A spinal column segmentation method
By fusing multi-scale features from the U-Net structure and the XUnet network, the problem of time-consuming and low-accuracy lumbar spine segmentation in existing technologies is solved, and efficient and accurate automatic vertebral segmentation is achieved.
Patent Information
- Application Number
- CN202210916681.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-01
- Publication Date
- 2025-11-07
- Estimated Expiration
- 2042-08-01
AI Technical Summary
In existing technologies, vertebral body segmentation methods are time-consuming and have low accuracy. In particular, the skip connection method based on the Unet model ignores the fusion of semantic information at different scales, resulting in the loss of detailed information and affecting the accuracy of lumbar vertebral body segmentation.
The U-Net structure is used to roughly locate the lumbar spine, and the bounding box is calculated by combining 4-neighborhood. The XUnet network is used for feature extraction. A 3D XUnet network is built by Inception block and multi-scale feature cross-fusion to achieve multi-scale feature fusion.
It improves the accuracy and efficiency of lumbar spine segmentation, reduces redundant network parameters, enhances network expressive power, and achieves automatic and efficient vertebral segmentation.
Smart Images
Figure CN115272373B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of spinal column segmentation, and particularly relates to a spinal column segmentation method. BACKGROUND
[0002] CT imaging technology is used to intuitively observe internal information of human body, to count volume information of various tissues, and to become an effective method for evaluating three-dimensional morphology of vertebral bodies due to its fast scanning speed and clear image. Precise lumbar vertebral body segmentation is an important step for subsequent analysis and treatment, including checking abnormal conditions of vertebral body fracture, biomechanical modeling, lumbar interbody fusion, and anterior and posterior lumbar interbody fusion. In particular, the effect of these analyses and interventions puts high requirements on the accuracy of lumbar intervertebral body segmentation. Before the maturity of artificial intelligence technology, automatic segmentation of vertebral bodies is mostly completed by manual delineation of multiple rich radiologists.
[0003] Manual vertebral body delineation method has huge workload and takes a long time, and the delineation standards of different doctors are inconsistent, resulting in large differences in the delineation results of vertebral bodies. Traditional image segmentation methods such as threshold-based and region-based segmentation can be applied to medical image segmentation problems. However, due to the influence of imaging devices, imaging principles and other factors, the content and morphology of medical images are relatively complex, and these traditional methods still have great challenges in segmentation accuracy. Image segmentation technology based on deep learning has made significant progress, and the Unet model has achieved good results in many medical segmentation tasks because it uses a symmetrical U-shaped structure to consider using symmetrical up-sampling to compensate for the information lost by down-sampling. However, the skip-layer connection method forces the encoder and decoder to only perform feature fusion at the same depth, ignoring the fusion of semantic information with different scales, which can cause the loss of some detailed information. SUMMARY
[0004] In view of the deficiencies of the prior art, the present application provides a spinal column segmentation method to solve the above problems.
[0005] To achieve the above purpose, the present application provides the following technical scheme: a spinal column segmentation method, comprising the following steps:
[0006] Step 1: data preprocessing, generating a spinal column heat map and windowing the image, setting the window level and window width of the lumbar skeletal tissue to 400 and 1800 respectively;
[0007] Step 2: using a U-Net structure to roughly locate the lumbar vertebrae;
[0008] Step 3: calculating the image bounding box, obtaining the selected lumbar vertebrae range using 4-neighborhood, and selecting voxels with HU values greater than 0.3 times the maximum HU value of the image as possible lumbar vertebrae, and then expanding the obtained bounding box in each dimension by 0.05-0.2 times the size of the corresponding image.
[0009] Step 4: Network segmentation based on XUnet network, XUnet uses Inception block to extract features;
[0010] Step 5: Build 3D XUnet network through multi-scale feature cross fusion.
[0011] Further, in step 2, using the U-Net structure to roughly locate the lumbar vertebrae further comprises the following steps:
[0012] Step 2.1: Set input X e R H*W*D*C , after feature extraction by the encoder, obtain The decoding feature is represented as:
[0013] Y i =σ(f 3*3*3 (Concat{downsample(X i-1 ),X i ,upsample(X i+1 ),upsample(Y i+1 )}))
[0014] Where Concat represents concatenating features, downsample represents downsampling, upsample represents upsampling, and σ is an activation function, and f 3*3*3 represents a convolution operation with a 3*3*3 convolution kernel;
[0015] Step 2.2: Multi-scale feature fusion during decoding according to the method described in step 2.1.
[0016] Further, in step 2.2, the specific flow of data in the network is as follows:
[0017] The first layer of the encoder, the input image with a size of 96*96*128*1, extracts features through two convolutions to obtain a feature map X1, and obtains a feature map X1' with 64 channels and a size of 48*48*64 through once downsampling;
[0018] The second layer of the encoder, the input feature map X1' with a size of 48*48*64, extracts features through two convolutions to obtain a feature map X2, and obtains a feature map X2' with 64 channels and a size of 24*24*32 through once downsampling;
[0019] The third layer of the encoder, the input feature map X2' with a size of 24*24*32, extracts features through two convolutions to obtain a feature map X3, and obtains a feature map X3' with 64 channels and a size of 12*12*16 through once downsampling;
[0020] The fourth layer of the encoder, the feature map X3' with an input size of 12*12*16, extracts features through two convolutions to obtain a feature map X4, and obtains a feature map X4' with 64 channels and a size of 6*6*8 through one down-sampling;
[0021] The fifth layer of the encoder, the feature map X4' with an input size of 6*6*8, extracts features through two convolutions to obtain a feature map X5 with 64 channels and a size of 6*6*8;
[0022] The first layer of the decoder, the multi-scale feature X4[16:48] and the feature map X5 after deconvolution are spliced along the channel direction to obtain a feature Y1, and then two convolutions are performed for feature fusion to obtain a feature map Y1' with 64 channels and a size of 12*12*16;
[0023] The second layer of the decoder, the multi-scale feature X3[16:48] and the up-sampled feature map Y1' are spliced along the channel direction to obtain a feature Y2, and then two convolutions are performed for feature fusion to obtain a feature map Y2' with 64 channels and a size of 24*24*32;
[0024] The third layer of the decoder, the multi-scale feature X2[16:48] and the up-sampled feature map Y2' are spliced along the channel direction to obtain a feature Y3, and then two convolutions are performed for feature fusion to obtain a feature map Y3' with 64 channels and a size of 48*48*64;
[0025] The fourth layer of the decoder, the multi-scale feature X1[16:48] and the up-sampled feature map Y3' are spliced along the channel direction to obtain a feature map Y4, and then two convolutions are performed for feature fusion to obtain a feature map Y4' with 64 channels and a size of 96*96*128, and finally a positioning heat map is generated using a convolution kernel with a size of 1*1*1.
[0026] In step 4, the XUnet uses an Inception block to extract features, wherein the Inception block includes two feature extraction layers.
[0027] Further, the Inception block includes a first feature extraction layer and a second feature extraction layer, the first feature extraction layer includes three convolution blocks and one maximum pooling, and the second feature extraction layer includes three convolution blocks.
[0028] Further, the first feature extraction layer includes three convolutional blocks, respectively convblock1-1, convblock1-2 and convblock1-3, wherein convblock1-1 is composed of 64 convolutional kernels with a size of 1*1*1, convblock1-2 is composed of 32 convolutional kernels with a size of 1*1*1, and convblock1-3 is composed of 16 convolutional kernels with a size of 3*3*3; the second feature extraction layer includes three convolutional blocks, respectively convblock2-1, convblock2-2 and convblock2-3, wherein convblock2-1 is composed of 128 convolutional kernels with a size of 3*3*3, convblock2-2 is composed of 32 convolutional kernels with a size of 5*5*5, and convblock2-3 is composed of 32 convolutional kernels with a size of 1*1*1; the feature extraction results of convblock1-1, convblock2-1, convblock2-2 and convblock2-3 are connected and fused as the final output of the Inception block.
[0029] Further, in step 5, the 3D XUnet network further includes:
[0030] Step 5.1: Set input X∈R H*W*D*C , after feature extraction by the encoder, obtain The decoding feature representation is Y i =σ(f 3*3*3 (Concat{downsample(X i-1 ),X i ,upsample(X i+1 ),upsample(Y i+1 )})), wherein downsample represents down-sampling, upsample represents up-sampling, σ is an activation function, and f 3*3*3 represents a convolution operation with a convolutional kernel with a size of 3*3*3.
[0031] Step 5.2: realize multi-scale feature fusion in the decoding process in the manner described in a, and the specific flow of data in the network is as follows:
[0032] The first layer of the encoder, an image with a size of 64*64*128*1 is input, and the feature map X1 is obtained after twice convolutional feature extraction, and the 32-channel feature map X1' with a size of 32*32*64 is obtained after once down-sampling;
[0033] The second layer of the encoder, the feature map X1' with the input size of 32*32*64, extracts features through two times of convolution to obtain a feature map X2, and obtains a feature map X2' with 64 channels and the size of 16*16*32 through one time of down-sampling;
[0034] The third layer of the encoder, the feature map X2' with the input size of 16*16*32, extracts features through two times of convolution to obtain a feature map X3, and obtains a feature map X3' with 128 channels and the size of 8*8*16 through one time of down-sampling;
[0035] The fourth layer of the encoder, the feature map X3' with the output size of 8*8*16, extracts features through one time of convolution and one Inception block to obtain a feature map X4 with 256 channels and the size of 4*4*8;
[0036] The first layer of the decoder, the multi-scale features X2 and X3 and the feature map X4 after deconvolution are spliced along the channel direction to obtain a feature Y1, and then two convolutions are performed for feature fusion to obtain a feature map Y1' with 128 channels and the size of 16*16*32;
[0037] The second layer of the decoder, the multi-scale features X1, X2 and X3 and the feature map Y1' after up-sampling are spliced along the channel direction to obtain a feature Y2, and then two convolutions are performed for feature fusion to obtain a feature map Y2' with 64 channels and the size of 32*32*64;
[0038] The third layer of the decoder, the multi-scale features X1 and X2 and the feature map after up-sampling are spliced along the channel direction to obtain a feature map Y3, and then two convolutions are performed for feature fusion to obtain a feature map Y3' with 32 channels and the size of 64*64*128, and finally the segmentation result map is returned by combining the Softmax function.
[0039] Further, the step of building the 3D XUnet network further comprises:
[0040] Step 5.3: The Inception block includes an a feature extraction layer and a b feature extraction layer, the a feature extraction layer includes three convolutional blocks of convblock1-1, convblock1-2 and convblock1-3 and a maximum pooling, wherein convblock1-1 is composed of 64 convolutional kernels with a size of 1*1*1, convblock1-2 is composed of 32 convolutional kernels with a size of 1*1*1, and convblock1-3 is composed of 16 convolutional kernels with a size of 3*3*3; the b feature extraction layer includes three convolutional blocks of convblock2-1, convblock2-2 and convblock2-3, wherein convblock2-1 is composed of 128 convolutional kernels with a size of 3*3*3, convblock2-2 is composed of 32 convolutional kernels with a size of 5*5*5, and convblock2-3 is composed of 32 convolutional kernels with a size of 1*1*1; the feature extraction results of convblock1-1, convblock2-1, convblock2-2 and convblock2-3 are connected and fused as the final output of the Inception block.
[0041] The application has the advantages that: by aggregating features to different semantic scales and using the Inception block to replace the convolution operation, the network width and depth are deepened, the network expression capability is improved, the redundant parameters in the network are further reduced, the deep learning and automatic and efficient automatic lumbar vertebra segmentation effect are realized, the effect of roughly positioning the lumbar vertebra to accurately positioning the lumbar vertebra is good, and outstanding progress is achieved, and the application has certain use value and popularization value. BRIEF DESCRIPTION OF DRAWINGS
[0042] Figure 1 It is a flowchart of the application.
[0043] Figure 2 It is a Unet structure diagram of the application.
[0044] Figure 3 It is a 3D XUnet structure diagram of the application. DETAILED DESCRIPTION
[0045] As shown in Figure 1 , 2 , a spine segmentation method includes the following steps:
[0046] Step 1: data preprocessing, generating a spine heat map and windowing the image, setting the window level and window width of the lumbar vertebrae tissue to 400 and 1800 respectively;
[0047] Step 2: roughly positioning the lumbar vertebrae using the U-Net structure;
[0048] Step 3: Calculate the image bounding box, get the candidate lumbar range with 4-neighborhood, and select the voxels with HU value greater than 0.3 times the maximum HU value of the image as the possible lumbar vertebrae, and then expand the obtained bounding box in x, y and z dimensions by 0.1 times the size of the image;
[0049] Step 4: Network segmentation based on XUnet network, XUnet uses Inception block for feature extraction;
[0050] Step 5: Build 3D XUnet network through multi-scale feature cross fusion.
[0051] In step 2, the U-Net structure is used to roughly locate the lumbar vertebrae, which further includes the following steps:
[0052] Step 2.1: Set input X∈R H*W*D*C , after feature extraction by the encoder, get The decoding feature is represented as:
[0053] Y i =σ(f 3*3*3 (Concat{downsample(X i-1 ),Xi,upsample(X i+1 ),upsample(Y i+1 )}))
[0054] Where Concat represents concatenating features, downsample represents downsampling, upsample represents upsampling, σ is the activation function, and f 3*3*3 represents convolution operation with 3*3*3 size convolution kernel;
[0055] Step 2.2: Multi-scale feature fusion during decoding according to the method described in step 2.1.
[0056] In step 2.2, the specific flow of data in the network is as follows:
[0057] The first layer of the encoder, the input image with a size of 96*96*128*1, extracts features through two convolutions to get feature map X1, and gets a feature map X1' with 64 channels and a size of 48*48*64 through once downsampling;
[0058] The second layer of the encoder, the input feature map X1' with a size of 48*48*64, extracts features through two convolutions to get feature map X2, and gets a feature map X2' with 64 channels and a size of 24*24*32 through once downsampling;
[0059] The third layer of the encoder, the feature map X2' with the input size of 24*24*32, extracts features through two times of convolution to obtain a feature map X3, and obtains a feature map X3' with 64 channels and the size of 12*12*16 through one time of down-sampling;
[0060] The fourth layer of the encoder, the feature map X3' with the input size of 12*12*16, extracts features through two times of convolution to obtain a feature map X4, and obtains a feature map X4' with 64 channels and the size of 6*6*8 through one time of down-sampling;
[0061] The fifth layer of the encoder, the feature map X4' with the input size of 6*6*8, extracts features through two times of convolution to obtain a feature map X5 with 64 channels and the size of 6*6*8;
[0062] The first layer of the decoder, the multi-scale feature X4[16:48] and the feature map X5 after the deconvolution are spliced along the channel direction to obtain a feature Y1, and then two convolutions are performed for feature fusion to obtain a feature map Y1' with 64 channels and the size of 12*12*16;
[0063] The second layer of the decoder, the multi-scale feature X3[16:48] and the feature map Y1' after the up-sampling are spliced along the channel direction to obtain a feature Y2, and then two convolutions are performed for feature fusion to obtain a feature map Y2' with 64 channels and the size of 24*24*32;
[0064] The third layer of the decoder, the multi-scale feature X2[16:48] and the feature map Y2' after the up-sampling are spliced along the channel direction to obtain a feature Y3, and then two convolutions are performed for feature fusion to obtain a feature map Y3' with 64 channels and the size of 48*48*64;
[0065] The fourth layer of the decoder, the multi-scale feature X1[16:48] and the feature map Y3' after the up-sampling are spliced along the channel direction to obtain a feature map Y4, and then two convolutions are performed for feature fusion to obtain a feature map Y4' with 64 channels and the size of 96*96*128, and finally, a positioning heat map is generated using one convolution kernel with the size of 1*1*1.
[0066] In step 4, the XUnet uses an Inception block to extract features, wherein the Inception block includes two feature extraction layers.
[0067] The Inception block includes a first feature extraction layer and a second feature extraction layer, the first feature extraction layer includes three convolution blocks and one maximum pooling, and the second feature extraction layer includes three convolution blocks.
[0068] The three convolutional blocks included in the first feature extraction layer are respectively convblock1-1, convblock1-2 and convblock1-3, wherein convblock1-1 is composed of 64 convolutional kernels with a size of 1*1*1, convblock1-2 is composed of 32 convolutional kernels with a size of 1*1*1, and convblock1-3 is composed of 16 convolutional kernels with a size of 3*3*3; the three convolutional blocks included in the second feature extraction layer are respectively convblock2-1, convblock2-2 and convblock2-3, wherein convblock2-1 is composed of 128 convolutional kernels with a size of 3*3*3, convblock2-2 is composed of 32 convolutional kernels with a size of 5*5*5, and convblock2-3 is composed of 32 convolutional kernels with a size of 1*1*1; the feature extraction results of convblock1-1, convblock2-1, convblock2-2 and convblock2-3 are connected and fused as the final output of the Inception block, so as to aggregate the features of different semantic scales; meanwhile, the Inception block is used to replace the convolution operation, which deepens the network width and depth, improves the network expression ability, and further reduces the redundant parameters in the network.
[0069] In step 5, the 3D XUnet network further includes:
[0070] In step 5.1, the input X is set to R H*W*D*C , and the feature extraction is performed through the encoder to obtain The decoding feature can be represented as: Y i =σ(f 3*3*3 (Concat{downsample(X i-1 ),X i ,upsample(X i+1 ),upsample(Y i+1 )})), wherein downsample represents down-sampling, upsample represents up-sampling, σ is an activation function, and f 3*3*3 represents a convolution operation with a 3*3*3 size convolutional kernel.
[0071] In step 5.2, the multi-scale feature fusion is realized in the decoding process in the manner described in a, and the specific flow of data in the network is as follows:
[0072] In the first layer of the encoder, an image with a size of 64*64*128*1 is input, and the feature map X1 is obtained through twice convolutional feature extraction, and the 32-channel feature map X1' with a size of 32*32*64 is obtained through once down-sampling;
[0073] The second layer of the encoder, the feature map X1' with an input size of 32*32*64, extracts features through two convolutions to obtain a feature map X2, and obtains a feature map X2' with 64 channels and a size of 16*16*32 through one down-sampling;
[0074] The third layer of the encoder, the feature map X2' with an input size of 16*16*32, extracts features through two convolutions to obtain a feature map X3, and obtains a feature map X3' with 128 channels and a size of 8*8*16 through one down-sampling;
[0075] The fourth layer of the encoder, the feature map X3' with an output size of 8*8*16, extracts features through one convolution and an Inception block to obtain a feature map X4 with 256 channels and a size of 4*4*8;
[0076] The first layer of the decoder, the multi-scale features X2 and X3 are spliced with the feature map X4 after deconvolution along the channel direction to obtain a feature Y1, and then two convolutions are performed for feature fusion to obtain a feature map Y1' with 128 channels and a size of 16*16*32;
[0077] The second layer of the decoder, the multi-scale features X1, X2 and X3 are spliced with the feature map Y1' after up-sampling along the channel direction to obtain a feature Y2, and then two convolutions are performed for feature fusion to obtain a feature map Y2' with 64 channels and a size of 32*32*64;
[0078] The third layer of the decoder, the multi-scale features X1 and X2 are spliced with the feature map after up-sampling along the channel direction to obtain a feature map Y3, and then two convolutions are performed for feature fusion to obtain a feature map Y3' with 32 channels and a size of 64*64*128, and finally the segmentation result map is returned by combining the Softmax function.
[0079] The steps of building the 3D XUnet network also include:
[0080] Step 5.3: The Inception block includes an a feature extraction layer and a b feature extraction layer, the a feature extraction layer includes three convolution blocks of convblock1-1, convblock1-2 and convblock1-3 and a maximum pooling, wherein the convblock1-1 is composed of 64 convolution kernels with a size of 1*1*1, the convblock1-2 is composed of 32 convolution kernels with a size of 1*1*1, and the convblock1-3 is composed of 16 convolution kernels with a size of 3*3*3; the b feature extraction layer includes three convolution blocks of convblock2-1, convblock2-2 and convblock2-3, wherein the convblock2-1 is composed of 128 convolution kernels with a size of 3*3*3, the convblock2-2 is composed of 32 convolution kernels with a size of 5*5*5, and the convblock2-3 is composed of 32 convolution kernels with a size of 1*1*1; the feature extraction results of the convblock1-1, the convblock2-1, the convblock2-2 and the convblock2-3 are connected and fused as the final output of the Inception block.
[0081] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement or improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A method for spinal segmentation, characterized in that, Comprising the following steps: Step 1: data preprocessing, generating a spine heat map and windowing the image, setting the window level and window width of the lumbar bone tissue to 400 and 1800 respectively; Step 2: roughly positioning the lumbar vertebrae using the U-Net structure; Step 3: calculating the image bounding box, obtaining the candidate lumbar vertebrae range using the 4-neighborhood, and selecting the voxels with HU values greater than 0.3 times the maximum HU value of the image as possible lumbar vertebrae, then expanding the obtained bounding box in each dimension by 0.05-0.2 times the size of the image; Step 4: network segmentation based on XUnet network, which builds a 3D XUnet network through multi-scale feature cross-fusion, and the building of the 3D XUnet network comprises: Let x be the input After feature extraction by the encoder, we get: The decoding feature representation is expressed as: , wherein represents down-sampling, represents up-sampling, is an activation function, represents a convolution operation with a 3*3*3 size of a convolution kernel, the XUnet network uses an Inception block for feature extraction, the Inception block includes a first feature extraction layer and a second feature extraction layer, the first feature extraction layer includes three convolution blocks and a maximum pooling, and the second feature extraction layer includes three convolution blocks.
2. The method of claim 1, wherein, In the step 2, the rough positioning of the lumbar vertebrae using the U-Net structure further comprises the following steps: Step 2.1: Set input , after feature extraction by the encoder , the decoded feature representation is wherein denotes down-sampling, denotes up-sampling, is an activation function, denotes a convolution operation with a 3*3*3 size kernel; Step 2.2: multi-scale feature fusion in the decoding process according to the manner described in step 2.
1.
3. The method of claim 2, wherein, In the step 2.2, the specific flow of data in the network is as follows: The first layer of the encoder, the image with the input size of 96*96*128*1, extracts features through twice convolution to obtain a feature map , obtains a feature map with 64 channels and the size of 48*48*64 through once down-sampling ; The second layer of the encoder, the input scale is 48*48*64 feature map , the feature map is obtained by twice convolution extraction , one down-sampling obtains 64 channels, and the size of the feature map is 24*24*32 ; The third layer of the encoder, the input scale is 24*24*32 feature map , the feature map is obtained by twice convolution extraction , and one down-sampling obtains a feature map with 64 channels and a size of 12*12*16 ; The fourth layer of the encoder, the feature map with an input size of 12*12*16 , the feature map obtained by extracting features twice , the feature map with 64 channels and a size of 6*6*8 obtained by once down-sampling ; The fifth layer of the encoder, the input scale is 6*6*8 feature map After twice convolution to extract features, 64 channels, 6*6*8 feature map ; Decoder first layer, multi-scale features [16:48] with deconvolved feature maps Along the channel direction splicing features , and then two convolution features are fused to obtain 64 channels, and the feature map is 12*12*16 ; Decoder second layer, multi-scale feature [16:48] with the up-sampled feature map Along the channel direction splicing feature , and then two convolution is carried out feature fusion to obtain 64 channels, size 24*24*32 feature map ; Decoder third layer, multi-scale feature [16:48] with the up-sampled feature map Along the channel direction splicing feature , and then two convolution is fused to obtain 64 channels, size 48*48*64 feature map ; decoder fourth layer, multi-scale feature [16:48] with the up-sampled feature map along the channel direction to obtain a feature map , and then two convolutions are performed for feature fusion to obtain a feature map with 64 channels and a size of 96*96*128 , and finally a positioning heat map is generated using a convolution kernel with a size of 1*1*1.
4. The method of claim 1, wherein, The three convolution blocks included in the first feature extraction layer are convblock1-1, convblock1-2 and convblock1-3, wherein convblock1-1 is composed of 64 convolution kernels with a size of 1*1*1, convblock1-2 is composed of 32 convolution kernels with a size of 1*1*1, and convblock1-3 is composed of 16 convolution kernels with a size of 3*3*3; the three convolution blocks included in the second feature extraction layer are convblock2-1, convblock2-2 and convblock2-3, wherein convblock2-1 is composed of 128 convolution kernels with a size of 3*3*3, convblock2-2 is composed of 32 convolution kernels with a size of 5*5*5, and convblock2-3 is composed of 32 convolution kernels with a size of 1*1*1; the feature extraction results of convblock1-1, convblock2-1, convblock2-2 and convblock2-3 are connected and fused as the final output of the Inception block.
5. The method of claim 1, wherein, The building of the 3D XUnet network further comprises: Multi-scale feature fusion is achieved in the decoding process according to the manner described in a, and the specific flow of data in the network is as follows: The first layer of the encoder, the image with the input size of 64*64*128*1, extracts features through twice convolution to obtain a feature map , obtains a feature map with 32 channels and the size of 32*32*64 through once down-sampling ; The second layer of the encoder, the input feature map is 32*32*64 , and the feature map is obtained by twice convolution extraction , and the feature map is obtained by once down-sampling ; The third layer of the encoder, the input scale is 16*16*32 feature map , the feature map is obtained by twice convolution extraction , 128 channels are obtained by once down-sampling, and the size of the feature map is 8*8*16 ; The fourth layer of the encoder outputs a feature map with a size of 8*8*16 After one convolution and one Inception block, 256 channels of a feature map with a size of 4*4*8 are obtained ; The decoder first layer fuses the multi-scale features with the deconvolved feature maps to obtain a feature map along the channel direction , and then two convolutions are performed to fuse the features to obtain a feature map with 128 channels and a size of 16*16*32 ; Decoder second layer, multi-scale feature with the up-sampled feature map along the channel direction to obtain a feature , and then two convolutions are performed for feature fusion to obtain a feature map with 64 channels and a size of 32*32*64 ; The third layer of the decoder incorporates multi-scale features. The feature map is obtained by concatenating the upsampled feature map along the channel direction. The features are then fused using two convolutions to obtain a 32-channel feature map with a size of 64*64*128. Finally, the Softmax function is used to complete the segmentation, and the segmentation result image is returned.
6. The method of claim 5, wherein, The step of building the 3D XUnet network further comprises: The Inception block comprises an a feature extraction layer and a b feature extraction layer, the a feature extraction layer comprises three convolution blocks of convblock1-1, convblock1-2 and convblock1-3 and a maximum pooling, wherein the convblock1-1 is composed of 64 convolution kernels with a size of 1*1*1, the convblock1-2 is composed of 32 convolution kernels with a size of 1*1*1, and the convblock1-3 is composed of 16 convolution kernels with a size of 3*3*3; the b feature extraction layer comprises three convolution blocks of convblock2-1, convblock2-2 and convblock2-3, wherein the convblock2-1 is composed of 128 convolution kernels with a size of 3*3*3, the convblock2-2 is composed of 32 convolution kernels with a size of 5*5*5, and the convblock2-3 is composed of 32 convolution kernels with a size of 1*1*1; the feature extraction results of the convblock1-1, the convblock2-1, the convblock2-2 and the convblock2-3 are connected and fused to serve as the final output of the Inception block.