CT / PET image registration method based on multi-scale features and progressive correlation correction

Abstract

The invention discloses a CT (Computed Tomography) / PET (Positron Emission Tomography) image registration method based on multi-scale characteristics and progressive correlation correction. Firstly, a PET image and a CT image are obtained; then, constructing a registration model consisting of an encoder and a decoder; the encoder is used for extracting multi-scale feature maps of a PET image and a CT image and comprises a convolution block and four feature encoding stages, and each feature encoding stage comprises a 3DConvNeXt module and a PMFA module; the decoder decodes and registers the multi-scale feature map extracted by the encoder and comprises four feature decoding stages, the first feature decoding stage comprises an AFMLP module, and each of the second to fourth feature decoding stages comprises two AFMLP modules; and finally, training the registration model, and using the trained registration model for registration of the CT image and the PET image. According to the method, hierarchical feature extraction from coarse to fine and progressive iterative optimization registration from coarse to fine are realized, and the precision and efficiency of image registration are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of medical image registration technology, specifically a CT / PET image registration method based on multi-scale features and progressive correlation correction. Background Technology

[0002] In the clinical diagnosis of myocardial ischemia, positron emission tomography (PET) images can display information on myocardial blood perfusion and metabolic function, while computed tomography (CT) images provide detailed anatomical information. Therefore, the combined application of PET and CT images provides an important means for the accurate assessment of myocardial ischemia. To accurately fuse the image information from these two modalities and achieve precise alignment of functional and anatomical information, image registration becomes a crucial step. Registration accuracy is an important prerequisite for the diagnosis and quantitative description of myocardial ischemia. However, in current clinical practice, multimodal image registration mainly relies on manual work by physicians, which not only consumes a significant amount of manpower and time but is also prone to introducing subjective errors. Therefore, utilizing artificial intelligence technology to assist physicians in completing registration tasks is of great significance for improving the efficiency and accuracy of clinical diagnosis.

[0003] In recent years, deep learning-based image registration methods have made significant progress, but still face a series of challenges. First, traditional convolutional neural networks (CNNs), due to their inherent limitations in local receptive fields, struggle to effectively capture long-range dependencies in images, particularly when dealing with complex anatomical structures. For example, while the U-Net architecture used in the VoxelMorph method integrates information through multi-scale fusion, its core convolutional operations remain local, potentially making it difficult to directly and efficiently model long-range spatial correspondences when faced with large-scale organ displacements. Second, although models like the Transformer can capture global contextual information through self-attention mechanisms, their enormous computational complexity typically forces them to operate on low-resolution feature maps, inevitably leading to the loss of fine-grained key details in the image. For instance, the TransMorph method, to control computational costs, first significantly downsamples high-resolution images before inputting them into the Transformer for processing, resulting in the loss of fine anatomical structural information in the early stages, thus affecting the final registration accuracy. Third, many existing models suffer from low feature extraction efficiency and a large number of parameters, a problem particularly pronounced in networks employing multi-stage or cascaded architectures. For example, Recursive Cascaded Networks decomposes a complex registration task into multiple steps of "coarse adjustment followed by fine adjustment," performed sequentially by a series of CNN networks. While this strategy may improve accuracy, it essentially strings multiple networks together, leading to a significant increase in the total number of model parameters and computational complexity, thus substantially increasing the training burden. Fourth, many mainstream methods lack explicit modeling of the spatial correspondence between moving and stationary images, limiting their registration performance in complex deformation scenarios. For instance, the VoxelMorph method directly feeds two concatenated input images into the network for "black box" prediction of the deformation field. The network must implicitly learn the correspondence between the two images without a specific module to explicitly compute and evaluate the matching degree between different locations, which is particularly challenging in sparsely textured or repetitive regions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a CT / PET image registration method based on multi-scale features and progressive correlation correction.

[0005] The present invention solves the aforementioned technical problem by adopting the following technical solution:

[0006] A CT / PET image registration method based on multi-scale features and progressive correlation correction, characterized by the following steps:

[0007] Step 1: Acquire PET and CT images;

[0008] Step 2: Construct the registration model, including the encoder and decoder;

[0009] The encoder includes a convolutional block and four feature encoding stages. The output feature map of the previous feature encoding stage is downsampled and used as the input feature map of the next feature encoding stage. The first feature encoding stage outputs a first-scale PET feature map and a first-scale CT feature map. The second feature encoding stage outputs a second-scale PET feature map and a second-scale CT feature map. The third feature encoding stage outputs a third-scale PET feature map and a third-scale CT feature map. The fourth feature encoding stage outputs a fourth-scale PET feature map and a fourth-scale CT feature map.

[0010] The decoder comprises four feature decoding stages. The first feature decoding stage includes one AFMLP module, while the second to fourth feature decoding stages each include two AFMLP modules. The fourth-scale PET feature map and the fourth-scale CT feature map are processed by the first feature decoding stage to obtain the output feature map of the first feature decoding stage. The output feature map of the first feature decoding stage is upsampled and then input into the second feature decoding stage to perform spatial registration on the third-scale PET feature map, resulting in a spatially registered third-scale PET feature map. The spatially registered third-scale PET feature map and the third-scale CT feature map are processed by the first AFMLP module. The output feature map of the first AFMLP module is compared with the output feature map of the upsampled first feature decoding stage. The output feature map is input into the second AFMLP module for processing. The output feature map of the second AFMLP module is concatenated with the output feature map of the first feature decoding stage after upsampling to obtain the output feature map of the second feature decoding stage. The output feature map of the second feature decoding stage is then upsampled and input into the third feature decoding stage for processing, and the same process is repeated to obtain the output feature map of the third feature decoding stage. The output feature map of the third feature decoding stage is then upsampled and input into the fourth feature decoding stage for processing, and the same process is repeated to obtain the output feature map of the fourth feature decoding stage. Finally, the output feature map of the fourth feature decoding stage is applied to the PET image to perform spatial registration on the PET image, generating a registered PET image aligned with the CT image.

[0011] In the AFMLP module, the spatially registered PET and CT feature maps undergo co-correlation calculation to obtain a correlation feature map. The correlation feature map, the spatially registered PET feature map, and the CT feature map are then concatenated along the channel dimension to obtain a composite feature map. This composite feature map is compressed and layer-normalized, then passed through a three-scale gated multilayer perceptron to generate three feature maps. These three feature maps are then concatenated and dynamically weighted to obtain importance weights for each feature map. These importance weights are multiplied by their corresponding feature maps to obtain channel-enhanced feature maps. The three channel-enhanced feature maps are then summed and residually concatenated with the compressed composite feature map. The resulting feature map is then passed through a residual channel attention module and residually concatenated with the compressed composite feature map to obtain the output feature map of the AFMLP module.

[0012] Step 3: Train the registration model and use the trained registration model for registration of CT images and PET images.

[0013] Furthermore, each feature encoding stage includes a PMFA module. In the PMFA module, the input feature map is processed through five branches for feature extraction. The output feature maps of the five branches are concatenated along the channel dimension. The concatenated feature maps are then calibrated using both channel attention and spatial attention mechanisms to obtain channel- and spatially calibrated feature maps. The channel- and spatially calibrated feature maps are then fused using an element-wise maximum value operation to obtain an enhanced feature map. The enhanced feature map is then subjected to a 1×1×1 convolution to obtain the output feature map.

[0014] Furthermore, each feature encoding stage includes a 3DConvNeXt module; in the 3DConvNeXt module, the input feature map is sequentially passed through depthwise separable convolution, batch normalization and GELU activation function, and then residually connected with itself. The feature map obtained by the residual connection is sequentially passed through two pointwise convolutional blocks and one convolutional block to obtain the output feature map.

[0015] Furthermore, the collaborative correlation calculation process is as follows: first, zero-padding of 1 unit is performed on the CT feature map in the height, width and depth directions to generate multiple spatial offsets; at each offset position, a region with the same size as the CT feature map is extracted, and the region is multiplied element-wise with the CT feature map, and then the mean is calculated in the channel dimension to form a correlation feature map.

[0016] Furthermore, the dynamic weighted fusion includes a series of global average pooling layers, a multilayer perceptron, a softmax activation function, and a linear layer.

[0017] Furthermore, in the residual channel attention module, the input feature map is sequentially processed through convolution, Leaky ReLU activation function, convolution and channel attention mechanism, and then residually connected to itself to obtain the output feature map.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. The PMFA module utilizes multi-branch parallel dilated convolution (including four 3×3×3 convolution branches with different dilation rates and a global average pooling branch) to perform deep processing and enhancement of features while maintaining the spatial resolution of the input features. This effectively captures multi-scale features from local details to global contextual information. After fusing multi-scale features through channel-dimensional concatenation, it further combines channel attention and spatial attention mechanisms for adaptive feature calibration and employs element-wise maximum fusion. This significantly enhances the representation ability of anatomical structure boundaries, functional regions, and key details in PET and CT images. It solves the problems of limited receptive fields in traditional convolutional neural networks and the high computational complexity of existing Transformer networks, which are forced to process low-resolution features and struggle to capture fine-grained long-distance dependencies at full resolution. This improves the discriminativeness and robustness of cross-modal feature representation, thereby enhancing registration accuracy.

[0020] 2. The 3DConvNeXt module achieves deep feature extraction and enhancement through efficient spatial feature extraction of depthwise separable convolution, gradient optimization of residual connections, and channel information fusion and transformation of pointwise convolution. At the same time, it keeps the spatial size and number of channels of input and output unchanged, ensuring the consistency of feature transmission and effectively reducing parameter complexity, providing stable and semantically rich feature representations for subsequent layers.

[0021] 3. The AFMLP module utilizes a multi-window gating mechanism and channel attention enhancement to explicitly calculate the local spatial correlation between moving and stationary images on a single medium-resolution feature map through a correlation layer. It also dynamically fuses multi-scale contextual information through three parallel windows of different sizes to efficiently model long-distance dependencies at full resolution. At the same time, it uses channel attention to filter key cross-modal features and combines residual connections to ensure information integrity. Finally, it achieves high-precision registration of CT / PET images through progressive upsampling, effectively overcoming the problems of traditional CNNs lacking explicit spatial modeling and Transformers losing detailed information due to being forced to run on low-resolution feature maps due to computational complexity.

[0022] 4. Through the synergistic effect of PMFA, 3DConvNeXt, and AFMLP modules, a multi-level (four-level), multi-scale (from full resolution to 1 / 8 resolution) registration model was constructed, achieving hierarchical feature extraction from coarse to fine. Within each level, the PMFA module utilizes multi-branch dilated convolution and global pooling to capture multi-scale features from local to global without changing the spatial resolution, thereby enabling in-depth analysis of information such as myocardial boundaries and perfusion zones. The decoder starts from the coarsest level. First, the AFMLP module calculates the correlation feature map to explicitly model the spatial correspondence between PET and CT feature maps. Then, it samples up level by level. At each finer level, using features from the encoder and the deformation field predicted by the previous level, the deformation field is optimized and corrected again by the AFMLP module. This progressive iterative optimization registration from coarse to fine achieves accurate alignment. Therefore, the method of the present invention significantly improves the accuracy and efficiency of PET and CT image registration. Especially in complex myocardial image registration tasks, it can perform hierarchical analysis and precise alignment of key information such as myocardial structural boundaries and perfusion areas, providing a reliable basis for the quantitative diagnosis of myocardial ischemia, effectively assisting clinical decision-making and reducing the burden of manual registration for doctors. Attached Figure Description

[0023] Figure 1 This is a structural diagram of the encoder of the registration model of the present invention;

[0024] Figure 2 This is a decoder structure diagram of the registration model of the present invention;

[0025] Figure 3 This is a structural diagram of the PMFA module of the present invention;

[0026] Figure 4 This is a structural diagram of the 3DConvNeXt module of the present invention;

[0027] Figure 5 This is a structural diagram of the AFMLP module of the present invention. Detailed Implementation

[0028] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to further illustrate the technical solution of the present invention in detail, and are not intended to limit the scope of protection of this application.

[0029] This invention provides a CT / PET image registration method based on multi-scale features and progressive correlation correction (see [link]). Figures 1-5 This includes the following steps:

[0030] Step 1: Acquire PET and CT images;

[0031] Step 2: Construct a registration model, including an encoder and a decoder; the encoder is used to extract multi-scale feature maps from PET and CT images, and the decoder decodes and registers the multi-scale feature maps extracted by the encoder.

[0032] like Figure 1 As shown, the encoder includes a convolutional block (comprising a cascaded 3×3×3 convolution, a ReLU activation function layer, and batch normalization) and four feature encoding stages. The convolutional block performs preliminary processing on the input image, and the four feature encoding stages extract multi-scale feature maps from the input image. The output feature map of the previous feature encoding stage is downsampled and used as the input feature map for the next feature encoding stage. Each feature encoding stage includes a 3DConvNeXt module and a pyramid multi-scale fusion attention (PMFA) module; the PET image I with dimensions H×W×D (H, W, and D represent length, width, and height, respectively) and 1 channel is used. m and CT images I f The initial PET feature map is obtained by processing it through convolutional blocks to increase the number of channels. and initial CT feature map Initial PET feature map and initial CT feature map After the first feature encoding stage, the first-scale PET feature map is obtained. and first-scale CT feature map Its dimensions are H×W×D, and it has 8 channels; first-scale PET feature map and first-scale CT feature map After downsampling, a second feature encoding stage is performed to obtain the second-scale PET feature map. Second-scale CT feature map Its dimensions are H / 2×W / 2×D / 2, and the number of channels is 16; similarly, the third feature encoding stage outputs a third-scale PET feature map. and third-scale CT feature maps Its dimensions are H / 4×W / 4×D / 4, and the number of channels is 32; the fourth feature encoding stage outputs a fourth-scale PET feature map. and fourth-scale CT feature maps Its dimensions are H / 8×W / 8×D / 8, and it has 64 channels.

[0033] like Figure 3 As shown, in the 3DConvNeXt module, the input feature map sequentially passes through depthwise separable convolution, batch normalization, and the GELU activation function, and then is residually concatenated with itself. The feature map obtained from the residual concatenation is then sequentially passed through two pointwise convolutional blocks and one convolutional block to obtain the output feature map of the 3DConvNeXt module. For example, the initial PET feature map... and initial CT feature map After the first feature encoding stage, the 3DConvNeXt module, the intermediate PET feature map is obtained. and intermediate CT feature map The input is then fed into the PMFA module of the first feature encoding stage. The pointwise convolutional block includes pointwise convolution, batch normalization, and the GELU activation function connected in sequence.

[0034] like Figure 4 As shown, in the PMFA module, multi-scale feature extraction is first implemented using five parallel branches. The first branch directly extracts the input feature map using a 1×1×1 convolutional kernel, maintaining the spatial scale. The second branch uses a 3×3×3 convolutional kernel with a dilation rate of 6 to appropriately expand the receptive field. The third branch uses a 3×3×3 convolutional kernel with a dilation rate of 12 to further expand the receptive field and capture broader contextual information. The fourth branch uses a 3×3×3 convolutional kernel with a dilation rate of 18 to provide the maximum receptive field. The fifth branch uses global averaging. Pooling is used to extract global contextual features, enhancing the model's understanding of the overall image layout. Subsequently, the multi-scale feature maps extracted from the five branches are concatenated along the channel dimension. The concatenated feature maps are then calibrated using channel attention and spatial attention mechanisms, resulting in channel-calibrated and spatially calibrated feature maps. The channel-calibrated and spatially calibrated feature maps are fused using an element-wise maximum value operation to generate an enhanced feature map. This enhanced feature map is then subjected to dimensionality reduction and feature integration using a 1×1×1 convolution to obtain the output feature map of the PMFA module.

[0035] like Figure 2 As shown, the decoder includes four feature decoding stages. The first feature decoding stage includes one AFMLP module, and the second to fourth feature decoding stages each include two AFMLP modules; the fourth-scale PET feature map... and fourth-scale CT feature maps After the first feature decoding stage, the output feature map of the first feature decoding stage is obtained. Output feature map of the first feature decoding stage After upsampling, the data is input into the second feature decoding stage, along with the third-scale PET feature map. After spatial transformation, the third-scale PET feature map Spatial registration was performed to obtain the third-scale PET feature map after spatial registration; the third-scale PET feature map and the third-scale CT feature map after spatial registration were also obtained. After processing by the first AFMLP module, the output feature map of the first AFMLP module and the upsampled output feature map of the first feature decoding stage are input into the second AFMLP module for processing. The output feature map of the second AFMLP module is then concatenated with the upsampled output feature map of the first feature decoding stage to obtain the output feature map of the second feature decoding stage. Output feature map of the second feature decoding stage After upsampling, the data is input into the third feature decoding stage, along with the second-scale PET feature map. After spatial transformation, the second-scale PET feature map Spatial registration was performed to obtain a second-scale PET feature map after spatial registration; the second-scale PET feature map and the second-scale CT feature map after spatial registration were also obtained. After processing by the first AFMLP module, the output feature map of the first AFMLP module and the upsampled output feature map of the second feature decoding stage are processed by the second AFMLP module of the third feature decoding stage. The output feature map of the second AFMLP module of the third feature decoding stage is then concatenated with the upsampled output feature map of the second feature decoding stage to obtain the output feature map of the third feature decoding stage. The output feature map from the third feature decoding stage is upsampled and then input into the fourth feature decoding stage for processing. Similarly, the output feature map of the fourth feature decoding stage is obtained. Finally, the output feature map of the fourth feature decoding stage is... Effect on PET image I m After spatial transformation, a registered PET image aligned with the CT image is generated.

[0036] The spatial transformation operation involves using the upsampled feature map as a displacement field to deform the PET feature map. Specifically, firstly, a regular initial coordinate grid is created based on the spatial dimensions of the PET feature map. This grid records the initial position of each pixel or voxel in the PET feature map. Then, the upsampled feature map is added to the initial coordinate grid to obtain a new set of coordinate positions. This set indicates the new position where each point in the PET feature map should be sampled. Finally, the PET feature map is resampled based on the new set of coordinate positions, and bilinear interpolation is used to handle the sampling problem of non-integer coordinates, generating a spatially registered PET feature map.

[0037] like Figure 5As shown, in the AFMLP module, the spatially registered PET and CT feature maps are processed through a correlation layer to calculate their co-correlation, resulting in a correlation feature map. The correlation feature map, the spatially registered PET feature map, and the CT feature map are then concatenated along the channel dimension to obtain a composite feature map. This composite feature map is compressed using a 3×3×3 convolutional kernel with a stride of 1, resulting in a compressed composite feature map. After layer normalization, the compressed composite feature map undergoes multi-window gating processing, using gated multilayer perceptrons (gMLPs) of three scales: 3×3×3, 5×5×5, and 7×7×7. Three feature maps G1, G2, and G3 with 128 channels and the same size are obtained. After concatenation, feature maps G1, G2, and G3 are assigned importance weights to each feature map through a dynamic weighted fusion module, resulting in importance weights ω1, ω2, and ω3. The importance weights are multiplied by the corresponding feature maps to obtain channel-enhanced feature maps. The three channel-enhanced feature maps are added together and then residually concatenated with the compressed composite feature map. The feature map obtained by the residual concatenation is then passed through the Residual Channel Attention (RCAM) module and then residually concatenated with the compressed composite feature map to obtain the output feature map of the AFMLP module.

[0038] The collaborative correlation calculation process is as follows: First, zero-padding of 1 unit is applied to the CT feature map in the height, width and depth directions to generate 27 spatial offsets (covering a 3×3×3 cubic neighborhood). At each offset position, a region with the same size as the CT feature map is extracted. This region is then multiplied element-wise with the CT feature map, and the mean is calculated in the channel dimension to form a correlation feature map containing 27 spatial offset relationships.

[0039] The dynamic weighted fusion module includes a series of global average pooling layers, a multilayer perceptron, a Softmax activation function, and a linear layer.

[0040] In the Residual Channel Attention (RCAM) module, the input feature map is sequentially processed through convolution, Leaky ReLU activation function, convolution and channel attention mechanism, and then residually connected to itself to obtain the output feature map of the Residual Channel Attention module.

[0041] Step 3: Train the registration model and use the trained registration model for registration of CT images and PET images.

[0042] Example

[0043] To verify the performance of this method, it was compared with VoxelMorph. [1] TransMorph [2] LapIRN [3]Three advanced deep learning image registration methods were compared, and the experimental results are shown in Table 1.

[0044] Table 1 Comparison of experimental results from different methods

[0045]

[0046]

[0047] The experimental results show that, compared with existing methods, this method achieves a DSC score of 77.85%, which is the best performance among all methods, demonstrating its superiority in CT / PET image registration tasks. The main reasons are: First, stronger feature representation capability; compared with traditional CNN methods such as VoxelMorph, this invention uses a PMFA module in each feature encoding stage of the encoder. This module, through multi-branch dilated convolution and attention mechanisms, can capture richer multi-scale features, enabling the model to better understand and represent the fine anatomical structures of CT images and the functional metabolic information of PET images, providing a high-quality feature foundation for accurate alignment. Second, superior long-distance dependency modeling capability; compared with Transformer-based methods such as TransMorph, the AFMLP module of this invention has a greater advantage in handling long-distance dependencies. To reduce computational complexity, TransMorph must run on low-resolution feature maps, leading to loss of detail information; while AFMLP, by explicitly calculating spatial correlations on higher-resolution feature maps and combining a multi-window gating mechanism, can effectively model long-distance relationships while avoiding detail loss, thereby improving registration accuracy. Third, the spatial correspondence modeling is more explicit. Unlike methods like VoxelMorph that directly stitch images together and input them into the network for "black box" prediction, the AFMLP module of this invention designs a correlation layer to explicitly calculate the local spatial correspondence between features of moving and fixed images. This explicit modeling provides the network with clearer alignment cues, especially when dealing with complex deformations, enabling more robust and accurate registration.

[0048] The source of the aforementioned existing method is:

[0049] [1] Balakrishnan G, Zhao A, Sabuncu MR, et al. Voxelmorph: a learning framework for deformable medical image registration [J]. IEEE transactions on medical imaging, 2019, 38(8): 1788-1800.

[0050] [2]Chen J, Frey EC, He Y, et al.Transmorph: Transformer for unsupervised medical image registration[J]. Medical image analysis, 2022,82:102615.

[0051] [3]Mok TCW,Chung AC S.Large deformation diffeomorphic imageregistration with laplacian pyramid networks[C] / / International Conference onMedical Image Computing and Computer-Assisted Intervention.Cham:SpringerInternational Publishing,2020:211-221.

[0052] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A CT / PET image registration method based on multi-scale features and progressive correlation correction, characterized in that, The method comprises the following steps: Step 1: obtaining PET images and CT images; Step 2: constructing a registration model, including an encoder and a decoder; The encoder comprises one convolutional block and four feature encoding stages, and the output feature map of the previous feature encoding stage is taken as the input feature map of the next feature encoding stage after down-sampling; The first feature encoding stage outputs a first scale PET feature map and a first scale CT feature map, the second feature encoding stage outputs a second scale PET feature map and a second scale CT feature map, the third feature encoding stage outputs a third scale PET feature map and a third scale CT feature map, and the fourth feature encoding stage outputs a fourth scale PET feature map and a fourth scale CT feature map; The decoder comprises four feature decoding stages, the first feature decoding stage comprises one AFMLP module, and the second to fourth feature decoding stages each comprise two AFMLP modules; The fourth scale PET feature map and the fourth scale CT feature map are input into the first feature decoding stage to obtain the output feature map of the first feature decoding stage; The output feature map of the first feature decoding stage is input into the second feature decoding stage after up-sampling, to perform spatial position registration on the third scale PET feature map, and obtain the spatial position registered third scale PET feature map; the spatial position registered third scale PET feature map and the third scale CT feature map are processed by the first AFMLP module, the output feature map of the first AFMLP module and the output feature map of the first feature decoding stage after up-sampling are input into the second AFMLP module for processing, the output feature map of the second AFMLP module and the output feature map of the first feature decoding stage after up-sampling are spliced to obtain the output feature map of the second feature decoding stage; the output feature map of the second feature decoding stage is input into the third feature decoding stage after up-sampling for processing, and the output feature map of the third feature decoding stage is obtained in the same way; The output feature map of the third feature decoding stage is input into the fourth feature decoding stage after up-sampling for processing, and the output feature map of the fourth feature decoding stage is obtained in the same way; finally, the output feature map of the fourth feature decoding stage is applied to the PET image to perform spatial position registration on the PET image, and a registration PET image aligned with the CT image is generated; In the AFMLP module, the spatial position registered PET feature map and the CT feature map are subjected to collaborative correlation calculation to obtain a correlation feature map; the correlation feature map, the spatial position registered PET feature map and the CT feature map are spliced along the channel dimension to obtain a composite feature map; the composite feature map is subjected to compression and layer normalization, and then is subjected to three scales of gated multi-layer perceptron to generate three feature maps; the three feature maps are spliced, and then are subjected to dynamic weighted fusion to obtain the importance weight of each feature map; The importance weight is multiplied by the corresponding feature map to obtain a channel-enhanced feature map. ​ The three enhanced feature maps are added together, and then are connected in residual connection with the compressed composite feature map, the feature map obtained through residual connection is input into the residual channel attention module, and then is connected in residual connection with the compressed composite feature map, to obtain the output feature map of the AFMLP module. Third step: training the registration model, and using the trained registration model for registration of the CT image and the PET image.

2. The CT / PET image registration method based on multi-scale features and progressive correlation correction according to claim 1, characterized in that, Each feature encoding stage includes a PMFA module; in the PMFA module, the input feature map is subjected to feature extraction through five branches, the output feature maps of the five branches are spliced in the channel dimension, and the spliced feature map is calibrated through the channel attention mechanism and the spatial attention mechanism to obtain a feature map calibrated in the channel and spatial dimensions; The feature map calibrated in the channel and spatial dimensions is fused through an element-by-element maximum value operation to obtain an enhanced feature map. The enhanced feature map is subjected to 1*1*1 convolution to obtain an output feature map.

3. The CT / PET image registration method based on multi-scale features and progressive correlation correction according to claim 1 or 2, characterized in that, Each feature encoding stage includes a 3DConvNeXt module; in the 3DConvNeXt module, the input feature map is subjected to residual connection with itself after deep separable convolution, batch normalization and GELU activation function, the feature map obtained through residual connection is subjected to two point-by-point convolution blocks and a convolution block in sequence to obtain an output feature map.

4. The method of claim 1, wherein, The cooperative correlation calculation process is: first, the CT feature map is zero-padded in the height, width and depth directions by 1 unit to generate multiple spatial offsets; a region with the same size as the CT feature map is extracted at each offset position, the region is multiplied element by element with the CT feature map, and then the mean value is calculated in the channel dimension to form a correlation feature map.

5. The method of claim 1, wherein, The dynamic weighted fusion includes a global average pooling layer, a multilayer perception, a Softmax activation function and a linear layer connected in series.

6. The CT / PET image registration method based on multi-scale features and progressive correlation correction according to claim 1, 4 or 5, characterized in that, In the residual channel attention module, the input feature map is subjected to residual connection with itself after convolution, Leaky ReLU activation function, convolution and channel attention mechanism in sequence to obtain an output feature map.

Images