CN-117315243-B - Brain tumor image region segmentation method and device, neural network and electronic equipment

CN117315243BCN 117315243 BCN117315243 BCN 117315243BCN-117315243-B

Abstract

The invention relates to a brain tumor image region segmentation method, a device, a neural network and electronic equipment, which comprise the steps of acquiring brain MRI images, forming a data set, and preprocessing the data set; the method comprises the steps of constructing improved residual attention blocks, replacing multi-layer perceptrons in the improved residual attention blocks, constructing an improved network model, adjusting the number and the replacement method of the improved residual attention blocks by introducing the improved residual attention blocks, inputting the trained network model into a test set for testing, and checking network effects. The brain tumor segmentation method based on the UNet improved residual error attention mechanism has the beneficial effects that the brain tumor segmentation method based on the UNet improved residual error attention mechanism, namely the model structure with double branches, is provided, the problem that UNet cannot utilize global information of images is solved, and the model segmentation quality can be effectively improved.

Inventors

LI YANJUN
Jin Linwei
CAI JIANPING

Assignees

浙江工业大学

Dates

Publication Date: 20260508
Application Date: 20230906

Claims (7)

1. The brain tumor image region segmentation method is characterized by comprising the following steps: S1, acquiring brain MRI images, forming a data set, and preprocessing the data set; s2, constructing an improved residual error attention block to replace the multi-layer perceptron; S2 comprises the following steps: s201, replacing a multi-layer perceptron layer in an original attention block, and adding residual connection on the basis of the multi-layer perceptron; s202, adding GELU an activation function, a random deactivation layer and a linear layer; s3, constructing an improved network model, and adjusting the number and the replacement method of the improved residual attention blocks by introducing the improved residual attention blocks; In S3, the improved network model comprises an encoder part and a decoder part, wherein the encoding blocks of the first three layers of the encoder part consist of a 3D convolution layer, a batch normalization layer and LeakyReLu layers, the decoder part comprises a decoding block used for decoding a feature map to obtain a brain tumor area, and the rear three layers of the encoder part use double-branch improved residual attention blocks, wherein the number of the improved residual attention blocks is (2, 4, 2), the singular residual attention blocks consist of a linear normalization layer, a W-MSA, a batch normalization layer and a residual multi-layer perception machine layer, and the double residual attention blocks consist of a linear normalization layer, a SW-MSA, a batch normalization layer and a residual multi-layer perception machine layer; s4, inputting the trained network model into a test set for testing, and checking the network effect.
2. The brain tumor image region segmentation method according to claim 1, wherein in S1, the brain MRI image has a plurality of modes, and the image of each mode is normalized by using a Z-score method, which is expressed as: X*=(X-μ)/σ where μ is the mean of all sample data and σ is the standard deviation of all sample data.
3. The brain tumor image region segmentation method according to claim 2, wherein in S1, the preprocessing comprises mirror image inversion, rotation, scaling, translation, elastic deformation and image cropping.
4. The brain tumor image region segmentation method according to claim 3, wherein in S4, the preprocessed picture is input into a network, the network parameter weight is updated to obtain an optimal network segmentation result, the segmentation result is subjected to sigmoid function operation, the segmentation result is changed into 0 and 1, the segmentation result is spliced, and the segmentation result is restored into a single channel according to three-channel definition, so that a segmentation result diagram is obtained.
5. A brain tumor image region segmentation apparatus for performing the brain tumor image region segmentation method according to any one of claims 1 to 4, comprising: The acquisition module is used for acquiring brain MRI images, forming a data set and preprocessing the data set; A first construction module for constructing an improved residual attention block to replace the multi-layer perceptron therein; A second construction module for constructing an improved network model, introducing the improved residual attention blocks, and adjusting the number and the replacement method of the improved residual attention blocks; And the test module is used for inputting the trained network model into the test set for testing and checking the network effect.
6. A computer storage medium, wherein a computer program is stored in the computer storage medium, and when the computer program runs on a computer, the computer program causes the computer to execute the brain tumor image region segmentation method according to any one of claims 1 to 4.
7. An electronic device characterized in that the electronic device comprises a processor and a memory, wherein the memory stores a program, and the processor is configured to call the program stored in the memory to cause the electronic device to execute the brain tumor image region segmentation method according to any one of claims 1 to 4.

Description

Brain tumor image region segmentation method and device, neural network and electronic equipment Technical Field The invention relates to the technical field of deep neural networks, in particular to a brain tumor image region segmentation method and device, a neural network and electronic equipment. Background Brain tumors refer to tumors that form within the brain or under the brain membrane, and currently, humans are affected by more than 100 types of brain tumors. Among them, gliomas are the most common malignant brain tumors, originating from glial cells, and often cause some invasion and compression of surrounding brain tissue. Diagnosis and surgical prediction of brain tumors has become increasingly important. With the rapid development of artificial intelligence, tumor prediction and diagnosis technologies based on artificial intelligence are becoming more and more sophisticated. Accurate and fine segmentation of brain tumors can be achieved using voxel analysis and the like, which aids in preoperative planning. In addition, life may also be predicted from segmented tumors. The fine segmentation of tumors from brain tumor images is the leading edge of current research. Magnetic Resonance Imaging (MRI) techniques play an important role in the diagnosis, treatment planning and monitoring of brain tumors. MRI can provide high resolution images that help doctors detect and differentiate brain tumor types, locate tumor locations, assess tumor size, morphology, wettability, and blood supply, and determine whether a tumor has an impact on surrounding structures. MRI scans can generate different image sequences including T1 weighting (T1), T1 enhanced contrast (T1-ce), T2 weighting (T2), and T2 fluid decay inversion recovery (Flair). These image sequences may provide different anatomical and physiological information. The T1 weighted image may show the anatomical location and size of the tumor, the T2 weighted image may show the wettability and cystic changes of the tumor, the FLAIR image may show peripheral edema and brain tissue destruction, and the contrast enhancement sequence may show the boundary and vascular richness of the tumor with the peripheral normal brain tissue. With the development of image segmentation technology based on a deep learning method, the automatic medical image segmentation technology can realize the accurate segmentation of brain tumor images. In recent years, brain tumor segmentation techniques based on deep learning have achieved the most advanced performance among various criteria. Mainly because convolutional neural networks have strong extraction capability for features. However, the convolution operation in the convolutional neural network is to perform feature extraction based on the pixel information of the local neighborhood, and the relationship between global pixels is not directly considered. This approach may not capture some global features, such as long-range dependencies or global semantic information, limiting the performance of convolutional neural networks. Yang et al add an expanding convolution block to the network to increase receptive fields. Still other efforts have increased the ability of networks to extract and integrate semantic information by adding receptive fields, but receptive fields in this manner are still limited to localized areas. Disclosure of Invention The invention aims at overcoming the defects of the prior art and provides a brain tumor image region segmentation method, a device, a neural network and electronic equipment. In a first aspect, a brain tumor image region segmentation method is provided, including: S1, acquiring brain MRI images, forming a data set, and preprocessing the data set; s2, constructing an improved residual error attention block to replace the multi-layer perceptron; s3, constructing an improved network model, and adjusting the number and the replacement method of the improved residual attention blocks by introducing the improved residual attention blocks; s4, inputting the trained network model into a test set for testing, and checking the network effect. Preferably, in S1, the brain MRI image has a plurality of modes, and the Z-score method is used to normalize the image of each mode, which is expressed as: X*=(X-μ)/σ where μ is the mean of all sample data and σ is the standard deviation of all sample data. Preferably, in S1, the preprocessing comprises mirror image overturning, rotation, zooming, translation, elastic deformation and image clipping. Preferably, S2 includes: s201, replacing a multi-layer perceptron layer in an original attention block, and adding residual connection on the basis of the multi-layer perceptron; s202, adding GELU an activation function, a random deactivation layer and a linear layer. Preferably, in S3, the improved network model includes an encoder portion and a decoder portion, wherein the encoding blocks of the first three layers of the encoder portion are composed of a 3D convol