CN114972382A - Brain tumor segmentation algorithm based on lightweight UNet + + network - Google Patents

Abstract

The invention proposes an improved lightweight brain tumor segmentation algorithm based on the UNet++ network model. For accurate multimodal segmentation of brain tumor magnetic resonance imaging (MRI), the UNet++ network model uses dense long and short connections to make the network structure and semantics closely connected, but the dense connections make the UNet++ network increase the amount of computation and the number of parameters, resulting in UNet++ The network training time is slow and puts higher demands on the hardware equipment. The lightweight UNet++ network model replaces the double-layer convolution structure of the UNet++ series with a lightweight residual module, reducing the computational complexity and parameter amount of the network. The dense connection in the network model leads to a large number of feature map channels obtained after each layer is spliced, and the features of some channels have no practical significance for the segmentation task. Add the CBAM attention mechanism after the feature map, learn the screening parameters, and pay attention to usefulness. information to improve the accuracy of network segmentation. The lightweight residual module is applied in the last downsampling, and the deep effective features are better preserved and utilized through the channel splicing of the lightweight residual module, which reduces the training time and further improves the accuracy of brain tumor segmentation.

Description

A Brain Tumor Segmentation Algorithm Based on Lightweight UNet++ Network

technical field

The invention proposes a brain tumor segmentation algorithm based on deep learning, and adopts an improved lightweight brain tumor segmentation algorithm based on the UNet++ network model. The improved lightweight UNet++ network model is applied to brain tumor MRI image segmentation. While ensuring the overall segmentation accuracy, it improves the accuracy of segmentation of brain tumor internal tissues. The application of lightweight modules effectively reduces the computational complexity of the entire model. and parameters, improve the training speed of the model, and solve the problem of slow model training due to the complex network structure of UNet++.

Background technique

At present, brain tumors are common malignant tumors that threaten human life, with high invasiveness and various histological subregions. Due to the inherent spatial heterogeneity of brain tumors and invasive growth, the internal tumor may Complex pathological changes occur, resulting in changes in the grayscale, shape, texture, and histological features of brain tumor MRI images, which make glioma MRI images with multiple modalities present diversity and complexity, which makes radiation Physicians and other clinicians have difficulty identifying and segmenting brain tumors. Manual brain tumor segmentation requires very professional prior knowledge, is time-consuming and labor-intensive, and is prone to errors. It is very dependent on the experience of doctors. Accurate segmentation of brain tumors is still one of the challenging tasks in medical image analysis.

Deep learning has shown a rapid development trend in recent years and is widely used in image segmentation. UNet++ network uses a series of grid-like dense skip paths and encoder-decoder end-to-end structure, and achieves good results in brain tumor segmentation tasks. However, it is precisely because of this structure that the network has a huge amount of parameters, which brings great challenges to the segmentation speed and device memory, and it is difficult to implement the model into practical applications. When segmenting 3D images with a huge amount of data, The UNet++ network training speed is slower and puts forward higher requirements on hardware equipment.

SUMMARY OF THE INVENTION

The invention mainly aims at the problem that the UNet++ network model has a large model structure and a large number of parameters in the processing of 3D brain tumor image segmentation tasks, which leads to slow model training, and proposes a lightweight 3D UNet++ network model. Optimize, retain its excellent dense connections, and use the improved lightweight residual module and lightweight residual module to make the model overall lightweight, reduce the amount of model parameters and improve the model training speed, and add in the structure. The CBAM attention mechanism enables the model to learn to pay attention to effective information, and filter out useful information through CBAM, which further improves the accuracy of network segmentation.

In order to achieve the above object, technical scheme of the present invention is as follows:

An improved lightweight brain tumor segmentation algorithm based on UNet++ network model, including the following steps:

Step 1: Data preprocessing, changing the data set composed of brain tumor MRI images to a size that can be trained by the network according to requirements;

Step 2: Establish a lightweight 3D UNet++ network model, and apply the lightweight residual module, the lightweight residual module and the CBAM attention mechanism in the model;

Step 3: Use the lightweight 3D UNet++ network model for training to obtain brain tumor image segmentation results;

The specific process in the step 1 is as follows:

(1) Perform cross-block processing on the brain tumor MRI image data, and divide the 155×160×160 size brain tumor image into seven 32×160×160 pixel blocks (the part that is not enough is filled with the background image);

(2) Because different imaging methods of brain tumor patients in the same period will produce brain tumor images of four modalities, each case in the BraTS2018 and BraTS2019 datasets has four modalities (t1, t2, flair, t1ce), and the difference between the modalities is There are differences in the imaging methods of the MRI pictures between different countries, resulting in different contrast of the images. First, the extreme value suppression should be applied to the data to prevent the image from having a great influence on the entire image due to the maximum or minimum value, and then the Z-score method is used to standardize the images of each modality (that is, the image minus the mean value). Divide by the standard deviation) to further solve the problem of contrast differences;

The Z-score normalization formula can be expressed as:

where X is the input sample, μ is the mean of all sample data, and σ is the standard deviation of all sample data;

(3) Crop the brain tumor MRI image data, and adjust the input to an appropriate scale through cropping. Because the background accounts for a large proportion of the entire image, and the background area is not the target area for segmentation, it can be identified as an invalid area, so Cropping it does not reduce the target area;

(4) Dicing and splicing, splicing 32×160×160 pixel blocks in the same position of the four modalities in a new dimension, and obtaining a 4×32×160×160 pixel block as the final input of the network. The brain tumor images annotated by the patient’s experts were cross-blocked, and the 155×160×160 image was divided into 7 image blocks of 32×160×160 size (the part that was not divided enough was filled with the background image), and two consecutive blocks were crossed. The number of channels is 8. Copy each 32×160×160 image block into three copies and perform the following operations respectively. Enhancing tumors, peritumoral edema, and non-enhancing tumors were set as 1, and the rest were background 0. Enhancing and non-enhancing tumors were set as 1 and the rest as background 0. Enhanced tumors were set to 1 and the rest to background 0. Through the above operations, three image blocks with a size of 32×160×160 are obtained, and the three pixel blocks are connected in a new dimension to obtain an image block with a size of 3×32×160×160. Finally, the obtained image block is used as labels for the entire network;

(5) Data enhancement, using affine transformation methods such as random cropping and random rotation, scaling, translation, and staggered cutting, to perform data enhancement on brain tumor images;

The specific situation in the second step is as follows:

(1) Apply the lightweight residual module and the lightweight residual module to the 3D UNet++ network to form a brain tumor segmentation network model:

①Lightweight improvement of residual module and residual module;

The specific process of implementing the lightweight residual module:

When performing convolutional feature extraction, the loss of deep feature information is greater than that of shallow convolutional feature extraction, and applying residual-like modules in deep networks can reduce the loss of feature information. The residual-like module first uses 1×1 convolution to expand the channel domain in the main branch, expands the channel domain to 2.5 times the original size, then uses 3×3 convolution for feature extraction, and finally uses 1×1 convolution. Perform channel domain information fusion. The input is not the superposition of the feature map pixels after the shortcut branch, but the splicing of the channel domain. This method is used to make full use of the feature maps before and after convolution;

The lightweight class residual module is further lightweight while retaining the advantages of the class residual module segmentation accuracy. The ordinary convolution with the original convolution kernel size of 3 is changed to grouped convolution to retain its structure, and the number of groups is the convolution kernel. is the number of convolution input channels of 3. Then, in order to solve the problem of inability to interact with channel domain information, after the module performs channel domain splicing, a convolution with a convolution kernel of 1 is used to exchange information between channels, and at the same time, the channel domain is reduced to reduce network parameters and computational effort.

The lightweight residual class module can be expressed as:

x _m+1 =Cat(x _m , F(x _m ; W _m ))

where x _m is the mapping part, F(x _m ; W _m ) is the class residual part, and Cat is the feature map channel domain splicing;

The specific process of implementing the lightweight residual module:

In the lightweight residual module, the number of input channels is changed to 1/4 of the original using the convolution with the convolution kernel of 1, and then the convolution with the convolution kernel of 3 is used for feature extraction, and finally the convolution with the convolution kernel of 1 is used. The product expands the number of channels to twice the number of original input channels, so as to achieve the purpose of reducing network parameters and calculation amount;

The lightweight residual module can be expressed as:

x _l+1 = x _l +F(x _l ; W _l )

where x _l is the direct mapping part, and F(x _l ; W _l ) is the residual part;

②During the training process, in order to reduce the impact of the class imbalance problem on the segmentation accuracy, the training adopts the binary cross entropy (binary_cross_entropy) and the medical image loss Dice Loss to form a hybrid loss function BCEDiceLoss:

The specific process of calculating the cross entropy of the binary classification:

First, the output of the model training is judged. Because the brain tumor segmentation pictures marked by the doctor are preprocessed, the target area is marked as 1, and the non-target area is marked as 0, so the judgment loss input is a binary classification problem. The training output of the network model, each A point is a node, and whether the node is greater than 0.5 is judged and classified;

The specific process of calculating cross entropy:

L(p,t)=[-plog(t)+(1-p)log(1-t)]

p is the expected output of the preprocessed doctor's labeled segmentation image, and t is the output of the actual network model training;

The specific process of calculating the medical image loss DiceLoss:

First understand the definition of Dice coefficient. Dice coefficient is a metric function used to measure the similarity of sets. It is usually used to calculate the similarity of two samples. The final value of s is in the range of [0,1]:

X represents the segmented image, Y represents the predicted segmented image, where |X∩Y| is the intersection between X and Y, and the coefficient 2 in the numerator is because X and Y are repeatedly calculated in the denominator;

The DiceLoss formula is defined as:

Add Laplace smoothing to Dice Loss. Since it is a modified value, the value is defined as 1e-5, that is, add 1e-5 to the numerator and denominator of Dice Loss:

Laplace smoothing can reduce overfitting and avoid the problem that the numerator is divided by 0 when both |X| and |Y| are 0;

The final mixing loss is defined as:

In summary, after using the mixed loss function BCEDiceLoss, the performance of the network model is improved, the accuracy of the Dice coefficient is ensured, the error between the model segmentation result and the expert outline result is reduced, and the segmentation accuracy is improved;

③ The constructed network model uses 3 downsampling and 6 upsampling, and uses a lightweight residual module to replace the UNet++ series double-layer convolution structure. Although direct replacement can make the network lightweight, but compared to the original double-layer volume Product structure, the residual structure only uses a convolution layer with a convolution kernel size of 3, and the rest are replaced by a convolution kernel with a convolution kernel size of 1, resulting in insufficient feature extraction, which may lead to a decrease in segmentation accuracy. In the improvement process Then add the CBAM attention mechanism and apply it to the outermost layer of the U-shaped structure, that is, the feature map obtained by splicing the channel domain after upsampling. At this time, the feature map is obtained by a series of long and short connections, although at this time due to a series of Long and short connections, the feature map semantic gap has been relatively small, but due to multiple splicing, the number of channels is large at this time, and the features of some channels have no practical significance for the segmentation task, so the CBAM attention module is used to learn the screening parameters, Pay attention to useful information and improve the accuracy of network segmentation. Because there is a certain semantic gap between the output feature map and the input feature map after the convolution feature extraction, the lightweight residual module is applied in the last downsampling, and the channel splicing of the lightweight residual module can improve the Good preservation and utilization of deep effective features can further improve the accuracy of brain tumor segmentation.

(2) Add a 3D convolution to the above lightweight UNet++ network model, the size of the convolution kernel is 1, and the number of channels is changed to 3, so that the output is consistent with the number of patient label channels marked by the preprocessed expert.

The specific situation in the third step is as follows:

(1) Use the lightweight 3D UNet++ network model for training, obtain the segmentation results of brain tumor images, and then perform a sigmoid on the segmentation results to determine whether the segmentation results are greater than 0.5, and change the results to 0 and 1, splicing, and then according to three The channel definition is restored to a single channel, that is, the brain tumor segmentation result map is obtained.

Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

(1) The present invention improves the weight of the UNet++ network model. While ensuring the overall segmentation accuracy, the accuracy of the segmentation of the internal tissue of the brain tumor is improved. The application of the lightweight module effectively reduces the computational complexity and parameters of the entire model. It can improve the training speed of the model and solve the problem of slow network training caused by the complex network of the UNet++ model.

Description of drawings

FIG. 1 is a flow chart of the method of the present invention.

FIG. 2 is a lightweight residual-like module of the present invention.

FIG. 3 is a lightweight residual module of the present invention.

Fig. 4 is the network model of the lightweight UNet++ improved by the present invention.

Detailed ways

It will be understood by those skilled in the art that some well-known structures and their descriptions may be omitted from the drawings. The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

The invention provides an improved lightweight brain tumor segmentation algorithm based on the UNet++ network model. The method of the invention can realize the segmentation of the whole brain tumor, the brain tumor core and the enhanced brain tumor core, and efficiently obtain high-precision brain tumor images. Segmentation map and application in brain tumor MRI image repeatability measurement and assessment.

Fig. 1 is a flow chart of the method of the present invention. First, the brain tumor MRI image is preprocessed, and BraTS2018 and BraTS2019 are changed into the input required by the network, and then a lightweight UNet++ network model is constructed, and it is used to train the data and save it. The best network weights to achieve the segmentation task.

The specific implementation steps are:

Step1.1 Perform cross-block processing on the input brain tumor MRI image data;

Step1.2 Use extreme value suppression to standardize the data, and use Z-score method to standardize the images of each modality respectively, subtract the mean and divide the image by the standard deviation;

Normalize with Z-score:

where μ is the mean of all sample data, σ is the standard deviation of all sample data;

Step1.3 Trim the brain tumor MRI image to an appropriate scale and remove the invalid area;

Step1.4 splicing, splicing 32×160×160 pixel blocks in the same position of the four modalities in a new dimension, and obtaining a 4×32×160×160 pixel block as the final input of the network. The brain tumor images annotated by the patient’s experts were cross-blocked, and the 155×160×160 image was divided into 7 image blocks of 32×160×160 size (the part that was not divided enough was filled with the background image), and two consecutive blocks were crossed. The number of channels is 8. Copy each 32×160×160 image block into three copies and perform the following operations respectively. Enhancing tumors, peritumoral edema, and non-enhancing tumors were set as 1, and the rest were background 0. Enhancing and non-enhancing tumors were set as 1 and the rest as background 0. Enhanced tumors were set to 1 and the rest to background 0. Through the above operations, three image blocks with a size of 32×160×160 are obtained, and the three pixel blocks are connected in a new dimension to obtain an image block with a size of 3×32×160×160. Finally, the obtained image block is used as labels for the entire network;

Step1.5 Data enhancement, using affine transformation methods such as random cropping and random rotation, scaling, translation and staggered cutting to enhance the data of brain tumor images;

Step2.1 Apply the lightweight residual module and the lightweight residual module in the 3D UNet++ network to form a brain tumor segmentation network model;

The network model constructed in Step 2.1.1 uses 3 downsampling and 6 upsampling, and uses a lightweight residual module to replace the UNet++ series double-layer convolution structure. Although direct replacement can make the network lightweight, but compared to the original dual Layer convolution structure, the residual structure only uses a convolution layer with a convolution kernel size of 3, and the rest are replaced by a convolution kernel with a convolution kernel size of 1, resulting in insufficient feature extraction, which may lead to a decrease in segmentation accuracy. In the improvement process, the CBAM attention mechanism is added and applied to the feature map obtained by the channel domain splicing after upsampling in the outermost layer of the U-shaped structure. At this time, the feature map is obtained by a series of long and short connections, although at this time due to a series of long and short connections , the semantic gap of the feature map is relatively small, but the number of channels is large due to multiple splicing, and the features of some channels have no practical significance for the segmentation task, so the CBAM attention module is used to learn the screening parameters, and attention is useful. information, improve the accuracy of network segmentation, and apply a lightweight class residual module in the last downsampling, because there is a certain semantic gap between the output feature map and the input feature map after convolution feature extraction. The channel splicing of the residual module can better preserve and utilize deep effective features, and further improve the accuracy of brain tumor segmentation;

The specific process of implementing the lightweight residual module:

When performing convolutional feature extraction, the loss of deep feature information is greater than that of shallow convolutional feature extraction, and applying residual-like modules in deep networks can reduce the loss of feature information. The original residual module first uses 1 × 1 convolution to expand the channel domain in the main branch, expands the channel domain to 2.5 times the original, then uses 3 × 3 convolution for feature extraction, and finally uses 1 × 1 volume. product for channel domain information fusion. The input is not the superposition of the feature map pixels after the shortcut branch, but the splicing of the channel domain. This method is used to make full use of the feature maps before and after convolution;

The lightweight residual module is further lightweight while retaining the segmentation accuracy advantage of the original residual module. The original convolution kernel size of 3 is changed to a grouped convolution to retain its structure, and the number of groups is convolution The number of convolution input channels with kernel 3. Then, in order to solve the problem of inability to interact with channel domain information, after the module performs channel domain splicing, a convolution with a convolution kernel of 1 is used to exchange information between channels, and at the same time, the channel domain is reduced to reduce network parameters and computational effort.

The lightweight residual class module can be expressed as:

x _m+1 =Cat(x _m , F(x _m ; W _m ))

where x _m is the direct mapping part, F(x _m ; W _m ) is the residual part, and Cat is the feature map channel domain splicing;

The specific process of implementing the lightweight residual module:

In the lightweight residual module, the number of input channels is changed to 1/4 of the original using the convolution with the convolution kernel of 1, and then the convolution with the convolution kernel of 3 is used for feature extraction, and finally the convolution with the convolution kernel of 1 is used. The product expands the number of channels to twice the number of original input channels, so as to achieve the purpose of reducing network parameters and calculation amount;

The lightweight residual module can be expressed as:

x _l+1 = x _l +F(x _l ; W _l )

where x _l is the direct mapping part, and F(x _l ; W _l ) is the residual part;

Step2.2 In the training process, in order to reduce the impact of the class imbalance problem on the segmentation accuracy, the training uses a combination of binary cross entropy (binary_cross_entropy) and medical image loss Dice Loss to form a hybrid loss function BCEDiceLoss;

The specific process of calculating the cross entropy of the binary classification:

First, the output of the model training is judged. Because the brain tumor segmentation pictures marked by the doctor are preprocessed, the target area is marked as 1, and the non-target area is marked as 0, so the judgment loss input is a binary classification problem. The training output of the network model, each A point is a node, and whether the node is greater than 0.5 is judged and classified.

The specific process of calculating cross entropy:

L(p,t)=[-plog(t)+(1-p)log(1-t)]

p is the expected output of the preprocessed doctor's labeled segmentation image, and t is the output of the actual network model training;

The specific process of calculating the medical image loss Dice Loss:

First understand the definition of Dice coefficient. Dice coefficient is a metric function used to measure the similarity of sets. It is usually used to calculate the similarity of two samples. The final value of s is in the range of [0,1]:

X represents the segmented image, Y represents the predicted segmented image, where |X∩Y| is the intersection between X and Y, and the coefficient 2 in the numerator is because X and Y are repeatedly calculated in the denominator;

The Dice Loss formula is defined as:

Add Laplace smoothing to Dice Loss. Since it is a modified value, the value is defined as 1e-5, that is, add 1e-5 to the numerator and denominator of Dice Loss:

Laplace smoothing can reduce overfitting and avoid the problem that the numerator is divided by 0 when both |X| and |Y| are 0;

The final mixing loss is defined as:

In summary, after using the mixed loss function BCEDiceLoss, the performance of the network model is improved, the accuracy of the Dice coefficient is ensured, the error between the model segmentation result and the expert outline result is reduced, and the segmentation accuracy is improved;

Step2.3 Add another 3D convolution to the above network model, so that the number of channels becomes 3, so that the output is consistent with the processed doctor's labeled picture;

Step3.1 Use the lightweight 3D UNet++ network model for training, obtain the segmentation results of brain tumor images, and then perform a sigmoid on the segmentation results to determine whether the segmentation results are greater than 0.5, and change the results to 0 and 1, splicing, and then according to three The channel definition is restored to a single channel, that is, the brain tumor segmentation result map is obtained.

Claims (4)

1. A lightweight brain tumor segmentation algorithm based on UNet + + network model improvement is characterized by comprising the following steps:

step 1: data preprocessing, namely changing a data set formed by the brain tumor nuclear magnetic resonance images into a size which can be trained by a network according to requirements;

step 2: establishing a lightweight 3D UNet + + network model, and applying a lightweight residual error module, a lightweight residual error module and a CBAM attention mechanism in the model;

step 3: and training by using a lightweight 3D UNet + + network model to obtain a brain tumor image segmentation result.

2. The UNet + + network model-based improved lightweight brain tumor segmentation algorithm according to claim 1, wherein the specific process in Step1 is as follows:

step1.1, performing cross blocking processing on input brain tumor nuclear magnetic resonance image data;

step1.2, carrying out standardization treatment on the data after extreme value inhibition, respectively standardizing the image of each mode by adopting a Z-score method, and subtracting the mean value from the image and dividing the mean value by the standard deviation;

normalization using Z-score:

wherein μ is the mean of all sample data and σ is the standard deviation of all sample data;

step1.3 cutting the nuclear magnetic resonance image of the brain tumor to a proper scale, and removing an invalid region;

and step1.4, splicing blocks, namely splicing 32 × 160 × 160 pixel blocks at the same positions of four modes in a new dimension to obtain 4 × 32 × 160 × 160 pixel blocks as the final input of the network. The patient expert-labeled brain tumor image was cross-blocked, from an image of 155 × 160 × 160 size, into 7 image blocks of 32 × 160 × 160 size (the insufficiently blocked portions were filled with background images), with 8 cross-channels in two consecutive blocks. The following operations are performed for each 32 × 160 × 160 size image block in triplicate. Enhanced tumors, peritumoral edema, and non-enhanced tumors were set to 1, with the remainder being background 0. Enhanced and non-enhanced tumors were set to 1, with the remainder being background 0. The enhanced tumor was set to 1, with the remainder being background 0. Obtaining three image blocks with the size of 32 multiplied by 160 through the operations, connecting the three pixel blocks in a new dimension to obtain the image blocks with the size of 3 multiplied by 32 multiplied by 160, and finally taking the obtained image blocks as the label of the whole network;

and (3) data enhancement is carried out on the brain tumor image by adopting affine transformation methods such as random cutting, random rotation, scaling, translation, miscut and the like.

3. The UNet + + network model-based improved lightweight brain tumor segmentation algorithm according to claim 1, wherein the specific process in Step2 is as follows:

step2.1 applies a lightweight residual module and a lightweight residual module to a 3D UNet + + network to form a brain tumor segmentation network model;

the network model constructed by Step2.1.1 uses 3 times of down sampling and 6 times of up sampling, and adopts a lightweight residual module to replace a UNet + + series double-layer convolution structure, so that the aim of lightening the network can be achieved, but compared with the original double-layer convolution structure, the residual structure only uses a convolution layer with a convolution kernel size of 3 once, and the rest is replaced by a convolution kernel with a convolution kernel size of 1, so that the feature extraction is insufficient, the segmentation precision is possibly reduced, a CBAM attention mechanism is added in the improvement process, and the method is applied to a feature map obtained by splicing channel domains after the up sampling of a U-shaped structure, at the moment, the feature map is obtained by a series of long and short connections, although the sense ditches of the feature map are relatively small due to the series of long and short connections, but the number of channels is large due to the multi-time splicing, and the features of some channels have no practical significance for the segmentation task, therefore, the CBAM attention module is used for learning and screening parameters, paying attention to useful information and improving the network segmentation precision, and meanwhile, the lightweight class residual error module is applied in the last downsampling, because certain semantic gaps exist between the output feature map and the input feature map after the convolution feature extraction, deep effective features can be better stored and utilized through the channel splicing of the lightweight class residual error module, and the brain tumor segmentation precision is further improved;

the specific process for realizing the lightweight residual module comprises the following steps:

when the convolution feature extraction is carried out, the loss of deep feature information is larger than that of shallow convolution feature extraction, and the feature information loss can be reduced by applying a similar residual error module in a deep network. The original residual module firstly uses convolution of 1 multiplied by 1 to expand the channel domain in the main branch, the channel domain is expanded to 2.5 times of the original domain, then uses convolution of 3 multiplied by 3 to extract the characteristics, and finally uses convolution of 1 multiplied by 1 to fuse the channel domain information. The input is not the superposition of the pixel points of the feature graph but the splicing of the channel domain after passing through the shortcut branch, and the mode is used for fully utilizing the feature graphs before convolution and after convolution;

the lightweight residual error module is further lightweight while retaining the advantage of segmentation precision of the original residual error module, the structure of the lightweight residual error module is retained by changing the common convolution with the original convolution kernel size of 3 into the grouped convolution, and the number of the grouped convolution is the number of convolution input channels with the convolution kernel size of 3. And then, in order to solve the problem that channel domain information cannot be interacted, after the modules are spliced, the modules adopt convolution with convolution kernel of 1 to carry out information interaction between channels, and simultaneously reduce the channel domain, thereby achieving the purpose of reducing network parameters and calculation amount. The lightweight class residual module can be expressed as:

x _m+1 ＝Cat(x _m ，F(x _m ；W _m ))

wherein x is _m For the direct mapping part, F (x) _m ；W _m ) Splicing the residual error part and the Cat is a characteristic diagram channel domain;

the specific process for realizing the lightweight residual module comprises the following steps:

the lightweight residual error module changes the convolution of the input channel number with convolution kernel 1 into original 1/4, then uses the convolution of convolution kernel 3 to extract the characteristics, and finally uses the convolution of convolution kernel 1 to expand the channel number to 2 times of the original input channel number, thereby achieving the purpose of reducing network parameters and calculation amount;

the lightweight residual module can be expressed as:

x _l+1 ＝x _l +F(x _l ；W _l )

wherein x _l For the direct mapping part, F (x) _l ；W _l ) Is a residual error part;

in the training process of Step2.2, in order to reduce the influence of the class imbalance problem on the segmentation accuracy, the training adopts the cross entropy (binary _ cross _ entropy) of the two classes and the medical image Loss Dice Loss to be combined into a mixed Loss function BCEDiceloss;

the specific process of calculating the cross entropy of the second classification comprises the following steps:

firstly, judging the output of model training, wherein the target area is marked as 1 and the non-target area is marked as 0 because a brain tumor segmentation picture marked by a doctor is preprocessed, so that the judgment that the loss input is a binary classification problem is judged, each point is a node in the training output of the network model, and judging and classifying whether the node is more than 0.5;

the specific process of calculating the cross entropy is as follows:

L(p,t)＝[-plog(t)+(1-p)log(1-t)]

p is the expected output of the preprocessed doctor labeling segmentation picture, and t is the output of the actual network model training;

the specific process of calculating the Loss Dice Loss of the medical image comprises the following steps:

first, understanding the definition of the Dice coefficient, the Dice coefficient is a measurement function for measuring the similarity of a set, and is usually used for calculating the similarity of two samples, and finally, the value range of s is [0,1 ]:

x represents a segmented image, and Y represents a predicted segmented image, wherein | X ≧ Y | is the intersection between X and Y, and the coefficient 2 in the numerator is because X and Y are repeatedly calculated in the denominator;

the Dice Loss formula is defined as:

laplace smoothing (Laplace smoothing) is added to the Dice Loss, and since the Laplace smoothing is a modified value, the value is defined as 1e-5, namely, 1e-5 is added to the denominator of the Dice Loss:

laplacian smoothing can reduce overfitting, avoiding the problem of dividing the molecule by 0 when | X | and | Y | are both 0;

the final mixing loss is defined as:

in conclusion, after the mixed loss function BCEDiceLoss is used, the performance of a network model is improved, the precision of the Dice coefficient is ensured, the error of the model segmentation result and the result sketched by an expert is reduced, and the segmentation precision is improved;

and adding a 3D convolution after the network model on the Step2.3 to change the number of channels into 3, so that the output is consistent with the processed doctor labeling picture.

4. The UNet + + network model-based improved lightweight brain tumor segmentation algorithm according to claim 1, wherein the specific process in Step3 is as follows:

step3.1 training by using a lightweight 3D UNet + + network model to obtain a brain tumor image segmentation result, performing sigmoid on the segmentation result once, judging whether the segmentation result is greater than 0.5, changing the result into 0 and 1, splicing, and restoring to a single channel according to three-channel definition to obtain a brain tumor segmentation result graph.

Images