampling operation on the third shrinking path feature image by the pooling submodule to obtain a third initial feature image;

[0084] Step S217: performing convolution operation and non-linear correction operation on the third initial feature image by the first preset convolution submodule to obtain a fourth shrinking path feature image;

[0085] Step S218: performing down-sampling operation on the fourth shrinking path feature image by the pooling submodule to obtain a fourth initial feature image;

[0086] Step S219: performing convolution operation and non-linear correction operation on the fourth initial feature image by the first preset convolution submodule to obtain a second target shrinking path feature image, and performing feature extraction on the second target shrinking path feature image by the compression attention submodule using a compression attention mechanism to obtain a fifth shrinking path feature image.

[0087] It should be noted that the first preset convolution submodule includes two 3*3 convolutions, each of which has a non-linear correction unit after the convolution. The convolution operation is a linear operation, and most of the data learned by the convolutional neural network in actual application is nonlinear, so a non-linear correction unit is used to explain the linear operation of convolution after convolution. In this embodiment, the two 3*3 convolutions are used for feature extraction of the image.

[0088] It can be understood that the down-sampling operation is also called pooling operation, which refers to dimensionality reduction of the feature image and retaining the most important information. In this embodiment, the most important information refers to the information useful for liver tumor image segmentation. The pooling submodule in this embodiment can be a 2*2 max pooling layer with a step of 2.

[0089] It should be understood that in the present embodiment, the compressed attention sub-module can be a compressed attention mechanism model, referring to Figure 4 , Figure 4 is a network structure diagram of the compressed attention mechanism model of the second embodiment of the liver tumor CT image segmentation device of the present application, and the formula of the compressed attention model is: , and is an up-sampling function for expanding the output of the attention channel, , represents the output of the attention convolution channel , which is parameterized by and the attention convolution layer structure . The non-full compression operation is performed using the average pooling layer , and then the output of the attention channel is up-sampled to match the output of the main convolution channel . In this way, the up-sampled output of the reserved spatial information and the attention channel is used, and the non-local extracted features are aggregated on the main channel.

[0090] Step S22: performing feature fusion and up-sampling operation on the first to fifth contraction path feature images through the decoding module to obtain fifth to eighth initial feature images and first to fourth expansion path feature images;

[0091] It should be noted that the feature extraction method adopted in the present embodiment is through convolution operation and non-linear correction operation, and the up-sampling is realized by bilinear interpolation method. The first to fourth expansion path feature images are respectively the feature images after the first, second, third and fourth complete feature extraction operations, and the fifth to eighth initial feature images are respectively the feature images obtained by up-sampling operation on the first contraction path feature image, the first to third expansion path feature images.

[0092] Further, in order to obtain more accurate liver tumor segmentation images, the step S22 can include:

[0093] Step S221: performing up-sampling operation on the fifth contraction path feature image through the bilinear interpolation submodule to obtain the fifth initial feature image, performing feature fusion on the fifth initial feature image and the fourth contraction path feature image through the jump splicing submodule to obtain the first initial expansion path feature image, and performing convolution operation and non-linear correction operation on the first initial expansion path feature image through the second preset convolution submodule to obtain the first expansion path feature image;

[0094] Step S222: performing an up-sampling operation on the first expansion path feature image by the bilinear interpolation submodule to obtain a sixth initial feature image, performing feature fusion on the sixth initial feature image and the third contraction path feature image by the jump concatenation submodule to obtain a second initial expansion path feature image, performing convolution operation and non-linear correction operation on the second initial expansion path feature image by the second preset convolution submodule to obtain a second expansion path feature image;

[0095] Step S223: performing an up-sampling operation on the second expansion path feature image by the bilinear interpolation submodule to obtain a seventh initial feature image, performing feature fusion on the seventh initial feature image and the second contraction path feature image by the jump concatenation submodule to obtain a third initial expansion path feature image, performing convolution operation and non-linear correction operation on the third initial expansion path feature image by the second preset convolution submodule to obtain a third expansion path feature image;

[0096] Step S224: performing an up-sampling operation on the third expansion path feature image by the bilinear interpolation submodule to obtain an eighth initial feature image, performing feature fusion on the eighth initial feature image and the first contraction path feature image by the jump concatenation submodule to obtain a fourth initial expansion path feature image, performing convolution operation and non-linear correction operation on the fourth initial expansion path feature image by the second preset convolution submodule to obtain a fourth expansion path feature image.

[0097] It should be noted that the preset second convolution submodule has the same function and structure as the above-mentioned preset first convolution submodule, and is used for feature extraction of an image. The bilinear interpolation submodule is used for up-sampling operation, that is, mapping a feature image to an original size and reducing a feature channel to restore resolution. The jump concatenation submodule is used for fusing deep features and shallow features to capture global feature information, that is, splicing deep feature images and shallow feature images.

[0098] Step S23: mapping the fourth expansion path feature image to a preset category tree by the first convolution module to obtain a liver tumor segmentation map having a foreground region and a background region.

[0099] It should be noted that the first convolution submodule can be a 1*1 convolution, which is used for mapping the features of the finally obtained feature image (the fourth expansion path feature image) to a corresponding number of categories to obtain a foreground region and a background region, and to form a liver tumor segmentation image.

[0100] In this embodiment, the feature extraction and down-sampling operation are performed on the target liver tumor feature image by the encoding module to obtain the first to fourth initial feature images and the first to fifth contraction path feature images; the feature fusion and up-sampling operation are performed on the first to fifth contraction path feature images by the decoding module to obtain the fifth to eighth initial feature images and the first to fourth expansion path feature images; and the fourth expansion path feature image is mapped to a preset category tree by the first convolution module to obtain a liver tumor segmentation map having a foreground region and a background region. By considering global and local factors, the feature extraction, the fusion of deep and shallow features, and the accuracy of the extracted features are further improved.

[0101] Referring to Figure 5 , Figure 5 FIG. 3 is a flowchart of a liver tumor CT image segmentation device according to a third embodiment of the present application.

[0102] Based on the above embodiments, before the step S10, the present embodiment further includes:

[0103] Step S01: constructing a plurality of initial liver tumor segmentation models based on compressed attention;

[0104] It should be noted that the initial liver tumor segmentation model based on compressed attention is a liver tumor segmentation model based on compressed attention constructed based on U-net, taking U-Net network as the backbone network, retaining the symmetric structure of the network, mainly including an encoding network and a decoding network two parts, learning more representative features in the semantic segmentation task by considering the reweighting mechanism of local and global factors, in order to improve the problem of inaccurate liver tumor boundary segmentation, a compressed attention mechanism is introduced in the encoding network, and an optimizer with embedded gradient centering is combined to fully utilize the improved hierarchical features. U-net is a new type of segmentation architecture based on FCN, which is different from FCN. U-Net is commonly used for medical image segmentation tasks, adopts a symmetric encoder-decoder structure, also known as a contraction path and an expansion path. The contraction path of U-Net follows the typical CNN architecture, and in the down-sampling process, the number of feature channels is doubled, and the resolution of the feature map is reduced. The expansion path includes feature mapping up-sampling, gradually mapping the feature to the original size and reducing the feature channel, restoring the resolution, and performing feature fusion with the corresponding cropped feature map in the contraction path, so that the deep and shallow features are combined to capture local and global feature information and enhance the segmentation effect. In the last layer, a 1x1 convolution is used to map each feature vector to the required category tree. The output segmentation map has two channels for the foreground region (segmentation target region) and the background region.

[0105] Further, in order to better construct the initial liver tumor segmentation model based on compressed attention, before the step S01, the method further comprises the following steps:

[0106] Step S001: constructing an initial liver tumor image segmentation training sample set and an initial liver tumor image segmentation test set;

[0107] Step S002: respectively pre-processing the initial liver tumor image segmentation training sample set and the initial liver tumor image segmentation test set to obtain a preset liver tumor image segmentation training sample set and a preset liver tumor image segmentation test set, the pre-processing comprising channel screening, CT value interception, data augmentation and data normalization.

[0108] It should be noted that the data set used in this embodiment is the MICCAI 2017 liver tumor segmentation (LiTS) data set and the liver07 public data set. The LiTS 2017 data set contains 131 training samples and 70 test samples. The liver07 data set contains 30 samples. Randomly extract 80% of the training samples from the LiTS 2017 as the initial liver tumor image segmentation training sample set, and the remaining 20% as the initial liver tumor image segmentation test set.

[0109] It can be understood that in order to speed up the construction speed of the model, the data needs to be pre-processed before the training model. The pre-processing of the data includes channel screening, CT value interception, data augmentation and normalization. Under the premise of ensuring the effect of semantic segmentation, the file in medical image format nii is processed into an image with a resolution of 512*512 pixels. And the liver data and liver tumor data in LiTS are divided into liver data set and liver tumor data set two data sets, from the multi-classification task to two binary classification tasks. Finally, the total number of liver images is 19163 (LiTS) + 2750 (liver07) = 21913, and the number of liver tumor images is 7190 (LiTS).

[0110] Step S02: training a plurality of initial liver tumor segmentation models based on compressed attention according to the preset liver tumor image segmentation training sample set, to obtain a plurality of liver tumor segmentation models based on compressed attention to be verified;

[0111] Step S03: verifying the plurality of liver tumor segmentation models based on compressed attention to be verified by the preset liver tumor image segmentation test set to obtain a verification result;

[0112] Step S04: selecting a preset liver tumor segmentation model based on compressed attention from the plurality of liver tumor segmentation models based on compressed attention to be verified according to the verification result.

[0113] It should be noted that the present embodiment is to use the Dice coefficient, the volume overlap error (VOE), the precision (Precision) and the recall (Recall) to measure the segmentation performance. The similarity coefficient (Dice) is used to measure the overlapping part of two samples, and the index range is from 0 to 1, wherein 1 represents a completely overlapping state, and 0 represents a completely non-overlapping state. The Dice represents the ratio of the intersection area between the segmentation result (S) and the label (T), and its expression is as follows:

[0114]

[0115] The volume overlap error (VOE) is to measure the correlation between the true value and the predicted value by using the ratio of the overlapping area of the segmentation region. Similar to Dice, it is represented by the error rate by changing multiplication into subtraction, and its expression is as follows:

[0116]

[0117] The precision (Precision) is represented by the ratio of the number of predicted positive examples to the actual positive examples, and its expression is as follows:

[0118]

[0119] The recall (Recall) is represented by the percentage of samples divided into predicted positive samples to actual positive samples, and its expression is as follows:

[0120]

[0121] In the above formula, TP, TN, FP and FN represent true positive, true negative, false positive and false negative, respectively. True positive represents the number of liver tumor pixel points correctly divided into tumor; true negative represents the number of background pixel points correctly divided into background area; false positive represents the number of background pixel points incorrectly divided into tumor; and false negative represents the number of liver tumor pixel points incorrectly divided into background. The above four indexes are used to select the preset liver tumor segmentation model based on compressed attention

[0122] It should be noted that the loss function used when constructing a plurality of the above initial liver tumor segmentation model based on compressed attention is the binary cross entropy loss function (BCEWithLogitsLoss) and the DiceLoss function to ensure numerical stability. The loss function is the difference between the network output value and the expectation. In the supervised learning process, the expectation is the label of the data. Since the loss function actually contains the network output value, the network output value is calculated by the network parameter. Through the loss function, the parameter can be updated to make the model segmentation more accurate. The binary cross entropy loss function (BCEWithLogitsLoss) and the DiceLoss function formula are as follows:

[0123]

[0124] wherein , p(y) represents the probability of the sample being predicted as positive, y represents the label of the sample, 0 is background, and 1 is liver or liver tumor; N is the total number of input pixels in the image.

[0125]

[0126] wherein, N represents the number of pixel points. K represents the number of categories, which is set to 2, including liver or liver tumor and background. , represents the probability of the pixel point being predicted as a category, that is, the probability value output by the last softmax layer of the network. , represents the label value of pixel point i belonging to category k. It is set .

[0127] The embodiment is constructed by building multiple initial liver tumor segmentation models based on compressed attention; the multiple initial liver tumor segmentation models based on compressed attention are trained respectively according to a preset liver tumor image segmentation training sample set, multiple liver tumor segmentation models based on compressed attention to be verified are obtained; the multiple liver tumor segmentation models based on compressed attention to be verified are verified through a preset liver tumor image segmentation test set, a verification result is obtained; and a preset liver tumor segmentation model based on compressed attention is selected from the multiple liver tumor segmentation models based on compressed attention to be verified according to the verification result. A liver tumor segmentation model with higher segmentation accuracy is obtained, and the effect of liver tumor CT image segmentation is effectively improved.

[0128] Referring to Figure 6 , Figure 6 is the structure block diagram of the first embodiment of the liver tumor CT image segmentation system of the application.

[0129] As Figure 6 shown, the liver tumor CT image segmentation system proposed by the embodiment of the application comprises an image reading module and an image segmentation module.

[0130] The image reading module 10 is used to read the liver tumor CT image.

[0131] The image segmentation module 20 is used to segment the liver tumor CT image through a preset liver tumor segmentation model, and obtain a liver tumor segmentation image.

[0132] The embodiment reads a liver tumor CT image, pre-processes the liver tumor CT image to obtain a target liver tumor feature image, and segments the target liver tumor feature image through a preset liver tumor segmentation model based on compressed attention, to obtain a liver tumor segmentation image. The liver tumor segmentation model based on compressed attention has a gradient centering optimizer embedded therein, and the optimizer is used to standardize a weight space and an output feature space and improve the generalization performance of the model. Compared with the existing liver tumor segmentation image segmentation mode, the liver tumor segmentation model based on compressed attention with the embedded gradient centering optimizer can learn more representative features in a semantic segmentation task, and obtain a liver tumor segmentation image with higher segmentation accuracy.

[0133] It should be noted that the above-described workflow is only illustrative and does not limit the protection scope of the present application. In actual applications, a person skilled in the art can select part or all of the above-described workflow to achieve the purpose of the embodiment according to actual needs, and the present application is not limited in this regard.

[0134] In addition, technical details not described in detail in the embodiment can be referred to the parameter operation method provided by any embodiment of the present application, and will not be described here.

[0135] Based on the first embodiment of the liver tumor CT image segmentation system described above, the second embodiment of the liver tumor CT image segmentation system of the present application is proposed.

[0136] In the embodiment, the image segmentation module 20 is further configured to perform feature extraction and down-sampling on the target liver tumor feature image through the encoding module to obtain first to fourth initial feature images and first to fifth contracted path feature images, perform feature fusion and up-sampling on the first to fifth contracted path feature images through the decoding module to obtain fifth to eighth initial feature images and first to fourth expanded path feature images, and map the fourth expanded path feature image to a preset category tree through the first convolution module to obtain a liver tumor segmentation image having a foreground region and a background region.

[0137] Further, the image segmentation module 20 is further configured to perform convolution operation and non-linear correction operation on the target liver tumor feature image by the first preset convolution submodule to obtain a first shrinkage path feature image; perform down-sampling operation on the first shrinkage path feature image by the pooling submodule to obtain a first initial feature image; perform convolution operation and non-linear correction operation on the first initial feature image by the first preset convolution submodule to obtain a second shrinkage path feature image; perform down-sampling operation on the second shrinkage path feature image by the pooling submodule to obtain a second initial feature image; perform convolution operation and non-linear correction operation on the second initial feature image by the first preset convolution submodule to obtain a first target shrinkage path feature image; perform feature extraction on the first target shrinkage path feature image by the compression attention submodule using a compression attention mechanism to obtain a third shrinkage path feature image; perform down-sampling operation on the third shrinkage path feature image by the pooling submodule to obtain a third initial feature image; perform convolution operation and non-linear correction operation on the third initial feature image by the first preset convolution submodule to obtain a fourth shrinkage path feature image; perform down-sampling operation on the fourth shrinkage path feature image by the pooling submodule to obtain a fourth initial feature image; perform convolution operation and non-linear correction operation on the fourth initial feature image by the first preset convolution submodule to obtain a second target shrinkage path feature image; and perform feature extraction on the second target shrinkage path feature image by the compression attention submodule using a compression attention mechanism to obtain a fifth shrinkage path feature image.

[0138] Further, the image segmentation module 20 is further configured to perform an up-sampling operation on the fifth contracted path feature image by the bilinear interpolation submodule to obtain a fifth initial feature image, perform feature fusion on the fifth initial feature image and the fourth contracted path feature image by the skip concatenation submodule to obtain a first initial expanded path feature image, perform convolution operation and non-linear correction operation on the first initial expanded path feature image by the second preset convolution submodule to obtain a first expanded path feature image; perform an up-sampling operation on the first expanded path feature image by the bilinear interpolation submodule to obtain a sixth initial feature image, perform feature fusion on the sixth initial feature image and the third contracted path feature image by the skip concatenation submodule to obtain a second initial expanded path feature image, and perform convolution operation and non-linear correction operation on the second initial expanded path feature image by the second preset convolution submodule to obtain a second expanded path feature image; perform an up-sampling operation on the second expanded path feature image by the bilinear interpolation submodule to obtain a seventh initial feature image, perform feature fusion on the seventh initial feature image and the second contracted path feature image by the skip concatenation submodule to obtain a third initial expanded path feature image, and perform convolution operation and non-linear correction operation on the third initial expanded path feature image by the second preset convolution submodule to obtain a third expanded path feature image; perform an up-sampling operation on the third expanded path feature image by the bilinear interpolation submodule to obtain an eighth initial feature image, perform feature fusion on the eighth initial feature image and the first contracted path feature image by the skip concatenation submodule to obtain a fourth initial expanded path feature image, and perform convolution operation and non-linear correction operation on the fourth initial expanded path feature image by the second preset convolution submodule to obtain a fourth expanded path feature image.

[0139] Other embodiments or specific implementations of the liver tumor CT image segmentation device can refer to the above-mentioned method embodiments, which will not be described here.

[0140] It should be noted that in this paper, the term "include", "contain" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or system. Without more limitations, the element defined by the statement "including a" does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0141] The above-mentioned embodiment number of the application is only for description, not representing the advantages and disadvantages of the embodiments.

[0142] Those skilled in the art can clearly understand the above-mentioned example method can be realized by means of software and the necessary general hardware platform, of course, also can be through hardware, but in many cases the former is a better implementation. Based on such understanding, the technical solutions of the present application essentially or say the part of the prior art contribution can be embodied in the form of software products, the computer software product is stored in a storage medium (such as read-only memory / random access memory, disk, optical disc), including a number of instructions to make a terminal device (may be a mobile phone, computer, server, air conditioner, or network equipment, etc.) executes the method described in various embodiments of the present application.

[0143] The above is only the preferred embodiment of the present application, not therefore limit the scope of the present application, all the equivalent structure or equivalent flow transformation made by using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the protection scope of the present application.

Claims

1. A liver tumor CT image segmentation apparatus characterized by comprising: The liver tumor CT image segmentation device comprises a memory, a processor and a liver tumor CT image segmentation program stored on the memory and executable on the processor, and the liver tumor CT image segmentation program is configured to implement the following steps: reading a liver tumor CT image, preprocessing the liver tumor CT image, and obtaining a target liver tumor feature image; segmenting the target liver tumor feature image through a preset liver tumor segmentation model based on compressed attention, obtaining a liver tumor segmentation image, and embedding a gradient centering optimizer in the liver tumor segmentation model based on compressed attention, wherein the optimizer is used to standardize the weight space and the output feature space, and improve the generalization performance of the model; The preset liver tumor segmentation model based on compressed attention comprises an encoding module, a decoding module and a first convolution module, and the liver tumor CT image segmentation program is configured to implement the following steps: feature extraction and downsampling operation of the target liver tumor feature image through the encoding module, obtaining first to fourth initial feature images and first to fifth contraction path feature images; feature fusion and upsampling operation of the first to fifth contraction path feature images through the decoding module, obtaining fifth to eighth initial feature images and first to fourth expansion path feature images; mapping the fourth expansion path feature image to a preset category tree through the first convolution module, and obtaining a liver tumor segmentation image with a foreground region and a background region; The formula of the gradient centering optimizer is: , represents a gradient, and subscript i represents a column vector of the ith column in the gradient matrix, The formula of the gradient centering optimizer is: The formula indicates that for each column vector of the gradient matrix, the average value of each element in the column vector is calculated Finally, the average value is removed from the column vector, and the above centering operation is performed on each column vector to obtain the centered gradient matrix.

2. The liver tumor CT image segmentation apparatus of claim 1, wherein The encoding module comprises a first preset convolution submodule, a pooling submodule and a compressed attention submodule, and The liver tumor CT image segmentation program is configured to implement the following steps: convolution operation and nonlinear correction operation of the target liver tumor feature image through the first preset convolution submodule, obtaining a first contraction path feature image; downsampling operation of the first contraction path feature image through the pooling submodule, obtaining a first initial feature image; convolution operation and nonlinear correction operation of the first initial feature image through the first preset convolution submodule, obtaining a second contraction path feature image; downsampling operation of the second contraction path feature image through the pooling submodule, obtaining a second initial feature image; convolution operation and nonlinear correction operation of the second initial feature image through the first preset convolution submodule, obtaining a first target contraction path feature image, and feature extraction of the first target contraction path feature image through the compressed attention submodule using a compressed attention mechanism, obtaining a third contraction path feature image; downsampling operation of the third contraction path feature image through the pooling submodule, obtaining a third initial feature image; convolution operation and nonlinear correction operation of the third initial feature image through the first preset convolution submodule, obtaining a fourth contraction path feature image; downsampling operation of the fourth contraction path feature image through the pooling submodule, obtaining a fourth initial feature image; The fourth initial feature image is subjected to convolution operation and non-linear correction operation by the first preset convolution submodule, so as to obtain a second target shrinking path feature image; and the second target shrinking path feature image is subjected to feature extraction by the compression attention submodule using a compression attention mechanism, so as to obtain a fifth shrinking path feature image.

3. The liver tumor CT image segmentation apparatus of claim 2, wherein The decoding module comprises a second preset convolution submodule, a jump splicing submodule and a bilinear interpolation submodule, and the liver tumor CT image segmentation program is configured to implement the following steps: The fifth shrinking path feature image is subjected to up-sampling operation by the bilinear interpolation submodule, so as to obtain a fifth initial feature image; the fifth initial feature image and the fourth shrinking path feature image are subjected to feature fusion by the jump splicing submodule, so as to obtain a first initial expansion path feature image; the first initial expansion path feature image is subjected to convolution operation and non-linear correction operation by the second preset convolution submodule, so as to obtain a first expansion path feature image; The first expansion path feature image is subjected to up-sampling operation by the bilinear interpolation submodule, so as to obtain a sixth initial feature image; the sixth initial feature image and a third shrinking path feature image are subjected to feature fusion by the jump splicing submodule, so as to obtain a second initial expansion path feature image; the second initial expansion path feature image is subjected to convolution operation and non-linear correction operation by the second preset convolution submodule, so as to obtain a second expansion path feature image; The second expansion path feature image is subjected to up-sampling operation by the bilinear interpolation submodule, so as to obtain a seventh initial feature image; the seventh initial feature image and a second shrinking path feature image are subjected to feature fusion by the jump splicing submodule, so as to obtain a third initial expansion path feature image; the third initial expansion path feature image is subjected to convolution operation and non-linear correction operation by the second preset convolution submodule, so as to obtain a third expansion path feature image; The third expansion path feature image is subjected to up-sampling operation by the bilinear interpolation submodule, so as to obtain an eighth initial feature image; the eighth initial feature image and a first shrinking path feature image are subjected to feature fusion by the jump splicing submodule, so as to obtain a fourth initial expansion path feature image; the fourth initial expansion path feature image is subjected to convolution operation and non-linear correction operation by the second preset convolution submodule, so as to obtain a fourth expansion path feature image.

4. The liver tumor CT image segmentation apparatus of claim 1, wherein The liver tumor CT image segmentation program is configured to implement the following steps: A plurality of initial compression attention-based liver tumor segmentation models are constructed; The plurality of initial compression attention-based liver tumor segmentation models are respectively trained according to a preset liver tumor image segmentation training sample set, so as to obtain a plurality of compression attention-based liver tumor segmentation models to be verified; The plurality of compression attention-based liver tumor segmentation models to be verified are verified by a preset liver tumor image segmentation test set, so as to obtain a verification result; and The plurality of compression attention-based liver tumor segmentation models to be verified are verified by a preset liver tumor image segmentation test set, so as to obtain a verification result. select a preset compression attention-based liver tumor segmentation model from the plurality of compression attention-based liver tumor segmentation models to be verified according to the verification result.

5. The liver tumor CT image segmentation apparatus of claim 4, wherein The liver tumor CT image segmentation program is configured to implement the following steps: An initial liver tumor image segmentation training sample set and an initial liver tumor image segmentation test set are constructed; The initial liver tumor image segmentation training sample set and the initial liver tumor image segmentation test set are respectively preprocessed to obtain a preset liver tumor image segmentation training sample set and a preset liver tumor image segmentation test set, and the preprocessing includes channel screening, CT value interception, data augmentation, and data normalization.

6. The liver tumor CT image segmentation apparatus of claim 4, wherein The liver tumor CT image segmentation program is configured to implement the following steps: According to the verification result, a target compression attention-based liver tumor segmentation model is selected from the plurality of compression attention-based liver tumor segmentation models to be verified; The image segmentation performance of the target compression attention-based liver tumor segmentation model is tested according to the Dice coefficient, the volume overlap error, the precision, and the recall; When the test result meets a preset condition, the target compression attention-based liver tumor segmentation model is taken as a preset compression attention-based liver tumor segmentation model.

7. The liver tumor CT image segmentation apparatus according to any one of claims 1 to 6, characterized by, The liver tumor CT image segmentation program is configured to implement the following steps: A binary cross-entropy function is used as a loss function of the preset compression attention-based liver tumor segmentation model.

8. A liver tumor CT image segmentation system, characterized by, The liver tumor CT image segmentation system comprises the liver tumor CT image segmentation device of any one of claims 1 to 7; The liver tumor CT image segmentation system comprises an image reading module and an image segmentation module. The image reading module is configured to read a liver tumor CT image. The image segmentation module is configured to segment the liver tumor CT image by using a preset liver tumor segmentation model to obtain a liver tumor segmentation image.

9. A storage medium, characterized by The storage medium stores a liver tumor CT image segmentation program, and the storage medium is used to be executed by the liver tumor CT image segmentation device of any one of claims 1 to 7. The liver tumor CT image segmentation program is executed by the processor to implement the following steps: A liver tumor CT image is read, and a target liver tumor feature image is obtained by preprocessing the liver tumor CT image; A liver tumor segmentation image is obtained by segmenting the target liver tumor feature image by using a preset compression attention-based liver tumor segmentation model, and a gradient centering optimizer is embedded in the compression attention-based liver tumor segmentation model, and the optimizer is used to standardize a weight space and an output feature space to improve the generalization performance of the model.

Images