Multi-task pneumothorax medical image semantic segmentation model system and method based on Unet

Abstract

The present invention discloses a multi-task pneumothorax medical image semantic segmentation model system and method based on Unet, and the multi-task pneumothorax medical image semantic segmentation model method based on Unet includes step S1: obtaining a pneumothorax CT image and encoding the pneumothorax CT image through a first model and a second model, thereby obtaining a binary classification of the image. The multi-task pneumothorax medical image semantic segmentation model system and method based on Unet disclosed by the present invention improve the traditional Unet semantic segmentation, adopt Resnet34 as a framework in the encoding stage, introduce the SCSE module to correct the image feature information, obtain the global information of the image from both the spatial and channel directions, and adopt the multi-task strategy of "classification" and "segmentation" fusion to learn the loss function, so as to perform semantic segmentation on the pneumothorax medical image.

Description

Technical Field

[0001] The present invention belongs to the technical field of pneumothorax image analysis, and specifically relates to a Unet-based multi-task pneumothorax medical image semantic segmentation model system and a Unet-based multi-task pneumothorax medical image semantic segmentation model method. Background Art

[0002] Pneumothorax refers to the state of gas accumulation caused by the entry of gas into the pleural cavity. Pneumothorax is a great danger to human life. The general diagnosis accuracy rate is 50%, which poses a great threat to humans. The usual method is to diagnose through X-rays, but X-rays must be judged by doctors with rich experience. Currently, with the continuous development of deep learning technology, deep learning models can quickly identify whether a patient has pneumothorax and can determine the location of the lesion. In addition, predicting whether a patient has pneumothorax through network models will greatly reduce the burden on doctors and provide effective auxiliary means.

[0003] The first problem encountered in the process of using the traditional Unet model for semantic segmentation is the low recall rate of positive samples. There are pneumothorax lesions in the X-ray film, but they are not identified during the prediction process of the semantic segmentation model. The second problem is whether it is possible to design a better network structure based on the original one to further improve the accuracy of semantic segmentation.

[0004] Therefore, further improvements are made to the above problems. Summary of the invention

[0005] The main purpose of the present invention is to provide a multi-task pneumothorax medical image semantic segmentation model system and method based on Unet, to improve the traditional Unet semantic segmentation, to use Resnet34 as the framework in the encoding stage, to introduce the SCSE module to correct the image feature information, to obtain the global information of the image from both the spatial and channel directions, and to use the multi-task strategy of "classification" and "segmentation" fusion to learn the loss function, to perform semantic segmentation on pneumothorax medical images, and to show that the semantic segmentation method has certain advantages over the traditional Unet semantic segmentation method in segmentation accuracy.

[0006] To achieve the above objectives, the present invention provides a multi-task pneumothorax medical image semantic segmentation model method based on Unet, which is used to perform semantic segmentation on pneumothorax medical images, comprising the following steps:

[0007] Step S1: Acquire a pneumothorax CT image and encode the pneumothorax CT image through a first model (Unet model with pre-trained parameters based on Resnet34 framework) and a second model (SCSE module structure) to obtain a binary classification of the image;

[0008] Step S2: During the decoding process, the feature information of each layer in the encoding process stage is fused and up-sampled, and the up-sampled information of each layer is continuously fused in combination with the second model to obtain a semantic segmentation image;

[0009] Step S3: The loss function of the binary classification of the image and the loss function of the semantic segmentation image are fused, and the loss function is learned using a multi-task strategy of classification and segmentation fusion. The classification of local pixels and global pixels is fused, and the global loss function is obtained after adjusting the weight relationship, thereby realizing the semantic segmentation of the pneumothorax CT image.

[0010] As a further preferred technical solution of the above technical solution, step S1 is specifically implemented as follows:

[0011] Step S1.1: The first model is used as a basic feature extraction network for pneumothorax CT images and basic feature extraction is performed by adding a residual network to the network;

[0012] Step S1.2: The second model corrects the extracted basic features through the space compression module and the channel compression module, thereby obtaining the corrected data of each layer corresponding to the first model;

[0013] Step S1.3: Encode the pneumothorax CT image using the first model and the second model to obtain a binary classification of the image.

[0014] As a further preferred technical solution of the above technical solution, step S2 is specifically implemented as the following steps:

[0015] Step S2.1: the first module (64, 32, 32) in the encoding process upsamples the module where the maximum pooling process in the encoding process is located (Bn1+Relu+Maxpool) to obtain a first feature map and performs a first correction process on the first feature map through a second model to obtain first corrected data;

[0016] Step S2.2: the second module (64, 64, 64) in the encoding process upsamples the first corrected data obtained by the first module and fuses the corrected data corresponding to the module where the fifth layer of the first model in the encoding process is located to obtain a second feature map and performs a second correction process on the second feature map through the second model to obtain second corrected data;

[0017] Step S2.3: the third module (64, 128, 128) in the encoding process upsamples the second corrected data obtained by the second module and fuses the corrected data corresponding to the module where the fourth layer of the first model in the encoding process is located to obtain a third feature map and performs a third correction process on the third feature map through the second model to obtain third corrected data;

[0018] Step S2.4: the fourth module (64, 256, 256) in the encoding process upsamples the third corrected data obtained by the third module and fuses the corrected data corresponding to the module where the third layer of the first model in the encoding process is located to obtain a fourth feature map and performs a fourth correction process on the fourth feature map through the second model to obtain fourth corrected data;

[0019] Step S2.5: the fifth module (64, 512, 512) in the encoding process upsamples the fourth corrected data obtained by the fourth module and fuses the corrected data corresponding to the module where the second layer of the first model in the encoding process is located to obtain a fifth feature map and performs a fifth correction process on the fifth feature map through the second model to obtain fifth corrected data;

[0020] Step S2.6: The corresponding corrected data obtained by each module in the encoding process are fused to complete the encoding process and obtain a semantic segmentation image.

[0021] As a further preferred technical solution of the above technical solution, the correction process performed by the space compression module and the channel compression module of the second model in step S1 and step S2 is:

[0022] Assume that the input pneumothorax CT image is X∈R H×W×C , assuming that the basic features obtained in the encoding process or the feature map obtained in the decoding process (collectively referred to as feature map for ease of description) is U∈R H×W×C' , H and W represent the spatial height and width of the image, C and C' represent the number of input and output channels. After convolution and nonlinear transformation operations, U contains the spatial and channel information in X. After each encoding or decoding process, the second model is introduced to correct U to U'.

[0023] For the spatial compression module, assume that the input feature map U = [u1,u2,…u C ], where u i ∈R H×W Representing the feature map of each channel, through the global average pooling (GAP) operation, we get the vector z∈R 1×1×C , the k-th channel feature map is transformed as follows:

[0024]

[0025] This transformation introduces global spatial feature information into z, and then z'=W 1 (δ(W 2 z)), we get z', where are two fully connected networks, δ(·) is the ReLU operation, and then through the sigmoid transformation, the normalized correction coefficient σ(z')∈[0,1] is obtained. Finally, the correction coefficient is multiplied by the input feature map U to obtain U′ cSE ,

[0026] U′ cSE =[σ(z′ 1 ) 1 ,σ(z′ 2 ) 2 ,…σ(z′ C ) C ];

[0027] (Through this spatial compression module, the information in unimportant channels will be reduced and suppressed, while the information in important channels will remain almost unchanged, thus being enhanced in disguise;

[0028] The above-mentioned index for measuring the importance of the channel is obtained by compressing the information in space, and the index for measuring the importance of the spatial position can also be obtained by compressing the channel information)

[0029] For the channel compression module, assume that the input feature map U = [u1,u2,…u C ], through conversion we get U = [u 1,1 ,u 1 ,1 …u i,j …u H,W ], where u i,j ∈R 1×1×C Represents the corresponding spatial position (i, j), through q = W sq *U convolution operation realizes spatial squeezing operation, and then through sigmoid transformation, the normalized correction coefficient σ(q)∈[0,1] is obtained. Finally, the correction coefficient is multiplied by the input feature map U to obtain U′ sSE ,

[0030] U′ sSE =[σ(q 1,1 ) 1,1 ,σ(q 1,2 ) 1,1 ,…σ(q h,w ) H,W ];

[0031] Finally, the fusion information of the two is obtained by adding them together, and the importance of spatial information and channel information is weighted, and the output is U′ sCSE =U′ cSE +U′ sSE .

[0032] To achieve the above objectives, the present invention also provides a Unet-based multi-task pneumothorax medical image semantic segmentation model system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the Unet-based multi-task pneumothorax medical image semantic segmentation model method are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the Unet-based multi-task pneumothorax medical image semantic segmentation model system and method of the present invention.

[0034] Figure 2 It is a structural diagram of the second model of the Unet-based multi-task pneumothorax medical image semantic segmentation model system and method of the present invention.

[0035] Figure 3 This is a multi-task learning structure diagram of the Unet-based multi-task pneumothorax medical image semantic segmentation model system and method of the present invention. DETAILED DESCRIPTION

[0036] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be applied to other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.

[0037] In the preferred embodiment of the present invention, those skilled in the art should note that the pneumothorax CT images and the like involved in the present invention may be regarded as prior art.

[0038] Preferred embodiments.

[0039] The present invention discloses a multi-task pneumothorax medical image semantic segmentation model method based on Unet, which is used for semantic segmentation of pneumothorax medical images, comprising the following steps:

[0040] Step S1: Acquire a pneumothorax CT image and encode the pneumothorax CT image through a first model (Unet model with pre-trained parameters based on Resnet34 framework) and a second model (SCSE module structure) to obtain a binary classification of the image;

[0041] Step S2: During the decoding process, the feature information of each layer in the encoding process stage is fused and up-sampled, and the up-sampled information of each layer is continuously fused in combination with the second model to obtain a semantic segmentation image;

[0042] Step S3: The loss function of the binary classification of the image and the loss function of the semantic segmentation image are fused, and the loss function is learned using a multi-task strategy of classification and segmentation fusion. The classification of local pixels and global pixels is fused, and the global loss function is obtained after adjusting the weight relationship, thereby realizing the semantic segmentation of the pneumothorax CT image.

[0043] Specifically, step S1 is implemented as follows:

[0044] Step S1.1: The first model is used as a basic feature extraction network for pneumothorax CT images and basic feature extraction is performed by adding a residual network to the network;

[0045] Step S1.2: The second model corrects the extracted basic features through the space compression module and the channel compression module, thereby obtaining the corrected data of each layer corresponding to the first model;

[0046] Step S1.3: Encode the pneumothorax CT image using the first model and the second model to obtain a binary classification of the image.

[0047] More specifically, step S2 is implemented as follows:

[0048] Step S2.1: the first module (64, 32, 32) in the encoding process upsamples the module where the maximum pooling process in the encoding process is located (Bn1+Relu+Maxpool) to obtain a first feature map and performs a first correction process on the first feature map through a second model to obtain first corrected data;

[0049] Step S2.2: the second module (64, 64, 64) in the encoding process upsamples the first corrected data obtained by the first module and fuses the corrected data corresponding to the module where the fifth layer of the first model in the encoding process is located to obtain a second feature map and performs a second correction process on the second feature map through the second model to obtain second corrected data;

[0050] Step S2.3: the third module (64, 128, 128) in the encoding process upsamples the second corrected data obtained by the second module and fuses the corrected data corresponding to the module where the fourth layer of the first model in the encoding process is located to obtain a third feature map and performs a third correction process on the third feature map through the second model to obtain third corrected data;

[0051] Step S2.4: the fourth module (64, 256, 256) in the encoding process upsamples the third corrected data obtained by the third module and fuses the corrected data corresponding to the module where the third layer of the first model in the encoding process is located to obtain a fourth feature map and performs a fourth correction process on the fourth feature map through the second model to obtain fourth corrected data;

[0052] Step S2.5: the fifth module (64, 512, 512) in the encoding process upsamples the fourth corrected data obtained by the fourth module and fuses the corrected data corresponding to the module where the second layer of the first model in the encoding process is located to obtain a fifth feature map and performs a fifth correction process on the fifth feature map through the second model to obtain fifth corrected data;

[0053] Step S2.6: The corresponding corrected data obtained by each module in the encoding process are fused to complete the encoding process and obtain a semantic segmentation image.

[0054] Furthermore, the correction process performed by the space compression module and the channel compression module of the second model in step S1 and step S2 is:

[0055] Assume that the input pneumothorax CT image is X∈R H×W×C , assuming that the basic features obtained in the encoding process or the feature map obtained in the decoding process (collectively referred to as feature map for ease of description) is U∈R H×W×C' , H and W represent the spatial height and width of the image, C and C' represent the number of input and output channels. After convolution and nonlinear transformation operations, U contains the spatial and channel information in X. After each encoding or decoding process, the second model is introduced to correct U to U'.

[0056] For the spatial compression module, assume that the input feature map U = [u1,u2,…u C ], where u i ∈R H×W Representing the feature map of each channel, through the global average pooling (GAP) operation, we get the vector z∈R 1×1×C , the k-th channel feature map is transformed as follows:

[0057]

[0058] This transformation introduces global spatial feature information into z, and then z'=W 1 (δ(W 2 z)), we get z', where are two fully connected networks, δ(·) is the ReLU operation, and then through the sigmoid transformation, the normalized correction coefficient σ(z')∈[0,1] is obtained. Finally, the correction coefficient is multiplied by the input feature map U to obtain U′ cSE ,

[0059] U′ cSE =[σ(z′ 1 ) 1 ,σ(z′ 2 ) 2 ,…σ(z′ C ) C ];

[0060] (Through this spatial compression module, the information in unimportant channels will be reduced and suppressed, while the information in important channels will remain almost unchanged, thus being enhanced in disguise;

[0061] The above-mentioned index for measuring the importance of the channel is obtained by compressing the information in space, and the index for measuring the importance of the spatial position can also be obtained by compressing the channel information)

[0062] For the channel compression module, assume that the input feature map U = [u1,u2,…u C ], through conversion we get U = [u 1,1 ,u 1 ,1 …u i,j …u H,W ], where u i,j ∈R 1×1×C Represents the corresponding spatial position (i, j), through q = W sq *U convolution operation realizes spatial squeezing operation, and then through sigmoid transformation, the normalized correction coefficient σ(q)∈[0,1] is obtained. Finally, the correction coefficient is multiplied by the input feature map U to obtain U′ sSE ,

[0063] U′ sSE =[σ(q 1,1 ) 1,1 ,σ(q 1,2 ) 1,1 ,…σ(q h,w ) H,W ];

[0064] Finally, the fusion information of the two is obtained by adding them together, and the importance of spatial information and channel information is weighted, and the output is U′ sCSE =U′ cSE +U′ sSE .

[0065] The present invention also discloses a Unet-based multi-task pneumothorax medical image semantic segmentation model system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the Unet-based multi-task pneumothorax medical image semantic segmentation model method are implemented.

[0066] Preferably, for the first problem in the background technology, the reason is that Unet is a semantic segmentation model for point-to-point pixel level, and it is not possible to judge whether there is a lesion from a global perspective. Therefore, for small lesions, simply using the Unet network for segmentation cannot obtain good prediction results. The present invention has made improvements on the basis of the classic Unet network, introducing "classification" and "segmentation" multi-task strategies for learning, by globally performing binary classification learning of the presence or absence of lesions on different images, and then performing semantic segmentation on a single pixel on each image, and fusing the loss functions of the two. This method can improve the accuracy of positive sample prediction. For the second problem, the traditional solution is to increase the size of the network, but increasing the number of neurons to design a deeper network structure will result in the occupation of very large hardware resources, and will also make training very difficult. In terms of improving the network structure, the present invention uses Resnet34 as the encoding framework, introduces the spatial channel compression and excitation (SCSE) module for correction, and improves the accuracy of training prediction.

[0067] Preferably, if Figure 1 As shown, the present invention discloses an end-to-end semantic segmentation model for pneumothorax CT images, which adopts a coding and decoding network structure. When designing a binary classification network and a semantic segmentation network, the basic feature extraction network in the coding stage is a very important part. Since it is difficult to obtain massive pneumothorax CT images, the present invention adopts a Unet model with pre-trained parameters based on the Resnet34 framework as a feature extraction network. Because it adds a residual network method in the network, it solves the problem that when the network depth reaches a certain level, the error increases, the effect deteriorates, and the gradient disappearance phenomenon becomes more obvious, making it difficult to achieve the minimum loss of the network back propagation. In the coding stage, the SCSE module structure is introduced to obtain the image binary classification loss. When decoding, the feature information of each layer of the coding stage is fused, up-sampled, and the feature information of each layer of the up-sampled layer is fused again to obtain a semantic segmentation image. Then, the loss functions of the image binary classification and image semantic segmentation are fused, and the weight relationship is adjusted to obtain a global loss function and then optimized.

[0068] Figure 1In the figure, Conv means convolution, Bn1 means batch normalization, Relu means activation function, Resnet_L1 means the first layer of Resnet34, SCSE means spatial channel compression and excitation module (Spatial-Channel Sequeeze & Excitation), and upsampling means upsampling operation.

[0069] Preferably, if Figure 3 As shown in the figure, when performing semantic segmentation on pneumothorax CT images, in order to reduce noise interference, when the sum of the image pixels obtained by segmentation is less than a certain number, we predict that the image has no lesion features, which is equivalent to using the semantic segmentation algorithm to perform a binary classification operation. However, the semantic segmentation model and loss function only consider the classification of a single pixel, and do not consider the overall classification of the sum of all pixels in the entire image. Therefore, in the optimization process, only local optimization is performed, and overall optimization is not considered. To address this problem, we introduced a multi-task learning strategy. While ensuring that the entire network structure remains unchanged, multiple tasks "classification" and "segmentation" are set. For binary classification, the logistic loss function loss1 is introduced, and for segmentation tasks, the cross entropy loss function loss2 is introduced. Finally, the fused loss function loss = a*loss1+b*loss2 is obtained. Where a and b are the weight values ​​of the two.

[0070] Preferably, when designing relevant experiments, the present invention adopts the pytorch framework for design. The graphics workstation used in the experiment is configured as follows: 4-core CPU, 16G memory, GeForceGTX1080Ti GPU, 16G video memory, operating system Ubuntu 16.04, and the network model is configured as follows: the optimizer adopts random gradient descent, momentum=0.9, weight_decay=0.0001, LearningRate=0.002, classification loss weight a=0.1, semantic segmentation loss weight b=0.9. Since a single picture is 512*512 and the picture is large, the batch size is set to 4. In order to increase the batch size, the present invention adopts gradient accumulation to achieve "video memory expansion". Each time a batch of data is obtained, the gradient is calculated by the gradient descent method, the gradient is not cleared, and it is continuously accumulated. After a certain number of accumulations, the network parameters are updated according to the accumulated gradient, and then the gradient is cleared for the next cycle. Under certain conditions, the larger the batch size, the better the training effect. Gradient accumulation achieves a disguised expansion of the batch size. The present invention sets the number of accumulations to 4 times, which is equivalent to setting the batch size to 16.

[0071] Preferably, the model designed by the present invention predicts pneumothorax CT images, wherein the Dice similarity coefficient score reaches 85.3%, taking 3 patients as an example for prediction.

[0072] The comparative analysis results of different network models show that using Resnet34 as the encoding framework and introducing the SCSE module will improve the performance by about 4% compared to the single Unet network. The prediction accuracy of using binary classification and semantic segmentation dual tasks will also be 1.2% higher than that of single task prediction.

[0073] Preferably, the deep learning model is a very powerful tool in the field of medical image detection. The use of semantic segmentation to predict the location of pneumothorax CT image lesions will play a good auxiliary role in doctors' identification of diseases. The semantic segmentation model designed by the present invention adopts resnet34 as the framework, which solves the problem of gradient disappearance when the network depth reaches a certain level. The SCSE module is introduced to obtain image information from both spatial and channel directions, thereby improving the accuracy of semantic segmentation. The loss function adopts the "classification" and "segmentation" multi-task strategy for learning, and is integrated from the classification of local pixels and global images, thereby further improving the accuracy of semantic segmentation.

[0074] It is worth mentioning that the technical features such as pneumothorax CT images involved in the patent application of this invention should be regarded as prior art. The specific structure, working principle and possible control method and spatial layout method of these technical features can be selected according to the conventional methods in the field, and should not be regarded as the inventive point of the patent of this invention. The patent of this invention will not be further elaborated.

[0075] For those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-task semantic segmentation method for pneumothorax medical images based on Unet, which is used to perform semantic segmentation on pneumothorax medical images, characterized in that: The following steps are involved: Step S1: Acquire a pneumothorax CT image and encode the pneumothorax CT image through a first model and a second model, thereby obtaining a binary classification of the image; Step S2: During the decoding process, the feature information of each layer in the encoding process stage is fused and up-sampled, and the up-sampled information of each layer is continuously fused in combination with the second model to obtain a semantic segmentation image; Step S3: The loss function of the binary classification of the image and the loss function of the semantic segmentation image are fused, and the loss function is learned using a multi-task strategy of classification and segmentation fusion, and the fusion is performed from the classification of local pixels and global pixels. After adjusting the weight relationship, the global loss function is obtained, thereby realizing the semantic segmentation of the pneumothorax CT image; Step S1 is specifically implemented as follows: Step S1.1: The first model is used as a basic feature extraction network for pneumothorax CT images and basic feature extraction is performed by adding a residual network to the network; Step S1.2: The second model corrects the extracted basic features through the space compression module and the channel compression module, thereby obtaining the corrected data of each layer corresponding to the first model; Step S1.3: encoding the pneumothorax CT image by using the first model and the second model to obtain a binary classification of the image; Step S2 is specifically implemented as the following steps: Step S2.1: the first module in the encoding process upsamples the module where the maximum pooling process is located in the encoding process to obtain a first feature map and performs a first correction process on the first feature map through a second model to obtain first corrected data; Step S2.2: the second module in the encoding process upsamples the first corrected data obtained by the first module and fuses the corrected data corresponding to the module where the fifth layer of the first model in the encoding process is located to obtain a second feature map and performs a second correction process on the second feature map through the second model to obtain second corrected data; Step S2.3: the third module in the encoding process upsamples the second corrected data obtained by the second module and fuses the corrected data corresponding to the module where the fourth layer of the first model in the encoding process is located to obtain a third feature map and performs a third correction process on the third feature map through the second model to obtain third corrected data; Step S2.4: the fourth module in the encoding process upsamples the third corrected data obtained by the third module and fuses the corrected data corresponding to the module where the third layer of the first model in the encoding process is located to obtain a fourth feature map and performs a fourth correction process on the fourth feature map through the second model to obtain fourth corrected data; Step S2.5: the fifth module in the encoding process upsamples the fourth corrected data obtained by the fourth module and fuses the corrected data corresponding to the module where the second layer of the first model in the encoding process is located to obtain a fifth feature map and performs a fifth correction process on the fifth feature map through the second model to obtain fifth corrected data; Step S2.6: Fusing the corresponding corrected data obtained by each module in the encoding process, thereby completing the encoding process and obtaining a semantic segmentation image; The second model in step S1.2 is modified by the space compression module and the channel compression module as follows: Assume that the input pneumothorax CT image is X∈R H×W×C , assuming that the basic features obtained in the encoding process or the feature map obtained in the decoding process is U∈R H×W×C' , H and W represent the spatial height and width of the image, C and C' represent the number of input and output channels. After convolution and nonlinear transformation operations, U contains the spatial and channel information in X. After each encoding or decoding process, the second model is introduced to correct U to U'. For the spatial compression module, assume that the input feature map U = [u1,u2,…u C ], where u i ∈R H×W Representing the feature map of each channel, through the global average pooling operation, we get the vector z∈R 1×1×C , the k-th channel feature map is transformed as follows: This transformation introduces global spatial feature information into z, and then z' is obtained by z'=W1(δ(W2z)), where are two fully connected networks, δ() is the ReLU operation, and then through the sigmoid transformation, the normalized correction coefficient σ(z')∈[0,1] is obtained. Finally, the correction coefficient is multiplied by the input feature map U to obtain U′ cSE , AT' cSE =[σ(z′1)u1,σ(z'2)u2,…σ(z' C )at C ]; For the channel compression module, assume that the input feature map U = [u1,u2,…u C ], through conversion we get U = [u 1,1 ,u 1,2 …u i ,j …u H,W ], where u i,j ∈R 1×1×C Represents the corresponding spatial position (i, j), through q = W sq *U convolution operation realizes spatial squeezing operation, and then through sigmoid transformation, the normalized correction coefficient σ(q)∈[0,1] is obtained. Finally, the correction coefficient is multiplied by the input feature map U to obtain U' sSE , U′ sSE =[σ(q 1,1 )u 1,1 ,σ(q 1,2 )u 1,2 ,…σ(q h,w )u H,W ]; Finally, the fusion information of the two is obtained by adding them together, and the importance of spatial information and channel information is weighted, and the output is U′ sCSE =U′ cSE +U′ sSE .

2. A multi-task pneumothorax medical image semantic segmentation system based on Unet, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the multi-task pneumothorax medical image semantic segmentation method based on Unet as claimed in claim 1 are implemented.

Images