CN110363797B - PET and CT image registration method based on excessive deformation inhibition - Google Patents

Abstract

The invention relates to the technical field of medical image registration, and provides a PET and CT image registration method based on excessive deformation inhibition. Firstly, acquiring a two-dimensional PET/CT sequence image to obtain a PET/CT sequence image set and preprocessing the PET/CT sequence image set to obtain a PET/CT image block training set; then constructing a PET/CT registration network based on the 3D U-Net convolution neural network, and constructing a cost function by combining an image similarity constraint term and an excessive deformation inhibition term; initializing a neural network weight parameter, setting a hyper-parameter, inputting a PET/CT image block training set into a PET/CT registration network, and performing iterative training on the PET/CT image block training set; and finally, inputting the PET/CT image to be registered into the trained PET/CT registration network to generate a registered PET image block. The invention can realize PET/CT elastic registration, improve registration efficiency and accuracy and reduce calculation cost for restraining excessive deformation.

Description

A Registration Method of PET and CT Images Based on Excessive Deformation Suppression

technical field

The invention relates to the technical field of medical image registration, in particular to a PET and CT image registration method based on excessive deformation suppression.

Background technique

Medical image registration plays an important role in many medical image processing tasks. Image registration is usually formulated as an optimization problem seeking a spatial transformation that establishes a pair of fixed and moving images by maximizing a surrogate measure of spatial correspondence between images (such as image intensity correlation between registered images) The pixel/voxel correspondence between them.

Positron Emission Computer Tomography (PET) uses a cyclotron to produce radioactive isotopes 18F and 13N, which participate in the metabolism of the human body after intravenous injection. Tissues or lesions with a high metabolic rate show a clear high metabolic bright signal on PET; tissues or lesions with a low metabolic rate show a low metabolic dark signal on PET. Computed Tomography (CT) uses X-ray beams to scan a certain part of the human body at a certain thickness. When X-rays are irradiated to human tissues, part of the rays are absorbed by the tissues, and part of the rays pass through the human body and are received by the detection organs. Signals can be generated to accurately position the image.

PET/CT can perform functional and anatomical image fusion on the same machine, which is an important progress in imaging medicine. Multimodal image registration utilizes the characteristics of various imaging methods to provide complementary information for different images, increase the amount of image information, and help to more comprehensively understand the nature of the lesion and its relationship with the surrounding anatomical structure, providing clinical diagnosis and treatment. The positioning provides an effective method. It is of great significance for medical image analysis and diagnosis to obtain a fusion image with both structural information and functional information by fusing PET/CT images of two different modalities. PET/CT registration is a challenging task due to the low similarity of pixel intensities between PET/CT images, which leads to excessive deformation after registration.

Most of the existing PET and CT image registration methods are based on iterative optimization. The most commonly used method among them is to transform the registration problem into an optimization problem to minimize a cost function. Commonly used cost functions include mean square error (MSE), mutual information (MI), normalized mutual information (NMI), normalized cross-correlation (NCC), and gradient correlation (GC). However, these similarity metrics directly compare images at the pixel level and cannot reflect higher-level structures in images. Although global optimization methods exist, such as simulated annealing algorithms and genetic algorithms, they require comprehensive sampling of the parameter space, which leads to prohibitively high computational costs. As a result, the existing PET and CT image registration methods have high computational costs for excessive deformation suppression and low efficiency and accuracy for image registration.

Contents of the invention

Aiming at the problems existing in the prior art, the present invention provides a PET and CT image registration method based on excessive deformation suppression, which can realize PET/CT elastic registration, improve registration efficiency and accuracy, and reduce the calculation of excessive deformation suppression cost.

Technical scheme of the present invention is:

A PET and CT image registration method based on excessive deformation suppression, is characterized in that, comprises the following steps:

Step 1: Collect 2D PET sequence images and 2D CT sequence images of m patients to obtain a PET/CT sequence image set;

Step 2: Preprocessing the PET/CT sequence image set to obtain a PET/CT image block training set; the preprocessing includes calculating the SUV value and hu value and limiting the threshold range, adjusting image resolution, generating image blocks and performing normalization One treatment;

Step 3: Construct PET/CT registration network based on 3DU-Net convolutional neural network;

Step 4: Construct the cost function of the PET/CT registration network by combining the image similarity constraint item and the excessive deformation suppression item; the similarity constraint item is normalized cross-correlation NCC, and the excessive deformation suppression item is based on the displacement vector The sum of the penalty term for the difference between the field element and the Gaussian distribution function element;

Step 5: Initialize the weight parameters of the neural network, set the batch size N in the batch processing, the regularization item weight λ, the maximum number of iterations COUNT, the network learning rate, the optimizer, and adopt the learning rate decay strategy;

Step 6: Use the PET/CT image block training set as the input of the PET/CT registration network, output the displacement vector field, input the displacement vector field and the PET image block together to the space transformer, and obtain the registered PET image block, According to the CT image block and the registered PET image block, the similarity constraint item is obtained, and the excessive deformation suppression item is obtained according to the displacement vector field, so as to update the weight parameters of the neural network through the cost function, and perform backpropagation, thus PET/CT The registration network is iteratively trained until the maximum number of iterations is COUNT, and the trained PET/CT registration network is obtained;

Step 7: Perform the preprocessing described in step 1 on the PET/CT image pair to be registered, and input the obtained PET/CT image block pair into the trained PET/CT registration network to generate a registered PET image block and visualize it.

Described step 2 comprises the following steps:

Step 2.1: Calculate the SUV value of the two-dimensional PET sequence image SUV=Pixels _PET ×LBM×1000/injecteddose

Calculate the hu value of the two-dimensional CT sequence image Hu=Pixels _CT ×slopes+intercepts

Among them, Pixels _PET is the pixel value of the PET sequence image, LBM is the lean body mass, injected dose is the tracer injection dose; Pixels _CT is the pixel value of the CT sequence image, slopes is the slope, and intercepts is the intercept;

Step 2.2: Perform image contrast enhancement processing on two-dimensional PET sequence images and two-dimensional CT sequence images, adjust the window width and window level of the hu value to [a ₁ , b ₁ ], and limit the SUV value to [a ₂ , b ₂ ] ; Among them, a ₁ , b ₁ , a ₂ and b ₂ are all constants;

Step 2.3: Adjust the image resolution of the two-dimensional CT sequence image, and adjust the size of the 512×512 two-dimensional CT sequence image to the size of the two-dimensional PET sequence image H×W=128×128;

Step 2.4: Generate three-dimensional volume data [H, W, D _{PET, i} ] and [H, W, D _{CT, i} ] for the two-dimensional PET sequence image and two-dimensional CT sequence image of the i-th patient respectively, and convert the three-dimensional volume The data is transformed into five-dimensional volume data [N, H, W, D _i , C], and the five-dimensional volume data is cut in the Z direction with d pixels as the sampling interval to generate multiple pairs of image blocks of size H×W×D , normalize the image blocks to obtain an image block set, randomly select l pairs of PET image blocks and CT image blocks from the image block set to form a PET/CT image block training set; where, i∈{1,2,..., m}, D _{PET, i} is the slice number of the PET sequence image of the i-th patient, D _{CT, i} is the slice number of the CT sequence image of the i-th patient, D _{PET, i} =D _{CT, i} =D _i ; N is the batch size in batch processing, C is the channel number of input neural network data, C=2.

In the step 2, [a ₁ , b ₁ ]=[-90,300], [a ₂ , b ₂ ]=[0,5], d=32, D=64.

使图像块的数据变为均值为0且标准差为1的正态分布；其中，x、x^*分别为图像块中归一化处理前、后的像素点，μ、σ分别为图像块中所有像素点的均值、标准差。In the step 2.4, the formula for normalizing the image block is

Make the data of the image block become a normal distribution with a mean value of 0 and a standard deviation of 1; among them, x, x ^* are the pixels before and after normalization processing in the image block, respectively, and μ, σ are the pixels in the image block, respectively. The mean and standard deviation of all pixels.

In the step 3, building a PET/CT registration network based on the 3DU-Net convolutional neural network includes an encoding path and a decoding path, and each path has 4 resolution levels; the encoding path has n ₁ layers, and the Each layer of the encoding path includes a convolutional layer with a convolution kernel of 3×3×3 and a step size of 2, and each convolutional layer is followed by a BN layer and a ReLU layer; the decoding path has n ₂ Layer, each layer of the decoding path includes a deconvolution layer with a convolution kernel of 3×3×3 and a step size of 2, and each deconvolution layer is followed by a BN layer and a ReLU layer; by shortcut, transfer the same resolution layer in the encoding path to the decoding path, and provide the original high-resolution features for the decoding path; the last layer of the PET/CT registration network is a 3×3×3 convolutional layer, Finally, the number of output channels is 3.

In the step 4, in combination with the image similarity constraint term and the excessive deformation suppression term, the cost function of constructing the PET/CT registration network is

为均值为μ、标准差为θ的高斯分布函数，λ为正则化项权重；Among them, F and M are CT image blocks and PET image blocks respectively, D _v is the displacement vector field matrix,

is a Gaussian distribution function with a mean of μ and a standard deviation of θ, and λ is the weight of the regularization term;

为相似性约束项

is a similarity constraint

Among them, S is the submap, T is the template image, (s, t) is the coordinate index, S(s, t) is the pixel value of the submap, T(s, t) is the pixel value of the template image, E(S ), E(T) are the average gray value of sub-image and template image respectively;

为过度形变抑制项

Excessive deformation suppression term

Among them, i is the element in the displacement vector field matrix _Dv , j is a random number following the Gaussian distribution function, f(i, j, θ) is a penalty item,

The beneficial effects of the present invention are:

The present invention builds a PET/CT registration network based on the 3DU-Net convolutional neural network, predicts the displacement vector field through the unsupervised end-to-end 3D elastic registration neural network based on deep learning, and performs voxel-by-voxel displacement prediction on the image to be registered , and using the normalized cross-correlation as the similarity constraint item, combined with the suppression of image deformation by the excessive deformation suppression item, the cost function of the PET/CT registration network is constructed, which can solve the registration problem caused by the low similarity of the PET/CT image itself. The quasi-excessive deformation problem can realize PET/CT elastic registration, improve registration efficiency and accuracy, and reduce the computational cost of excessive deformation suppression.

Description of drawings

Fig. 1 is the flow chart of the PET and CT image registration method based on excessive deformation suppression of the present invention;

Fig. 2 is a schematic structural diagram of the PET/CT registration network in the PET and CT image registration method based on excessive deformation suppression of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

As shown in FIG. 1 , it is a flow chart of the PET and CT image registration method based on excessive deformation suppression of the present invention. The PET and CT image registration method based on excessive deformation suppression of the present invention is characterized in that it comprises the following steps:

Step 1: Acquire 2D PET sequence images and 2D CT sequence images of m patients to obtain a PET/CT sequence image set.

Step 2: Preprocessing the PET/CT sequence image set to obtain a PET/CT image block training set; the preprocessing includes calculating the SUV value and hu value and limiting the threshold range, adjusting image resolution, generating image blocks and performing normalization One treatment.

Described step 2 comprises the following steps:

Step 2.1: Calculate the SUV value of the two-dimensional PET sequence image SUV=Pixels _PET ×LBM×1000/injecteddose

Calculate the hu value of the two-dimensional CT sequence image Hu=Pixels _CT ×slopes+intercepts

Among them, Pixels _PET is the pixel value of the PET sequence image, LBM is the lean body mass, injected dose is the tracer injection dose; Pixels _CT is the pixel value of the CT sequence image, slopes is the slope, and intercepts is the intercept;

Step 2.2: Perform image contrast enhancement processing on two-dimensional PET sequence images and two-dimensional CT sequence images, adjust the window width and window level of the hu value to [a ₁ , b ₁ ], and limit the SUV value to [a ₂ , b ₂ ] ; Among them, a ₁ , b ₁ , a ₂ and b ₂ are all constants;

Step 2.3: Adjust the image resolution of the two-dimensional CT sequence image, and adjust the size of the 512×512 two-dimensional CT sequence image to the size of the two-dimensional PET sequence image H×W=128×128;

Step 2.4: Generate three-dimensional volume data [H, W, D _{PET, i} ] and [H, W, D _{CT, i} ] for the two-dimensional PET sequence image and two-dimensional CT sequence image of the i-th patient respectively, and convert the three-dimensional volume The data is transformed into five-dimensional volume data [N, H, W, D _i , C], and the five-dimensional volume data is cut in the Z direction with d pixels as the sampling interval to generate multiple pairs of image blocks of size H×W×D , normalize the image blocks to obtain an image block set, randomly select l pairs of PET image blocks and CT image blocks from the image block set to form a PET/CT image block training set; where, i∈{1,2,..., m}, D _{PET, i} is the slice number of the PET sequence image of the i-th patient, D _{CT, i} is the slice number of the CT sequence image of the i-th patient, D _{PET, i} =D _{CT, i} =D _i ; N is the batch size in batch processing, C is the channel number of input neural network data, C=2.

In this embodiment, m=176; the processed SUV and Hu value images of PET and CT are respectively generated into three-dimensional volume data and stored in ndarray; in the step 2, [a ₁ , b ₁ ]=[-90,300 ], [a ₂ , b ₂ ]=[0,5], d=32, D=64.

For all 176 patients, randomly select l=900 pairs of SUV and Hu value image blocks generated by the volume data of 141 patients as the PET/CT image block training set, and select 259 pairs of SUV, Hu value generated by the volume data of 35 patients among them. Hu-valued image patches are used as the validation set.

使图像块的数据变为均值为0且标准差为1的正态分布；其中，x、x^*分别为图像块中归一化处理前、后的像素点，μ、σ分别为图像块中所有像素点的均值、标准差。In the step 2.4, the formula for normalizing the image block is

Make the data of the image block become a normal distribution with a mean value of 0 and a standard deviation of 1; among them, x, x ^* are the pixels before and after normalization processing in the image block, respectively, and μ, σ are the pixels in the image block, respectively. The mean and standard deviation of all pixels.

Step 3: Construct a PET/CT registration network based on the 3DU-Net convolutional neural network, as shown in Figure 2.

The PET/CT registration network of the present invention includes: (1) 3D U-Net for regressing the displacement vector field; (2) component spatial transformer (Spatial Transformer) for spatial transformation. In the present embodiment, in the step 3, the PET/CT registration network based on the 3DU-Net convolutional neural network construction includes an encoding path and a decoding path, and each path has 4 resolution levels; the encoding path has n ₁ layer, each layer of the encoding path includes a convolution kernel with a convolution kernel of 3×3×3 and a step size of 2, and each convolution layer is followed by a BN layer and a ReLU layer; the The decoding path has n ₂ layers, and each layer of the decoding path includes a deconvolution layer with a convolution kernel of 3×3×3 and a step size of 2, and each deconvolution layer is followed by a BN layer and ReLU layer; through shortcut, the same resolution layer in the encoding path is passed to the decoding path to provide the original high-resolution features for the decoding path; the last layer of the PET/CT registration network is 3×3×3 The convolutional layer, the final output channel number is 3.

Among them, the BN layer is a batch normalization layer, the ReLU layer is a rectified linear unit layer, and the shortcut is a skip connection. The last layer is a 3×3×3 convolutional layer, which reduces the number of output channels, and the final output channel number is 3 (that is, the three directions of x, y, and z).

Step 4: Construct the cost function of the PET/CT registration network by combining the image similarity constraint item and the excessive deformation suppression item; the similarity constraint item is normalized cross-correlation NCC, and the excessive deformation suppression item is based on the displacement vector Sum of the penalty term for the difference between the field elements and the Gaussian distribution function elements.

Among them, the "excessive deformation suppression" measure is defined based on the deformation degree of the 3D deformation field, and the "excessive deformation suppression term" is introduced in the cost function to optimize the registration network.

In the step 4, in combination with the image similarity constraint term and the excessive deformation suppression term, the cost function of constructing the PET/CT registration network is

为均值为μ、标准差为θ的高斯分布函数，λ为正则化项权重；其中，F、M分别为固定图像块、浮动图像块；Among them, F and M are CT image blocks and PET image blocks respectively, D _v is the displacement vector field matrix,

is a Gaussian distribution function with mean μ and standard deviation θ, and λ is the weight of the regularization term; where F and M are fixed image blocks and floating image blocks respectively;

为相似性约束项

is a similarity constraint

Among them, S is the submap, T is the template image, (s, t) is the coordinate index, S(s, t) is the pixel value of the submap, T(s, t) is the pixel value of the template image, E(S ), E(T) are the average gray value of sub-image and template image respectively;

为过度形变抑制项

Excessive deformation suppression term

本实施例中，根据经验k被设置为2，当|i-j|＞θ时，惩罚项为|i-j|^k，即通过k次幂放大惩罚项。Among them, i is the element in the displacement vector field matrix _Dv , j is a random number following the Gaussian distribution function, f(i, j, θ) is a penalty item,

In this embodiment, k is set to 2 based on experience, and when |ij|>θ, the penalty term is |ij| ^k , that is, the penalty term is amplified by the power of k.

Step 5: Use the global variable initialization (global_variables_initializer) to initialize the weight parameters of the neural network, set the batch size in batch processing N=16, the regularization item weight λ=0.3, the maximum number of iterations COUNT=1000, the network learning rate to 0.001, optimize The controller is Adam, and the learning rate decay strategy is adopted.

Step 6: Use the PET/CT image block training set as the input of the PET/CT registration network, output the displacement vector field, input the displacement vector field and the PET image block together to the space transformer, and obtain the registered PET image block, According to the CT image block and the registered PET image block, the similarity constraint item is obtained, and the excessive deformation suppression item is obtained according to the displacement vector field, so as to update the weight parameters of the neural network through the cost function, and perform backpropagation, thus PET/CT The registration network is iteratively trained until the maximum number of iterations is COUNT, and the trained PET/CT registration network is obtained.

Among them, a pair of PET/CT image blocks with a size of 128×128×64 are used as the input of the 3D U-Net network, and a displacement vector field of the same resolution size (128×128×64×3, corresponding to x, y, Z-direction displacement), the displacement vector field and the PET image block are jointly input to the space transformer, and the registered PET image block is output.

Step 7: Perform the preprocessing described in step 1 on the PET/CT image pair to be registered, and input the obtained PET/CT image block pair into the trained PET/CT registration network to generate a registered PET image block and visualize it.

In this embodiment, the PET and CT image registration method based on excessive deformation suppression of the present invention runs in the Windows 10 system environment of the Intel kernel, and performs medical image registration based on the Python and Tensorflow framework. The image registration algorithm based on deep learning adopted by the present invention converts image registration into a regression problem of displacement vector field, namely predicting the spatial correspondence between pixels/voxels from a pair of images. Image registration is achieved by simultaneously optimizing the similarity constraint term and the displacement vector field excessive deformation suppression term between fixed and floating image pairs through 3D U-Net convolutional neural network. Quantitative and qualitative results show that using the registration method of the present invention to perform 3D PET/CT image registration achieves good results. Among them, for the trained model, given a pair of new PET/CT volume data, the registration result can be obtained through one forward calculation within 10 seconds.

Apparently, the above-mentioned embodiments are only some of the embodiments of the present invention, but not all of them. The above-mentioned embodiments are only used to explain the present invention, and do not constitute a limitation to the protection scope of the present invention. Based on the above-mentioned embodiments, all other embodiments obtained by those skilled in the art without creative work, that is, all modifications, equivalent replacements and improvements made within the spirit and principles of this application are all Fall within the scope of protection required by the present invention.

Claims (5)

1. A PET and CT image registration method based on excessive deformation suppression, is characterized in that, comprises the steps: Step 1: Collect 2D PET sequence images and 2D CT sequence images of m patients to obtain a PET/CT sequence image set; Step 2: Preprocessing the PET/CT sequence image set to obtain a PET/CT image block training set; the preprocessing includes calculating the SUV value and hu value and limiting the threshold range, adjusting image resolution, generating image blocks and performing normalization One treatment; Step 3: Construct PET/CT registration network based on 3D U-Net convolutional neural network; Step 4: Construct the cost function of the PET/CT registration network by combining the image similarity constraint item and the excessive deformation suppression item; the similarity constraint item is normalized cross-correlation NCC, and the excessive deformation suppression item is based on the displacement vector The sum of the penalty term for the difference between the field element and the Gaussian distribution function element; The cost function of constructing the PET/CT registration network by combining the image similarity constraint term and the excessive deformation suppression term is:

Among them, F and M are CT image blocks and PET image blocks respectively, D _v is the displacement vector field matrix,

is a Gaussian distribution function with a mean of μ and a standard deviation of θ, and λ is the weight of the regularization term;

is a similarity constraint

Among them, S is the subimage, T is the template image, (s,t) is the coordinate index, S(s,t) is the pixel value of the subimage, T(s,t) is the pixel value of the template image, E(S ), E(T) are the average gray value of sub-image and template image respectively;

Excessive deformation suppression term

Among them, i is the element in the displacement vector field matrix D _v , and j is the Gaussian distribution function

random number, f(i,j,θ) is the penalty item,

Step 5: Initialize the weight parameters of the neural network, set the batch size N in the batch processing, the regularization item weight λ, the maximum number of iterations COUNT, the network learning rate, the optimizer, and adopt the learning rate decay strategy; Step 6: Use the PET/CT image block training set as the input of the PET/CT registration network, output the displacement vector field, input the displacement vector field and the PET image block together into the space transformer, and obtain the registered PET image block, According to the CT image block and the registered PET image block, the similarity constraint item is obtained, and the excessive deformation suppression item is obtained according to the displacement vector field, so as to update the weight parameters of the neural network through the cost function, and perform backpropagation, thus PET/CT The registration network is iteratively trained until the maximum number of iterations is COUNT, and the trained PET/CT registration network is obtained; Step 7: Perform the preprocessing described in step 1 on the PET/CT image pair to be registered, and input the obtained PET/CT image block pair into the trained PET/CT registration network to generate a registered PET image block and visualize it. 2. The PET and CT image registration method based on excessive deformation suppression according to claim 1, wherein said step 2 comprises the following steps: Step 2.1: Calculate the SUV value of the two-dimensional PET sequence image SUV=Pixels _PET ×LBM×1000/injecteddose Calculate the hu value of the two-dimensional CT sequence image Hu=Ptxels _CT ×slOpeS+interCepts Among them, Pixels _PET is the pixel value of the PET sequence image, LBM is the lean body mass, injecteddose is the tracer injection dose; Pixels _CT is the pixel value of the CT sequence image, slopes is the slope, and intercepts is the intercept; Step 2.2: Perform image contrast enhancement processing on two-dimensional PET sequence images and two-dimensional CT sequence images, adjust the window width and window level of the hu value to [a ₁ , b ₁ ], and limit the SUV value to [a ₂ , b ₂ ] ; Among them, a ₁ , b ₁ , a ₂ and b ₂ are all constants; Step 2.3: Adjust the image resolution of the two-dimensional CT sequence image, and adjust the size of the 512×512 two-dimensional CT sequence image to the size of the two-dimensional PET sequence image H×W=128×128; Step 2.4: Generate 3D volume data [H,W,D _PET,i ] and [H,W,D _CT,i ] for the 2D PET sequence image and 2D CT sequence image of the i-th patient respectively, and convert the 3D volume The data is transformed into five-dimensional volume data [N, H, W, D _i , C], and the five-dimensional volume data is cut in the Z direction with d pixels as the sampling interval, and multiple pairs of image blocks of H×W×D size are generated , normalize the image blocks to obtain an image block set, randomly select l pairs of PET image blocks and CT image blocks from the image block set to form a PET/CT image block training set; where, i∈{1,2,..., m}, D _{PET, i} is the number of slices of the PET sequence image of the i-th patient, D _{CT, i} is the slice number of the CT sequence image of the i-th patient, D _{PET, i} = D _{CT, i} = D _i ; N is the batch size in batch processing, C is the channel number of input neural network data, C=2. 3. The PET and CT image registration method based on excessive deformation suppression according to claim 2, characterized in that, in the step 2.2, [a ₁ , b ₁ ]=[-90,300], [a ₂ , b ₂ ]=[0,5], d=32, D=64. 4. The PET and CT image registration method based on excessive deformation suppression according to claim 2, characterized in that, in the step 2.4, the formula for normalizing the image block is

Make the data of the image block become a normal distribution with a mean of 0 and a standard deviation of 1; where x, x* are the pixels before and after normalization in the image block, respectively, and μ, σ are the pixels in the image block, respectively. The mean and standard deviation of all pixels. 5. The PET and CT image registration method based on excessive deformation suppression according to any one of claims 2 to 4, characterized in that, in the step 3, the PET/CT image registration method is constructed based on the 3D U-Net convolutional neural network. The CT registration network includes an encoding path and a decoding path, each of which has 4 resolution levels; the encoding path has n ₁ layers, and each layer of the encoding path includes a convolution kernel of 3×3× 3. A convolutional layer with a step size of 2, each convolutional layer is followed by a BN layer and a ReLU layer; the decoding path has n ₂ layers, and each layer of the decoding path includes a convolution kernel as 3×3×3 deconvolution layers with a step size of 2, each deconvolution layer is followed by a BN layer and a ReLU layer; through shortcut, the same resolution layer in the encoding path is passed to the decoding path, as The decoding path provides the original high-resolution features; the last layer of the PET/CT registration network is a 3×3×3 convolutional layer, and the number of final output channels is 3.

Images