A dual-view CT reconstruction method based on geometric guided diffusion and frequency domain perception
By employing a dual-view CT reconstruction method based on geometrically guided diffusion and frequency domain sensing, and utilizing discrete wavelet transform and cross-attention mechanism, the problems of high-frequency detail loss and Z-axis discontinuity artifacts in sparse-view CT reconstruction are solved, achieving high-resolution 3D CT volume reconstruction, which is suitable for low-dose, low-cost medical image reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-03-27
- Publication Date
- 2026-06-09
AI Technical Summary
Existing sparse view CT reconstruction models suffer from high-frequency detail loss and Z-axis discontinuity artifacts caused by slice-by-slice generation. Traditional methods cause spatial aliasing and loss of high-frequency signals of microscopic anatomical structures during feature extraction.
A dual-view CT reconstruction method based on geometrically guided diffusion and frequency domain awareness is adopted. By acquiring anteroposterior and lateral X-ray projection data, features are extracted using a dual-domain residual network of discrete wavelet transform, and feature alignment and stitching are performed by combining a cross-attention mechanism to generate high-resolution three-dimensional CT volume data.
It restores high-fidelity microscopic anatomical structures, ensures strict consistency in three-dimensional space, and provides a low-dose, low-cost, high-quality medical image reconstruction solution.
Smart Images
Figure CN122176323A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer vision technology, and in particular to a dual-view CT reconstruction method based on geometrically guided diffusion and frequency domain perception. Background Technology
[0002] Computed tomography (CT) is an important tool for clinical diagnosis, but its high radiation dose and expensive, non-removable equipment limit its application. In contrast, orthogonal biplane X-ray imaging has the advantages of low radiation, low cost, and easy accessibility. Therefore, reconstructing high-fidelity 3D CT volumes from sparse 2D biplane projections has become a key inverse problem in medical image computation.
[0003] Early deep learning methods often employed convolutional neural networks or generative adversarial networks. However, due to the voxel-level regression objective, these models often fell into the "regression mean" effect, resulting in severe smoothing and blurring of high-frequency textures such as bone trabeculae. In recent years, denoising diffusion probability models have shown great potential in medical image synthesis, with frameworks such as DX2CT and DVG-Diffusion emerging for CT reconstruction.
[0004] Although existing high-resolution 3D diffusion frameworks have broken through the "fuzzy bottleneck," they still have the following drawbacks: First, to avoid the cubic memory expansion caused by full 3D U-Nets, existing methods generally adopt a slice-wise generation paradigm. This approach completely severs the contextual dependencies between the coronal and sagittal planes, leading to morphological abrupt changes along the Z-axis and severe "inter-slice jittering" artifacts, compromising the integrity of clinical diagnosis. Second, during front-end feature extraction, traditional stride-based downsampling causes spatial aliasing and loses high-frequency signals of microscopic anatomical structures. Summary of the Invention
[0005] This application provides a dual-view CT reconstruction method based on geometry-guided diffusion and frequency domain sensing to solve the problems of high-frequency detail loss and Z-axis discontinuity artifacts caused by slice-by-slice generation in existing sparse-view CT reconstruction models. The method includes:
[0006] Acquire the frontal and lateral X-ray projection data of the target object;
[0007] The frontal X-ray projection data is input into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain the first projection feature map. The lateral X-ray projection data is input into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain the second projection feature map.
[0008] The two-dimensional features of the first projection feature map and the two-dimensional features of the second projection feature map are projected into three-dimensional space to obtain aligned three-dimensional space features.
[0009] The three-dimensional spatial features of three adjacent feature slices are extracted to form a feature set. The three-dimensional spatial features in the feature set are spliced and fused along the channel dimension to obtain the context condition feature map.
[0010] The contextual condition feature map is used as a priori guided input diffusion model for analysis and processing to generate high-resolution 3D CT volume data.
[0011] Optionally, the step of inputting the orthogonal X-ray projection data into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain a first projection feature map includes:
[0012] The first intermediate feature map is obtained by downsampling the orthogonal X-ray projection data using two-dimensional discrete wavelet transform.
[0013] The first intermediate feature map is processed sequentially through a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation function to obtain the second intermediate feature map;
[0014] The second intermediate feature map is processed sequentially through a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function to obtain the third intermediate feature map;
[0015] The third intermediate feature map is processed sequentially through a 1×1 convolutional layer and a batch normalization layer to obtain the fourth intermediate feature map;
[0016] The input orthogonal X-ray projection data is processed through a 1×1 convolutional layer and then added to the fourth intermediate feature map to obtain the first projection feature map.
[0017] Optionally, the step of projecting the two-dimensional features of the first projection feature map and the two-dimensional features of the second projection feature map onto three-dimensional space to obtain aligned three-dimensional spatial features includes:
[0018] A physical mesh is constructed based on the preset target 3D CT resolution. Any feature slice that needs to be reconstructed in the physical mesh is uniformly sampled to obtain the absolute coordinates of multiple discrete 3D voxels on the slice.
[0019] The absolute coordinates of three-dimensional voxels are transformed into a query matrix in the cross-attention mechanism through sine and cosine position encoding and linear mapping.
[0020] Using the physical perspective projection matrix of an actual X-ray device, the absolute coordinates of three-dimensional voxels are back-projected along the real physical ray trajectory to the first projection feature map and the second projection feature map to obtain the first perspective projection coordinates and the second perspective projection coordinates.
[0021] A first local perceptual feature block of a preset size is extracted from the first projection feature map with the first perspective projection coordinates as the center, and a second local perceptual feature block of a preset size is extracted from the second projection feature map with the second perspective projection coordinates as the center.
[0022] After feature splicing and flattening, the first and second local sensing feature blocks are linearly mapped into a key matrix and a value matrix.
[0023] The attention score is obtained by performing a dot product on the query matrix and the value matrix, and the attention score is normalized by the Softmax function to generate attention weights;
[0024] Based on the attention weights, the value matrix is weighted and summed to obtain the high-dimensional physical features of the three-dimensional voxel points after cross-dimensional alignment.
[0025] The high-dimensional physical features of all three-dimensional voxel points on the feature slice are combined to obtain the complete feature of the feature slice.
[0026] Optionally, the step of stitching and fusing the three-dimensional spatial features in the feature set along the channel dimension to obtain the contextual condition feature map includes:
[0027] The feature maps of the previous layer, the current layer, and the next layer are concatenated and stitched together along the channel dimension to obtain a high-dimensional hybrid tensor.
[0028] The high-dimensional hybrid tensor is processed by a 1×1 convolutional layer, normalization, and nonlinear activation function to compress the channel dimension of the high-dimensional hybrid tensor back to the target dimension, thus obtaining the context conditional feature map.
[0029] Optionally, the step of using the contextual condition feature map as a priori guided input diffusion model for analysis and processing to generate high-resolution three-dimensional CT volume data includes:
[0030] An initial pure noise slice is randomly sampled from a standard normal distribution according to a preset number of sampling steps;
[0031] The denoising network is trained by using the initial pure noise slice and the context conditional feature map as prior guided inputs. Based on the Gaussian noise predicted at the current time step, the noise slice at the current time step is denoised by inverse reparameterization to obtain the denoised feature slice.
[0032] The denoised feature slices are written into the initialized three-dimensional empty tensor to obtain the reconstructed high-resolution three-dimensional CT volume.
[0033] Optionally, after training the denoising network using the contextual conditional feature map as prior guided input, the method further includes:
[0034] The context conditional feature map and the noise slice at the current time step are concatenated along the channel dimension to obtain the fused initial input tensor;
[0035] The aligned hidden layer tensor is obtained by linearly projecting the initial input tensor.
[0036] The aligned hidden layer tensor is input into the encoder of the denoising network to extract denoising features, and the Gaussian noise predicted at the current time step is output.
[0037] Optionally, the Gaussian noise predicted based on the current time step is denoised by inverse reparameterization of the noise slice at the current time step to obtain a denoised feature slice that satisfies the following formula:
[0038]
[0039] in, This is the denoised slice from the previous time step. The noise scheduling parameters are preset for the current time step. The noise scheduling parameters preset in the previous time step, This is a noise slice for the current time step. The Gaussian noise predicted at the current time step. For contextual conditional feature maps, Indicates the current time step.
[0040] Optionally, the constraints for the training process of the denoising network satisfy the following formula:
[0041]
[0042] in, This represents the mean square error loss of the denoising autoencoder. This represents a noisy sample in the latent space. This represents true Gaussian noise. Represents the standard normal distribution. This represents the Gaussian noise predicted at the current time step. This represents the noise sample at the current time step. Represents the contextual conditional feature map. Indicates the current time step. This represents the square of the L2 norm.
[0043] This application provides a dual-view CT reconstruction method based on geometrically guided diffusion and frequency domain sensing. The method acquires both anteroposterior and lateral X-ray projection data of the target object. The anteroposterior X-ray projection data is input into a dual-domain residual network based on discrete wavelet transform for feature extraction, yielding a first projection feature map. Similarly, the lateral X-ray projection data is input into the same network for feature extraction, yielding a second projection feature map. The two-dimensional features of the first and second projection feature maps are projected into three-dimensional space to obtain aligned three-dimensional spatial features. Three-dimensional spatial features from three adjacent feature slices are extracted to form a feature set. These features are then stitched and fused along the channel dimension to obtain a context-conditional feature map. This context-conditional feature map is used as a priori guided input diffusion model for analysis and processing, generating high-resolution three-dimensional CT volume data. This method not only recovers high-fidelity microscopic anatomical structures from a minimal amount of two-dimensional X-ray projections but also ensures strict consistency in three-dimensional space, providing a reliable technical solution for low-dose, low-cost, high-quality medical image reconstruction in clinical settings. Attached Figure Description
[0044] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0045] Figure 1 This is a flowchart illustrating a dual-view CT reconstruction method based on geometry-guided diffusion and frequency domain sensing provided in an embodiment of this application. Figure 1 ;
[0046] Figure 2 This is a flowchart illustrating a dual-view CT reconstruction method based on geometry-guided diffusion and frequency domain sensing provided in an embodiment of this application. Figure 2 ;
[0047] Figure 3 This is a schematic diagram of the structure of the dual-domain residual network provided in the embodiments of this application.
[0048] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0050] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.
[0051] In this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.
[0052] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0053] Figure 1 This is a flowchart illustrating a dual-view CT reconstruction method based on geometry-guided diffusion and frequency domain sensing provided in an embodiment of this application. Figure 1 . Figure 2 This is a flowchart illustrating a dual-view CT reconstruction method based on geometry-guided diffusion and frequency domain sensing provided in an embodiment of this application. Figure 2 .like Figures 1-2 As shown, the dual-view CT reconstruction method based on geometry-guided diffusion and frequency domain sensing provided in this embodiment includes:
[0054] S1: Obtain the frontal and lateral X-ray projection data of the target object.
[0055] Understandably, frontal projection involves X-rays being projected from the front to the back of the target object, or from the back to the front, to show the front and back views of the target object. Side projection involves X-rays being projected from the side of the target object (e.g., from left to right or from right to left), to show the side view of the target object.
[0056] In an optional embodiment, after acquiring the frontal and lateral X-ray projection data of the target object, the method further includes: preprocessing the data samples, the preprocessing including data normalization and image enhancement operations, wherein the image enhancement operations include at least one of rotation, flipping, and scaling.
[0057] S2: Input the orthogonal X-ray projection data into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain the first projection feature map. Input the lateral X-ray projection data into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain the second projection feature map.
[0058] Specifically, the orthogonal X-ray projection data is input into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain a first projection feature map. This includes: downsampling the orthogonal X-ray projection data using two-dimensional discrete wavelet transform to obtain a first intermediate feature map; processing the first intermediate feature map sequentially through a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation function to obtain a second intermediate feature map; processing the second intermediate feature map sequentially through a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function to obtain a third intermediate feature map; processing the third intermediate feature map sequentially through a 1×1 convolutional layer and a batch normalization layer to obtain a fourth intermediate feature map; and adding the input orthogonal X-ray projection data after processing it through a 1×1 convolutional layer to the fourth intermediate feature map to obtain the first projection feature map.
[0059] In this embodiment, the process of inputting the lateral X-ray projection data into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain the second projection feature map is similar to the above-described process of "inputting the anterior X-ray projection data into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain the first projection feature map", and will not be repeated here.
[0060] Understandably, in traditional convolutional networks, pooling or stride convolution downsampling inevitably leads to the loss of high-frequency signals such as image edges and micro-textures. In order for the network to capture sharp edge features, this application designs a dual-domain residual network based on discrete wavelet transform for feature downsampling. Figure 3 This is a schematic diagram of the structure of the dual-domain residual network provided in the embodiments of this application.
[0061] For example, let the input features be... ,in, For feature map height, The width of the feature map. The spatial resolution is halved after two-dimensional discrete wavelet transform processing, but the number of channels is expanded to four times, i.e., the output features... The discrete wavelet transform can losslessly decompose the input two-dimensional feature map into four sub-bands: a low-frequency sub-band (LL) containing global approximation information and three high-frequency sub-bands (LH, HL, HH) containing edge details. Output features Channel integration is performed sequentially through convolutional layers with kernel sizes of 1×1, 3×3, and 1×1 and a stride of 1. After the convolutional layers, batch normalization layers and ReLU activation functions are applied to perform nonlinear feature mapping.
[0062] Furthermore, the dual-domain residual network uses a residual connection mechanism to alleviate the "spectral bias" of the neural network, ensuring that high-frequency cues such as bone contours and fine blood vessels in the dual-view X-ray are completely preserved, laying a solid frequency domain foundation for subsequent high-fidelity generation.
[0063] S3: Project the two-dimensional features of the first projection feature map and the two-dimensional features of the second projection feature map into three-dimensional space to obtain aligned three-dimensional space features.
[0064] Specifically, it includes the following steps:
[0065] S31: Construct a physical mesh according to the preset target 3D CT resolution, and uniformly sample any feature slice that needs to be reconstructed in the physical mesh to obtain the absolute coordinates of multiple discrete 3D voxels on the slice.
[0066] For example, the 3D CT voxel is divided into 128×128 slices, for a total of 128 slices. The depth of the z-th slice that needs to be reconstructed is locked as z, and the discrete 3D voxel absolute coordinates on the slice are obtained by uniform sampling. .
[0067] S32: Transform the absolute coordinates of three-dimensional voxels into a query matrix in the cross-attention mechanism through sine and cosine position encoding and linear mapping.
[0068] Understandably, coordinates Through sine and cosine position encoding (PE) and linear mapping, it is transformed into a high-dimensional query vector in the cross-attention mechanism, which contains the precise three-dimensional spatial geometric prior of the current voxel.
[0069] S33: Using the physical perspective projection matrix of the actual X-ray equipment, the absolute coordinates of the three-dimensional voxels are back-projected along the real physical ray trajectory to the first projection feature map and the second projection feature map to obtain the first perspective projection coordinates and the second perspective projection coordinates.
[0070] Understandably, by using the physical perspective projection matrix of an actual X-ray device, the coordinates... The perspective projection coordinates are obtained by back-projecting the actual physical ray trajectory onto the two-dimensional feature maps of the frontal and lateral positions extracted by the dual-domain residual network. and .
[0071] Its mathematical projection equation in the homogeneous coordinate system is:
[0072]
[0073]
[0074] in, This is the depth scaling factor along the positive orientation. This is the depth scaling factor along the lateral direction. This is the perspective projection matrix in the positive orientation. This is the perspective projection matrix in the lateral direction. voxel point Continuous subpixel perspective coordinates on the orthogonal feature map voxel point Continuous subpixel perspective coordinates on the lateral feature map.
[0075] S34: Using the first perspective projection coordinates as the center, extract a first local perceptual feature block of a preset size from the first projection feature map, and using the second perspective projection coordinates as the center, extract a second local perceptual feature block of a preset size from the second projection feature map.
[0076] Understandably, using the projection coordinates calculated in step S33 above as the center, local perceptual feature blocks of a fixed size (such as 3×3) are extracted from the two-dimensional feature maps of the frontal and lateral views respectively.
[0077] S35: After feature splicing and flattening, the first local sensing feature block and the second local sensing feature block are linearly mapped into a key matrix and a value matrix.
[0078] Understandably, after the extracted dual-view feature blocks are spliced and flattened, they are mapped to a key matrix and a value matrix through a linear layer to construct a two-dimensional data dictionary containing local anatomical details.
[0079] S36: Calculate the attention score by performing a dot product on the query matrix and the value matrix, and normalize the attention score using the Softmax function to generate attention weights.
[0080] Understandably, the process involves performing a dot product between a query matrix representing 3D location and a key matrix representing 2D features, and then generating attention weights using the Softmax function. The goal of this step is to autonomously evaluate and find the pixel features that best match the true semantics of the target 3D slice point within a local 2D neighborhood.
[0081] S37: Based on the attention weights, the value matrix is weighted and summed to obtain the high-dimensional physical features of the three-dimensional voxel points after cross-dimensional alignment.
[0082] Understandably, based on the calculated attention weights, the value matrix is weighted and summed, and the final output feature is the 3D voxel point. High-dimensional physical features after precise geometric error correction and cross-dimensional alignment.
[0083] S38: Combine the high-dimensional physical features of all three-dimensional voxel points on the feature slice to obtain the complete feature of the feature slice.
[0084] S4: Extract the three-dimensional spatial features of adjacent three-layer feature slices to form a feature set, and stitch and fuse the three-dimensional spatial features in the feature set along the channel dimension to obtain the context condition feature map.
[0085] Specifically, the three-dimensional spatial features in the feature set are spliced and fused along the channel dimension to obtain the context condition feature map, including: cascading and splicing the feature map of the previous layer, the feature map of the current layer, and the feature map of the next layer along the channel dimension to obtain a high-dimensional hybrid tensor; the high-dimensional hybrid tensor is processed by a 1×1 convolutional layer, normalization, and nonlinear activation function to compress the channel dimension of the high-dimensional hybrid tensor back to the target dimension to obtain the context condition feature map.
[0086] For example, the aligned upper-layer feature map Current layer feature map and the next layer feature map (The dimension of a single-layer feature map is 1) ,in (where the number of channels is 1), and then a strict concatenation and splicing is performed along the channel dimension to obtain a dimension of 1. The high-dimensional hybrid tensor is used, and a 1×1 convolutional layer is employed, combined with operations such as normalization and nonlinear activation functions, to losslessly compress the dilated 3D channels back to the target dimension, restoring it to d-dimensionality, and finally outputting a 2.5D context-conditional feature map. .
[0087] This method extracts the geometric alignment features of the current target slice and its physically adjacent slices, and then splices and fuses them along the channel dimension into contextual condition features, which serve as a priori guide for the denoising U-Net. While avoiding the huge memory overhead of the full 3D diffusion model, it injects strong Z-axis vertical topological constraints into the generation process, completely eliminating inter-layer jump artifacts.
[0088] S5: The contextual feature map is used as a priori guided input diffusion model for analysis and processing to generate high-resolution 3D CT volume data.
[0089] In this embodiment, the diffusion model is based on the standard Denoising Diffusion Probabilistic Models (DDPM) framework and the U-Net architecture of Conditional Generation.
[0090] Specifically, the contextual feature map is used as a priori guided input diffusion model for analysis and processing to generate high-resolution 3D CT volume data, including the following steps:
[0091] S51: Randomly sample an initial pure noise slice from a standard normal distribution according to a preset number of sampling steps.
[0092] For example, the number of sampling steps Sample initial pure noise slices from a standard normal distribution .
[0093] S52: The denoising network is trained by using the initial pure noise slice and the contextual conditional feature map as prior guided inputs. Based on the Gaussian noise predicted at the current time step, the noise slice at the current time step is denoised by inverse reparameterization to obtain the denoised feature slice.
[0094] In an optional embodiment, after training the denoising network using the context conditional feature map as a priori guiding input, the method further includes: concatenating the context conditional feature map and the noise slice of the current time step along the channel dimension to obtain a fused initial input tensor; performing linear projection on the initial input tensor to obtain an aligned hidden layer tensor; inputting the aligned hidden layer tensor into the encoder of the denoising network to extract denoising features, and outputting the Gaussian noise predicted at the current time step.
[0095] For example, at the input of the denoising U-Net, the context conditional feature map Noisy slice at the current time step (The initial images are all Gaussian noise) They are directly concatenated along the channel dimension to obtain the fused initial input tensor. The feature dimension is expanded to .
[0096] Furthermore, to address the feature heterogeneity of the concatenated tensor across channels and prevent shallow networks from falling into distributional chaos, an initial feature projection layer is explicitly introduced at the input. This projection layer performs a weighted linear combination of channels with a 3×3 local receptive field. In essence, this convolutional kernel acts as a learnable affine transformation matrix, simultaneously mapping the low-level pixel intensity scalar and the high-dimensional semantic feature vector to the underlying hidden space of U-Net, generating an aligned hidden tensor. :
[0097]
[0098] in, For intermediate features of diffusion denoising networks, This represents the initial feature projection layer. This represents the number of hidden channels in the first hidden layer of U-Net. Through this initial projection operation, the network achieves deep semantic fusion of pure Gaussian noise distribution and rigorous 3D physical topology priors in the first layer. Subsequently, It was officially sent to the U-Net encoder to perform multi-scale noise reduction and feature extraction.
[0099] After deep feature extraction and time-step modulation of multi-scale residual blocks, U-Net finally outputs the Gaussian noise predicted by the network at the current time step. .
[0100] In an optional embodiment, after acquiring the predicted noise, the diffusion model progressively eliminates Gaussian noise in the image through a reverse reparameterization process. Specifically, during the inference phase, a denoising diffusion implicit model (DDIM) is employed to accelerate the sampling strategy, and its deterministic single-step reverse evolution formula is defined as:
[0101]
[0102] in, This is the denoised slice from the previous time step. The noise scheduling parameters are preset for the current time step. The noise scheduling parameters preset in the previous time step, This is a noise slice for the current time step. The Gaussian noise predicted at the current time step. For contextual conditional feature maps, Indicates the current time step.
[0103] In an optional embodiment, the constraints of the denoising network training process satisfy the following formula:
[0104]
[0105] in, This represents the mean square error loss of the denoising autoencoder. This represents a noisy sample in the latent space. This represents true Gaussian noise. Represents the standard normal distribution. This represents the Gaussian noise predicted at the current time step. This represents the noise sample at the current time step. Represents the contextual conditional feature map. Indicates the current time step. This represents the square of the L2 norm.
[0106] Understandably, this loss function measures the difference between predicted noise and actual noise, and by minimizing this difference, the diffusion model learns to inversely recover the original signal from the noisy data.
[0107] S53: Write the denoised feature slices into the initialized three-dimensional empty tensor to obtain the reconstructed high-resolution three-dimensional CT volume.
[0108] For example, the generated clean slice Write the initial three-dimensional empty tensor This yields the final reconstructed high-resolution three-dimensional CT volume.
[0109] This application provides a dual-view CT reconstruction method based on geometrically guided diffusion and frequency domain perception. The method acquires both anteroposterior and lateral X-ray projection data of the target object. The anteroposterior X-ray projection data is input into a dual-domain residual network based on discrete wavelet transform for feature extraction, yielding a first projection feature map. Similarly, the lateral X-ray projection data is input into the same network for feature extraction, yielding a second projection feature map. The two-dimensional features of the first and second projection feature maps are projected into three-dimensional space to obtain aligned three-dimensional spatial features. Three-dimensional spatial features from three adjacent feature slices are extracted to form a feature set. These features are then stitched and fused along the channel dimension to obtain a contextual conditional feature map. This contextual conditional feature map is used as a priori guided input diffusion model for analysis and processing, generating high-resolution three-dimensional CT volume data. This method not only recovers high-fidelity microscopic anatomical structures from a minimal amount of two-dimensional X-ray projections but also ensures strict consistency in three-dimensional space, providing a reliable technical solution for low-dose, low-cost, high-quality medical image reconstruction in clinical settings.
[0110] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.
[0111] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A dual-view CT reconstruction method based on geometrically guided diffusion and frequency domain sensing, characterized in that, The method includes: Acquire the frontal and lateral X-ray projection data of the target object; The frontal X-ray projection data is input into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain the first projection feature map. The lateral X-ray projection data is input into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain the second projection feature map. The two-dimensional features of the first projection feature map and the two-dimensional features of the second projection feature map are projected into three-dimensional space to obtain aligned three-dimensional space features. The three-dimensional spatial features of three adjacent feature slices are extracted to form a feature set. The three-dimensional spatial features in the feature set are spliced and fused along the channel dimension to obtain the context condition feature map. The contextual condition feature map is used as a priori guided input diffusion model for analysis and processing to generate high-resolution 3D CT volume data.
2. The method according to claim 1, characterized in that, The step of inputting the orthogonal X-ray projection data into a dual-domain residual network based on discrete wavelet transform for feature extraction to obtain a first projection feature map includes: The first intermediate feature map is obtained by downsampling the orthogonal X-ray projection data using two-dimensional discrete wavelet transform. The first intermediate feature map is processed sequentially through a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation function to obtain the second intermediate feature map; The second intermediate feature map is processed sequentially through a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function to obtain the third intermediate feature map; The third intermediate feature map is processed sequentially through a 1×1 convolutional layer and a batch normalization layer to obtain the fourth intermediate feature map; The input orthogonal X-ray projection data is processed through a 1×1 convolutional layer and then added to the fourth intermediate feature map to obtain the first projection feature map.
3. The method according to claim 1, characterized in that, The step of projecting the two-dimensional features of the first projection feature map and the two-dimensional features of the second projection feature map into three-dimensional space to obtain aligned three-dimensional spatial features includes: A physical mesh is constructed based on the preset target 3D CT resolution. Any feature slice that needs to be reconstructed in the physical mesh is uniformly sampled to obtain the absolute coordinates of multiple discrete 3D voxels on the slice. The absolute coordinates of three-dimensional voxels are transformed into a query matrix in the cross-attention mechanism through sine and cosine position encoding and linear mapping. Using the physical perspective projection matrix of an actual X-ray device, the absolute coordinates of three-dimensional voxels are back-projected along the real physical ray trajectory to the first projection feature map and the second projection feature map to obtain the first perspective projection coordinates and the second perspective projection coordinates. A first local perceptual feature block of a preset size is extracted from the first projection feature map with the first perspective projection coordinates as the center, and a second local perceptual feature block of a preset size is extracted from the second projection feature map with the second perspective projection coordinates as the center. After feature splicing and flattening, the first and second local sensing feature blocks are linearly mapped into a key matrix and a value matrix. The attention score is obtained by performing a dot product on the query matrix and the value matrix, and the attention score is normalized by the Softmax function to generate attention weights; Based on the attention weights, the value matrix is weighted and summed to obtain the high-dimensional physical features of the three-dimensional voxel points after cross-dimensional alignment. The high-dimensional physical features of all three-dimensional voxel points on the feature slice are combined to obtain the complete feature of the feature slice.
4. The method according to claim 1, characterized in that, The step of stitching and fusing the three-dimensional spatial features in the feature set along the channel dimension to obtain the context condition feature map includes: The feature maps of the previous layer, the current layer, and the next layer are concatenated and stitched together along the channel dimension to obtain a high-dimensional hybrid tensor. The high-dimensional hybrid tensor is processed by a 1×1 convolutional layer, normalization, and nonlinear activation function to compress the channel dimension of the high-dimensional hybrid tensor back to the target dimension, thus obtaining the context conditional feature map.
5. The method according to claim 1, characterized in that, The step of using the contextual condition feature map as a priori guided input diffusion model for analysis and processing to generate high-resolution three-dimensional CT volume data includes: An initial pure noise slice is randomly sampled from a standard normal distribution according to a preset number of sampling steps; The denoising network is trained by using the initial pure noise slice and the context conditional feature map as prior guided inputs. Based on the Gaussian noise predicted at the current time step, the noise slice at the current time step is denoised by inverse reparameterization to obtain the denoised feature slice. The denoised feature slices are written into the initialized three-dimensional empty tensor to obtain the reconstructed high-resolution three-dimensional CT volume.
6. The method according to claim 5, characterized in that, After training a denoising network using contextual conditional feature maps as prior guided input, the method further includes: The context conditional feature map and the noise slice at the current time step are concatenated along the channel dimension to obtain the fused initial input tensor; The aligned hidden layer tensor is obtained by linearly projecting the initial input tensor. The aligned hidden layer tensor is input into the encoder of the denoising network to extract denoising features, and the Gaussian noise predicted at the current time step is output.
7. The method according to claim 5, characterized in that, The Gaussian noise predicted based on the current time step is denoised by inverse reparameterization on the noise slice of the current time step to obtain a denoised feature slice that satisfies the following formula: ； in, This is the denoised slice from the previous time step. The noise scheduling parameters are preset for the current time step. The noise scheduling parameters preset in the previous time step, This is a noise slice for the current time step. The Gaussian noise predicted at the current time step. For contextual conditional feature maps, Indicates the current time step.
8. The method according to claim 5, characterized in that, The constraints for training the denoising network satisfy the following formula: ； in, This represents the mean square error loss of the denoising autoencoder. This represents a noisy sample in the latent space. This represents true Gaussian noise. Represents the standard normal distribution. This represents the Gaussian noise predicted at the current time step. This represents the noise sample at the current time step. Represents the contextual conditional feature map. Indicates the current time step. This represents the square of the L2 norm.