A chest CT volume data reconstruction method based on positive and lateral projection interactive enhancement and multi-plane constraint

Abstract

The application provides a method for recovering chest CT volume data by using front and lateral X-ray projections, and relates to the fields of medical image reconstruction and intelligent imaging processing. The method first adjusts the gray scale range, unifies the voxel scale and retains the chest cavity region of the chest CT volume data, and obtains two X-ray projection images according to the front-back and left-right projection relationships; then, a front encoding channel and a lateral encoding channel are set, and a triple interactive attention feature enhancement block is configured at the encoding output to strengthen the channel-height, channel-width and spatial plane feature responses. After the enhanced double-view features are mapped to three-dimensional coordinates, aligned in the voxel space, fused and decoded in three dimensions, the chest CT volume data is obtained. During training, the voxel gray scale error, the front and lateral projection error, the local volume block discrimination error and the three-direction slice perception error are combined, so that the output volume data is consistent in terms of gray scale distribution, projection structure and multi-plane texture.

Description

Technical Field

[0001] This invention belongs to the field of medical image volume data recovery and intelligent reconstruction processing, specifically involving a reconstruction method for generating chest CT volume data using anteroposterior and lateral X-ray projection information. Background Technology

[0002] X-ray examinations have advantages such as low cost, fast imaging speed, widespread availability of equipment, and relatively low radiation dose, making them a commonly used medical imaging method for chest disease screening and preliminary clinical diagnosis. However, X-ray images are essentially two-dimensional projection images, and chest anatomical structures are prone to overlap during imaging, leading to a loss of depth information and insufficient spatial localization ability, which in turn affects the observation of complex lesions or fine structures.

[0003] CT volumetric data provides three-dimensional tomographic information, offering high spatial resolution and strong tissue visualization capabilities, making it invaluable for diagnosing chest diseases, lesion localization, and three-dimensional structural analysis. However, CT examinations typically involve high radiation doses, high equipment costs, and complex procedures, limiting the direct acquisition of high-quality CT volumetric data in primary healthcare settings, bedside examinations, for patients with limited mobility, or in scenarios requiring low-dose imaging.

[0004] By inferring CT volume data from X-ray projection images, an approximate three-dimensional structural representation can be obtained without directly performing a full CT scan. Anteroposterior and lateral images record the superimposed projections of chest structures in the anteroposterior and lateral directions, respectively. These two viewpoints are complementary, providing additional spatial clues for restoring the missing depth dimension.

[0005] Existing dual-view projection reconstruction methods still have shortcomings when applied to the recovery of chest CT volume data: on the one hand, two-dimensional coded features often lack joint modeling between channels and different spatial directions before being elevated to three-dimensional space, which can easily lead to insufficient response in areas such as lung field edges, rib contours, and mediastinal structures; on the other hand, when training solely based on voxel grayscale errors, although the reconstructed volume data can maintain a similar overall intensity distribution, continuous textures and local boundaries in the axial, coronal, and sagittal planes are easily smoothed.

[0006] Therefore, there is a need to provide a method for reconstructing chest CT volume data based on anteroposterior and lateral X-ray projection inputs, so that the model can obtain a more complete cross-dimensional structural expression in the two-dimensional encoding stage, and improve the spatial continuity and detail readability of the reconstructed volume data by utilizing multi-anatomical plane constraints in the training stage. Summary of the Invention

[0007] The purpose of this invention is to provide a chest CT volume data reconstruction method based on frontal and lateral projection interaction enhancement and multi-plane constraints, which solves the problems of insufficient cross-dimensional expression of two-dimensional coded features, insufficient alignment and fusion of dual-view volume features, and insufficient multi-plane texture details in the reconstruction results during the existing X-ray projection reconstruction of CT volume data.

[0008] To address the aforementioned issues, this invention employs a combined input method of frontal and lateral projections to recover chest CT body data. The specific process includes the following steps.

[0009] Step S1: Acquire chest CT volume data, perform grayscale standardization, voxel spacing uniformity, chest cavity target area preservation, boundary cropping and background area removal on the chest CT volume data, and generate anteroposterior X-ray images and lateral X-ray images according to the two sets of projection geometry in the anterior-posterior and left-right directions, respectively, to form paired training data between anteroposterior X-ray images, lateral X-ray images and three-dimensional CT volume data.

[0010] Step S2: Construct a three-dimensional CT volume data reconstruction model, which includes an anterior two-dimensional coding branch, a lateral two-dimensional coding branch, a triple interactive attention feature enhancement block, a two-dimensional to three-dimensional coordinate mapping module, a dual-view spatial alignment fusion module, and a three-dimensional decoding module.

[0011] Step S3: After at least one coding scale of the orthogonal two-dimensional coding branch and the lateral two-dimensional coding branch, set a triple interactive attention feature enhancement block, and perform attention modeling on the two-dimensional coding features by combining CH, CW and HW dimensions respectively to obtain orthogonal attention enhancement features and lateral attention enhancement features.

[0012] Step S4: The positive and lateral attention enhancement features are upgraded to corresponding three-dimensional volume features through the two-dimensional to three-dimensional coordinate mapping module, and the two three-dimensional volume features are spatially aligned and fused in a unified three-dimensional coordinate system. Then, the predicted three-dimensional CT volume data is output through the three-dimensional decoding module.

[0013] Step S5: Train the three-dimensional CT volume data reconstruction model based on the joint loss function. The joint loss function includes at least a multi-plane perceptual loss function, which calculates the perceptual feature differences on the axial, coronal, and sagittal slices of the predicted three-dimensional CT volume data and the real three-dimensional CT volume data, respectively.

[0014] Step S6: Input the anteroposterior X-ray image and lateral X-ray image of the object to be reconstructed into the trained three-dimensional CT volume data reconstruction model, output the corresponding three-dimensional CT volume data, and display the three-dimensional CT volume data as axial, coronal, sagittal slices or three-dimensional visualization results.

[0015] Furthermore, in step S1, the processing includes grayscale standardization, voxel spacing unification, chest cavity target region screening, non-human background removal, abnormal boundary clipping, and uniform volume data size processing of the chest CT volume data. Through the above processing, the CT volume data of different cases are kept consistent in grayscale range, voxel spacing, and spatial coverage area, reducing the impact of differences in data sources on model training.

[0016] Furthermore, the anteroposterior X-ray image is generated from 3D CT volume data based on the anteroposterior ray integration path, and the lateral X-ray image is generated from 3D CT volume data based on the left-right ray integration path. The anteroposterior X-ray image is used to characterize the superposition relationship of the frontal chest structures, and the lateral X-ray image is used to characterize the lateral depth distribution information of the chest. Together, they provide complementary spatial constraints for CT volume data reconstruction.

[0017] Furthermore, both the orthogonal and lateral two-dimensional coding branches employ a two-dimensional convolutional coding structure. Each coding branch includes multiple coding stages, and each coding stage includes a convolutional downsampling layer and a dense feature extraction layer. The shallow coding stage mainly extracts edge, texture, and local contour information, while the deep coding stage mainly extracts organ contours, global structure, and high-level semantic information.

[0018] In a preferred embodiment, the triple interactive attention feature enhancement block is embedded after multiple coding scales of the orthogonal two-dimensional coding branch and the lateral two-dimensional coding branch, respectively. For an encoder containing five downsampling coding stages, the triple interactive attention feature enhancement block can be arranged at the feature output of each downsampling stage, so that each viewpoint obtains multi-scale enhanced features from shallow edges to deep semantics.

[0019] For any encoding stage, let the two-dimensional features input to the triple interactive attention feature enhancement block be:

[0020] Where C represents the number of channels, H represents the feature height, and W represents the feature width. The triple interactive attention feature enhancement block includes a CH permutation gating branch, a CW permutation gating branch, and an HW spatial gating branch, which respectively obtain the highly correlated channel response, the width-correlated channel response, and the structural response in the two-dimensional spatial plane.

[0021] The triple interactive attention feature enhancement block employs a dual statistical pooling aggregation operator to jointly describe the mean and extreme value statistical components on the target dimension. The dual statistical pooling aggregation operator is expressed as follows:

[0022] Where d represents the dimension for feature aggregation. This represents average pooling along dimension d. This indicates max pooling along dimension d, and Concat indicates feature concatenation.

[0023] The CH permutation-gated branch is used to obtain channel responses highly correlated with the features. This branch first performs a dimensional permutation on the input feature F, ensuring the channel and height dimensions are interactively arranged; then, it performs double statistical pooling aggregation along the target dimension, followed by convolutional mapping and sigmoid gating to obtain attention weights; these weights are then applied to the permuted features, and finally, the original dimensional order is restored through inverse permutation. Its expression is:

[0024] The CW permutation-gated branch is used to obtain the channel response related to the feature width. This branch aligns the channel dimensions with the width dimension through another set of dimensional permutations, and then sequentially performs double statistical pooling aggregation, convolutional mapping, and sigmoid gating to generate the corresponding attention weights. After element-wise weighting, the original dimensional order is restored through inverse permutation, and its expression is:

[0025] The HW spatial gating branch is used to obtain the two-dimensional structural response in the original spatial plane. This branch preserves the dimensional arrangement of the input features, performs double statistical pooling aggregation along the channel dimension, and then generates spatial attention weights through convolutional mapping and sigmoid gating. Its expression is:

[0026] Among them, the dimension permutation operation is used to adjust the arrangement relationship between the channel dimension and the spatial dimension, the inverse permutation operation is used to restore the weighted features to the original dimensional order; the sigmoid activation function is used to generate gate weights, the convolutional mapping is used to model the local response of the concatenated statistical features, and the element-wise multiplication is used to apply the gate weights to the corresponding features.

[0027] Finally, the outputs of the CH permutation-gated branch, the CW permutation-gated branch, and the HW spatial gated branch are fused to obtain the output features of the triple interactive attention feature enhancement block:

[0028] Through the aforementioned three-branch attention enhancement method, the model can simultaneously enhance the channel response, lung field boundaries, bony contours, mediastinal regions, and local texture details in the projection image during the two-dimensional encoding stage, thereby improving the structural representation ability when transferring two-dimensional projection features to three-dimensional CT volume data recovery tasks.

[0029] Furthermore, in the two-dimensional convolutional coding stage, the dense feature extraction layer can be represented as:

[0030] Where s represents the s-th encoding stage, and n represents the n-th dense layer. This represents a feature transformation consisting of convolution, normalization, and nonlinear activation. This represents the number of dense layers in the s-th encoding stage. Dense connections can preserve low-level texture information and promote feature reuse.

[0031] After obtaining the frontal and lateral attention-enhanced features, the 2D-to-3D coordinate mapping module unfolds the 2D encoded features along the missing depth dimension according to the corresponding viewpoint direction, and adjusts the number of channels and voxel positions through 3D convolution to form 3D volumetric features corresponding to the frontal and lateral images. Subsequently, the dual-view spatial alignment fusion module transforms the two 3D volumetric features to a unified voxel coordinate system and completes the fusion using channel stitching, attention weighting, or skip connection methods.

[0032] In a preferred embodiment, the frontal 2D coding branch and the lateral 2D coding branch each output corresponding 3D volumetric features. Through dimension transposition, coordinate alignment, and weighted fusion, the anterior-posterior projection information and the lateral-posterior projection information are complementaryly expressed in a unified 3D coordinate system. The fused 3D volumetric features are input into the 3D decoding module, undergo multi-level 3D upsampling and 3D convolution processing, and finally output predicted 3D CT volumetric data.

[0033] To further improve the structural continuity and texture readability of reconstructed volume data across different anatomical directions, this invention introduces a multi-plane perceptual loss function. This loss does not solely rely on voxel space error; instead, it slices the predicted 3D CT volume data and the actual 3D CT volume data from the axial, coronal, and sagittal planes respectively, and calculates the differences between corresponding slices in the perceptual feature space.

[0034] For any anatomical plane p, the sensory loss is defined as:

[0035] Multi-plane sensing loss is defined as:

[0036] in, This represents the predicted 3D CT volume data. Represents actual 3D CT volume data, where 'a' represents the axial plane, 'c' the coronal plane, and 's' the sagittal plane. This represents the operator for extracting the i-th slice from the p-th anatomical plane. This represents the number of slices at the p-th anatomical plane. This represents the output of the k-th layer of the perceptual feature extraction network, where K represents the number of perceptual feature layers. This represents the weighting coefficient of the p-th anatomical plane.

[0037] In one specific embodiment, the perceptual feature extraction network is a deep feature extraction network (e.g., a pre-trained VGG network or LPIPS network) with a multi-level cascaded convolutional architecture, pre-trained on a large image dataset. The outputs of convolutional layers of different depths in this network are used as the perceptual feature space to extract multi-scale structural and texture features of the image. For grayscale CT slices, single-channel slices can be copied as three-channel inputs to adapt to the perceptual feature extraction network. To reduce memory consumption, a portion of the slice can be uniformly sampled in each anatomical plane to calculate the multi-planar perceptual loss.

[0038] During the training phase, the joint loss function also includes voxel reconstruction loss, projection consistency loss, and generative adversarial loss. Voxel reconstruction loss is used to constrain the difference in voxel grayscale between the predicted 3D CT volumetric data and the real 3D CT volumetric data, and is expressed as:

[0039] The projection consistency loss can be expressed as:

[0040] in, This represents the projection operator in the q-th direction. It represents the set of projection directions, which may include the orthographic direction and the lateral direction.

[0041] The generative adversarial loss can be calculated by a 3D image patch discriminator. For the generator, the generative adversarial loss can be expressed as:

[0042] Finally, the joint loss function of the generator can be expressed as:

[0043] in, , respectively, represent the weight coefficients of the generative adversarial loss, voxel reconstruction loss, projection consistency loss, and multi-planar perception loss. Through the above joint loss function, the model can jointly optimize the prediction of 3D CT volume data from four aspects: voxel grayscale consistency, frontal and lateral projection consistency, local block realism, and multi-anatomical planar perception structure.

[0044] Compared with the prior art, the present invention has the following beneficial effects: First, the present invention uses frontal and lateral X-ray images as inputs, which, compared with single-view X-ray reconstruction methods, can provide more sufficient spatial complementary information and reduce structural ambiguity in the two-dimensional to three-dimensional reconstruction process.

[0045] Second, the present invention sets up a triple interactive attention feature enhancement block in the positive two-dimensional coding branch and the lateral two-dimensional coding branch. By combining the three types of dimensions CH, CW and HW, the response of the highly correlated channel, the response of the width correlated channel and the spatial plane response are enhanced respectively, so that the lung field contour, rib boundary, mediastinal region and local texture in the projection image are more fully expressed.

[0046] Third, this invention utilizes axial, coronal, and sagittal slices to construct multi-planar perceptual constraints, enabling the training process to not only focus on voxel grayscale errors but also constrain boundary continuity and texture consistency in different anatomical directions, thereby reducing the oversmoothing phenomenon of reconstructed volume data.

[0047] Fourth, this invention only requires inputting anteroposterior and lateral X-ray images during the reasoning stage to obtain corresponding chest CT body data, which can be used as an auxiliary reconstruction scheme for low-dose image supplementation, primary screening, and medical image scientific research analysis. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0049] Figure 2 This is a structural diagram of the dual-view three-dimensional CT volume data reconstruction model of the present invention.

[0050] Figure 3 This is a structural diagram of the triple interactive attention feature enhancement block of the present invention.

[0051] Figure 4 This is a schematic diagram of the reconstructed CT slices and three-dimensional visualization results of the present invention. Detailed Implementation

[0052] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0053] like Figure 1As shown, this embodiment provides a method for reconstructing chest CT volume data based on interactive enhancement of anteroposterior and lateral projection and multi-plane constraints. The method first performs grayscale normalization, voxel spacing unification, preservation of the target region of the thoracic cavity, and background removal on the chest CT volume data. Then, it generates anteroposterior and lateral X-ray images based on two sets of projection geometry: anteroposterior and lateral. Subsequently, a dual-branch 3D CT volume data reconstruction model is constructed, introducing triple interactive attention feature enhancement blocks into the anteroposterior and lateral 2D coding branches. The dual-view enhancement features are then mapped from 2D to 3D coordinates, spatially aligned and fused, and 3D decoded to generate predicted 3D CT volume data. Finally, a joint loss function including multi-plane perceptual loss is used to train the model, and the 3D CT volume data and its visualization results are output during the inference phase.

[0054] In one embodiment, the data processing procedure includes: first, performing grayscale standardization on the original chest CT volume data; then, standardizing the voxel spacing of the CT volume data to ensure consistent spatial sampling scale across different cases; next, screening the target area of ​​the thoracic cavity and removing the examination bed, atmospheric background, and other non-human areas; and finally, performing boundary clipping and size regularization on the CT volume data to obtain standardized three-dimensional CT volume data.

[0055] After obtaining standardized 3D CT volume data, an anteroposterior X-ray image is generated based on the anterior-posterior ray integration path, and a lateral X-ray image is generated based on the left-right ray integration path. This constructs paired training data between the anteroposterior X-ray image, the lateral X-ray image, and the actual 3D CT volume data.

[0056] like Figure 2 As shown, the 3D CT volume data reconstruction model in this embodiment includes an anteroposterior 2D coding branch and a lateral 2D coding branch. The two coding branches have the same or similar 2D convolutional coding structures, used to extract 2D structural features from the anteroposterior and lateral X-ray images, respectively. Each coding branch includes multiple downsampling coding stages, in which convolutional downsampling, dense feature extraction, and attention enhancement are performed sequentially.

[0057] like Figure 3 As shown, this embodiment introduces a triple interactive attention feature enhancement block in the generator encoding stage. The triple interactive attention feature enhancement block includes a CH permutation gated branch, a CW permutation gated branch, and a HW spatial gated branch; wherein, the CH permutation gated branch is used to obtain highly correlated channel responses, the CW permutation gated branch is used to obtain width-correlated channel responses, and the HW spatial gated branch is used to obtain two-dimensional structural responses in the original spatial plane. All three branches generate attention weights through double statistical pooling aggregation, convolutional mapping, and sigmoid gating, and then multiply them element-wise with their corresponding features, finally fusing the outputs of the three branches.

[0058] In a preferred embodiment, the triple interactive attention feature enhancement block is incorporated into the five downsampling encoding stages of both the orthogonal and lateral 2D encoding branches. For the final encoding stage, the triple interactive attention feature enhancement block is placed before the adaptive average pooling operation to prevent premature compression of the feature map to 1×1, which could lead to degradation of the spatial attention branch. Thus, both viewpoint encoding branches can obtain cross-dimensional attention enhancement features at multiple scales.

[0059] After obtaining the enhanced encoded features from both perspectives, the 2D features are upscaled to 3D volumetric features using a 2D-to-3D coordinate mapping module. Subsequently, the 3D volumetric features from the frontal and lateral branches are unified into the same voxel coordinate system through dimension transposition, coordinate alignment, and feature fusion. The fused 3D volumetric features are then input into a 3D decoding module, which gradually restores the spatial resolution through multi-level 3D transposed convolution or 3D upsampling, ultimately outputting the predicted 3D CT volumetric data.

[0060] During the training phase, predicted 3D CT volume data and real 3D CT volume data are jointly input into a joint loss function for optimization. Voxel reconstruction loss is used to constrain the grayscale consistency of volume data, projection consistency loss is used to constrain the projection structure consistency of predicted volume data in the anteroposterior and lateral directions, generative adversarial loss is used to improve the realism of local volumes, and multi-plane perception loss is used to constrain the structural continuity and texture details on axial, coronal, and sagittal slices.

[0061] In one specific embodiment, multi-planar perceptual loss extracts axial, coronal, and sagittal slices from both predicted and actual 3D CT volume data. For each anatomical plane, the predicted and actual slices are input into a pre-trained perceptual feature extraction network, the L1 distance between the outputs of corresponding feature layers is calculated, and the perceptual errors of the three anatomical planes are weighted and summed. In this way, the model not only approximates actual CT volume data in voxel grayscale space but also in multi-planar structure and texture perceptual space.

[0062] like Figure 4 As shown, during the inference phase, the anteroposterior and lateral X-ray images of the object to be reconstructed are input into the trained 3D CT volume data reconstruction model, and the model outputs the corresponding 3D CT volume data. This 3D CT volume data can be directly stored or further displayed as axial, coronal, sagittal slices or 3D visualization results.

Claims

1. A method for reconstructing chest CT volume data based on interactive enhancement of anteroposterior and lateral projection and multiplanar constraints, characterized in that, Includes the following steps: S1: Acquire chest CT volume data, perform grayscale standardization, voxel spacing unification, chest cavity target area preservation, boundary clipping and background area removal on the chest CT volume data, and generate frontal X-ray images and lateral X-ray images according to the two sets of projection geometry in the anterior-posterior and left-right directions respectively, forming paired training data between frontal X-ray images, lateral X-ray images and three-dimensional CT volume data. S2: Construct a three-dimensional CT volume data reconstruction model, which includes an anteroposterior two-dimensional coding branch, a lateral two-dimensional coding branch, a triple interactive attention feature enhancement block, a two-dimensional to three-dimensional coordinate mapping module, a dual-view spatial alignment fusion module, and a three-dimensional decoding module. S3: After at least one coding scale of the orthogonal two-dimensional coding branch and the lateral two-dimensional coding branch, a triple interactive attention feature enhancement block is set to perform attention modeling on the two-dimensional coding features by combining CH, CW and HW dimensions respectively, to obtain orthogonal attention enhancement features and lateral attention enhancement features. S4: The two-dimensional to three-dimensional coordinate mapping module enhances the positive and lateral attention enhancement features into corresponding three-dimensional volume features, and performs spatial alignment and fusion of the two three-dimensional volume features under a unified three-dimensional coordinate system. Then, the three-dimensional decoding module outputs the predicted three-dimensional CT volume data. S5: Train the three-dimensional CT volume data reconstruction model based on the joint loss function, wherein the joint loss function includes at least a multi-plane perception loss function, and the multi-plane perception loss function calculates the perceptual feature differences on the axial, coronal and sagittal slices of the predicted three-dimensional CT volume data and the real three-dimensional CT volume data respectively; S6: Input the anteroposterior X-ray image and lateral X-ray image of the object to be reconstructed into the trained three-dimensional CT volume data reconstruction model, output the corresponding three-dimensional CT volume data, and display the three-dimensional CT volume data as axial, coronal, sagittal slices or three-dimensional visualization results.

2. The method according to claim 1, characterized in that, The processing of chest CT volume data in S1 includes: grayscale standardization, voxel spacing unification, chest cavity target region screening, non-human background removal, abnormal boundary clipping, and uniform volume data size processing of the original CT volume data, so that CT volume data from different sources have a consistent grayscale range, voxel spacing, and spatial coverage area.

3. The method according to claim 1, characterized in that, The frontal X-ray image is generated from chest CT volume data based on the anterior-posterior ray integration path, and the lateral X-ray image is generated from chest CT volume data based on the left-right ray integration path; wherein, the frontal X-ray image represents the superposition relationship of the frontal chest structures, and the lateral X-ray image represents the lateral depth distribution information of the chest.

4. The method according to claim 1, characterized in that, The 3D CT volume data reconstruction model includes multiple 2D convolutional coding stages in both the frontal and lateral 2D coding branches. Each 2D convolutional coding stage includes a convolutional downsampling layer and a dense feature extraction layer. The triple interactive attention feature enhancement block is set after one or more 2D convolutional coding stages and is used to perform attention-weighted enhancement on the projected structural features at different scales.

5. The method according to claim 4, characterized in that, The input two-dimensional features of the triple interactive attention feature enhancement block are denoted as follows: Where C represents the number of channels, H represents the feature height, and W represents the feature width; the triple interactive attention feature enhancement block includes a CH permutation gated branch, a CW permutation gated branch, and an HW spatial gated branch, which are used to obtain the height-related channel response, the width-related channel response, and the structural response in the two-dimensional spatial plane, respectively.

6. The method according to claim 5, characterized in that, The triple interactive attention feature enhancement block employs a dual statistical pooling aggregation operator, which includes mean statistical components and extreme statistical components obtained along the target dimension. The mean statistical components and extreme statistical components are concatenated and then input into the convolutional gating layer to generate corresponding attention weights.

7. The method according to claim 5, characterized in that, The CH permutation-gated branch, CW permutation-gated branch, and HW spatial gated branch respectively calculate their corresponding attention enhancement features in the following ways: Among them, the dimension permutation operation is used to adjust the arrangement relationship between the channel dimension and the spatial dimension, and the inverse permutation operation is used to restore the attention-enhanced features to the original dimensional order; the dual statistical pooling aggregation operator executed along the width dimension W, the height dimension H, or the channel dimension C is used to generate the statistical description of the corresponding branch, the sigmoid activation function is used to obtain the gate weights, the convolutional mapping is used to model the local response of the concatenated statistical features, and the element-wise multiplication is used to apply the gate weights to the corresponding features.

8. The method according to claim 7, characterized in that, The triple interactive attention feature enhancement block fuses the outputs of the CH permutation gating branch, the CW permutation gating branch, and the HW spatial gating branch to obtain the enhanced two-dimensional coding features: The enhanced two-dimensional coding features are used for subsequent two-dimensional to three-dimensional coordinate mapping, dual-view spatial alignment and fusion, and three-dimensional CT volume data reconstruction.

9. The method according to claim 1, characterized in that, The multi-plane perceptual loss function extracts axial, coronal, and sagittal slices from both predicted and actual 3D CT volume data, and inputs the corresponding slices into the perceptual feature extraction network to calculate the structural texture differences under different anatomical planes; for any anatomical plane p, the perceptual loss is: The multi-plane sensing loss function is: in, This represents the predicted 3D CT volume data. This represents actual 3D CT volume data, where a, c, and s represent the axial, coronal, and sagittal planes, respectively. This represents the operator for extracting the i-th slice from the p-th anatomical plane. This represents the number of slices at the p-th anatomical plane. This represents the output of the k-th layer of the perceptual feature extraction network, where K represents the number of perceptual feature layers. This represents the weighting coefficient of the p-th anatomical plane.

10. The method according to claim 1, characterized in that, The joint loss function is composed of a weighted average of generative adversarial loss, voxel reconstruction loss, projection consistency loss, and multi-plane perceptual loss, and is expressed as: Among them, generative adversarial loss is used to constrain the local realism of the predicted volume data, voxel reconstruction loss is used to constrain the grayscale difference between the predicted 3D CT volume data and the real 3D CT volume data, projection consistency loss is used to constrain the consistency of the projection structure of the predicted volume data in the frontal and lateral directions, multi-plane perception loss is used to constrain the differences in perception features in the axial, coronal and sagittal planes, and the weight coefficients of each loss term are used to adjust the proportion of the corresponding constraint in the joint training.

Images