Low-dose CT image three-dimensional reconstruction enhancement method based on lightweight TriAD-Net network

Abstract

The application discloses a kind of low-dose CT image three-dimensional reconstruction enhancement methods based on light TriAD-Net network, comprising: the three-dimensional body data provided by low-dose Mayo Clinic LDCT Dataset public dataset is matched with full-dose / low-dose CT slice and is preprocessed, obtains training dataset and test dataset;TriAD-Net network model is constructed;TriAD-Net network model is trained on the training dataset obtained;The low-dose CT three-dimensional body data slice to be enhanced is input into the TriAD-Net network model trained, and the enhanced CT slice is output by model inference, and three-dimensional reconstruction is completed based on the slice enhanced.The application can accurately retain anatomical structure edge and tissue texture details, guarantee the integrity and accuracy of overall structure after three-dimensional reconstruction, significantly improve the clinical usability of low-dose CT three-dimensional reconstruction image.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and medical CT three-dimensional reconstruction technology, and more specifically, relates to a method for low-dose CT image three-dimensional reconstruction enhancement based on a lightweight TriAD-Net network. Background Technology

[0002] Computed tomography (CT) is an indispensable imaging tool in modern clinical diagnosis. CT scans are essentially three-dimensional volumetric scans, and the generated volumetric data is the core data source for clinical three-dimensional reconstruction (such as the three-dimensional structural reconstruction of organs and bones). Since the 1980s, the number of CT examinations has increased dramatically; in the United States alone, the annual number of CT examinations climbed from approximately 3 million in 1980 to nearly 70 million in 2007. CT three-dimensional reconstruction technology plays a crucial role in early disease screening, preoperative three-dimensional planning, and three-dimensional evaluation of treatment outcomes, and is widely recognized as one of the most important technological advancements in modern medicine.

[0003] However, the radiation dose from CT scans is increasingly concerning. A single routine chest CT scan delivers over 100 times the radiation dose of a standard anteroposterior or lateral chest X-ray. With the widespread use of high-dose CT techniques such as angiography, cardiac imaging, and multi-phase contrast-enhanced CT, patients' cumulative ionizing radiation exposure continues to increase, and long-term exposure may increase the risk of cancer. Therefore, minimizing radiation dose while ensuring the quality of 3D reconstructed images has become one of the core research directions in the field of medical imaging. Low-dose CT (LDCT) technology reduces radiation dose by lowering tube current, but this comes at the cost of significant noise and artifacts in the slices of the 3D volumetric data. This not only affects the visual quality of individual slices but also leads to blurred edges and artifact superposition in the reconstructed anatomical structures, severely reducing the clinical diagnostic value of the 3D reconstruction results.

[0004] Traditional methods for noise suppression in low-dose CT images mainly include median filtering, nonlocal mean denoising (NLM), and bilateral filtering. While these methods are simple in principle and do not require training data, they have inherent limitations: median filtering tends to blur tissue boundary details, disrupting the structural continuity of the 3D reconstruction; NLM denoising has high computational complexity and is parameter-sensitive, making it difficult to adapt to the processing of large-scale CT 3D volumetric data; bilateral filtering has limited edge preservation under strong noise conditions and cannot solve the problem of spatially uneven noise distribution in 3D volumetric data. Overall, traditional methods struggle to achieve a satisfactory balance between noise suppression and 3D detail preservation, and cannot adaptively handle heterogeneous noise from different layers and regions in CT 3D volumetric data.

[0005] With the rapid development of deep learning technology, image denoising methods based on neural networks have made significant progress. Among them, U-Net, with its encoder-decoder symmetric structure and skip connection mechanism, can effectively extract multi-scale features and maintain accurate spatial localization capabilities, and has been widely adopted in the field of medical image processing. However, U-Net and its improved variants usually have a large number of model parameters (generally over 3 million), requiring high computational resources and storage space. More importantly, these models are mostly designed for two-dimensional slices and do not fully consider the inter-slice correlation of CT three-dimensional volume data. When deployed in three-dimensional reconstruction scenarios, not only does the computational cost double, but it is also prone to causing the loss of inter-slice consistency of the three-dimensional structure. In addition, large parameter models are at risk of overfitting when training data is limited, and the acquisition of medical CT three-dimensional volume data is often restricted by factors such as ethical approval and privacy protection, and the available sample size is usually small.

[0006] Therefore, how to maintain the model's lightweight design while ensuring that the network can simultaneously consider the inter-slice correlation of CT 3D volumetric data and the ability to perceive multi-scale noise within slices, so as to achieve efficient enhancement of 3D reconstructed slices under low parameter conditions and even surpass the denoising performance of large networks, is a key technical problem that urgently needs to be solved in the field of low-dose CT 3D reconstruction image enhancement. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a method for 3D reconstruction enhancement of low-dose CT images based on a lightweight three-scale atrous denoising network (TriAD-Net). This method utilizes the Mayo Clinic LDCT Dataset public dataset (which contains full-dose / low-dose CT volumetric data from patients, including continuous slice sequences) to enhance the core unit of 3D reconstruction—the CT slice. By employing three parallel atrous convolutional branches, it simultaneously covers the receptive fields at three scales: local details, mid-range context, and long-range context. Furthermore, it incorporates a channel attention mechanism to adaptively fuse multi-scale features, achieving denoising performance that surpasses that of large encoder-decoder networks with several times the number of parameters with an extremely low parameter count (approximately 0.9M). This invention effectively suppresses multi-scale noise such as fine noise, stripe artifacts, and uneven regional noise in low-dose CT three-dimensional volumetric data slices, while accurately preserving the edges of anatomical structures and tissue texture details. It ensures the integrity and accuracy of the overall structure after three-dimensional reconstruction, significantly improving the clinical usability of low-dose CT three-dimensional reconstruction images and providing an efficient technical solution for balancing radiation dose optimization and three-dimensional reconstruction image quality assurance.

[0008] To address at least one of the aforementioned technical problems, according to one aspect of the present invention, a method for three-dimensional reconstruction and enhancement of low-dose CT images based on a lightweight TriAD-Net network is provided, comprising the following steps:

[0009] S1. Preprocess the 3D volumetric data provided by the low-dose Mayo Clinic LDCT Dataset public dataset by pairing full-dose / low-dose CT slices to obtain training and testing datasets; S2. Construct a TriAD-Net network model, including a shared feature extraction head, three-way parallel multi-scale dilated convolution branches, and a scale attention fusion module; S3. Train the TriAD-Net network model obtained in step S2 on the training dataset obtained in step S1; S4. Input the low-dose CT 3D volumetric data slices to be enhanced into the trained TriAD-Net network model, output the enhanced CT slices through model inference, and complete the 3D reconstruction based on the enhanced slices.

[0010] Step S1 includes the following steps:

[0011] S1.1 Obtain the Mayo Clinic LDCTDataset public dataset containing paired full-dose and low-dose CT scans with 3D volumetric data, and extract continuous CT slices from the dataset; perform grayscale normalization on each slice, uniformly scaling pixel values ​​to the 0,1 range, and uniformly setting the image resolution to 512×512 pixels; retain the 3D inter-slice index information of the slices; S1.2 Randomly divide the patients in the dataset according to random seeds, using an 8:2 ratio to divide them into training and test sets; S1.3 Perform data augmentation processing on the training set slices, including random 128×128 cropping, random horizontal flipping, random vertical flipping, and random 90° rotation.

[0012] Further, step S2 involves constructing the TriAD-Net network model, specifically as follows:

[0013] S2.1, Construct a shared feature extraction head, consisting of a single-layer 3×3 convolution and a ReLU activation function, to map a single-channel CT image into a multi-channel feature map, expressed as the formula:

[0014]

[0015] in For the input low-dose CT images, To share the initial feature map;

[0016] S2.2, construct three parallel dilated convolution branches, each branch consisting of 5 layers of "3×3 dilated convolution-batch normalization-ReLU" stacked together. The three branches adopt different dilation rates. , , To obtain effective receptive fields of different sizes without increasing the number of parameters;

[0017] The principle of three-way parallel dilated convolution branches is as follows:

[0018] For the Road branch ( Its void ratio is The feature extraction process can be expressed as a formula:

[0019]

[0020] in , Indicates the void ratio The 3×3 convolution operation, For batch normalization operations;

[0021] The outputs of the three branches are as follows: Corresponding to local detail features, (corresponding to mid-range context features) Corresponding remote context features;

[0022] S2.3, construct a scale attention fusion module, concatenate the outputs of the three branches along the channel dimension, compress the channels through 1×1 convolution, and then adaptively weight them through the channel attention mechanism;

[0023] The principle of the scale attention fusion module is as follows:

[0024] First, the three branch outputs are concatenated along the channel dimension and compressed using a 1×1 convolution, as shown in the formula:

[0025]

[0026] in This indicates a channel-level concatenation operation;

[0027] Then, the importance weight of each channel is calculated using the SE channel attention mechanism, expressed as the formula:

[0028]

[0029] in This is a global average pooling operation. and These are the learnable weight matrices for compressing and activating the fully connected layers, respectively. Use the Sigmoid activation function;

[0030] The final fused features are weighted by attention and superimposed with local residual connections, expressed as the formula:

[0031]

[0032] in This indicates a channel-by-channel multiplication operation. For the shared initial features in step S2.1, residual connections ensure that the network can learn identity mappings;

[0033] S2.4, construct the noise prediction tail and residual subtraction output; the tail consists of a 3×3 convolution layer, which maps the fused features to a single-channel noise estimation map, and the final output obtains the enhanced image through residual subtraction, expressed as the formula:

[0034]

[0035]

[0036] in The noise component estimated for the network. This is a CT image after noise reduction and enhancement.

[0037] Furthermore, in S3, the loss function used in the network training process is a weighted combination of L1 loss and structural similarity loss, expressed as the formula:

[0038]

[0039] in For the corresponding full-dose reference image, For structural similarity loss weights, The structure similarity index is used; the AdamW optimizer and cosine annealing learning rate scheduling strategy are employed for training.

[0040] Furthermore, in S3 and S4, the image enhancement effect evaluation metrics include: Peak Signal-to-Noise Ratio (PSNR); Structural Similarity Index (SSIM); and Root Mean Square Error (RMSE).

[0041] The calculation formulas are as follows:

[0042]

[0043]

[0044] .

[0045] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the low-dose CT image three-dimensional reconstruction enhancement method based on a lightweight TriAD-Net network of the present invention.

[0046] According to another aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the low-dose CT image three-dimensional reconstruction enhancement method based on a lightweight TriAD-Net network of the present invention.

[0047] Compared with existing technologies, the beneficial effects of the above-described method of the present invention are as follows:

[0048] Regarding multi-scale noise perception and 3D compatibility, this invention, based on the Mayo Clinic LDCTDataset dataset from 3D scanning, designs a three-way parallel dilated convolution for CT 3D volumetric data slices. It simultaneously extracts noise features at three scales—local, mid-range, and long-range—with three different dilation rates (1, 2, and 4), effectively covering receptive fields of approximately 35, 69, and 137 pixels, respectively. This design overcomes the limitations of traditional single-kernel receptive fields while ensuring the structural continuity of the 3D reconstruction after slice enhancement, comprehensively capturing multi-scale noise components from fine noise to large-scale artifacts in low-dose CT 3D volumetric data. In terms of adaptive feature fusion, this invention introduces an SE channel attention mechanism to adaptively weight multi-scale fusion features, enabling the network to dynamically adjust the contribution ratio of the three scale branches according to the noise characteristics of different regions, rather than simply adding fixed weights. This mechanism can specifically enhance key anatomical structural features in 3D reconstruction, significantly improving the enhancement effect of low-dose CT slices and ensuring the accuracy of the reconstructed structure. Regarding model lightweighting and 3D processing efficiency, the TriAD-Net network of this invention has only about 0.9M parameters, far less than encoder-decoder architectures such as U-Net (typically over 3M), yet it achieves denoising performance surpassing the latter. The extremely low parameter count and computational overhead make this model easy to batch process CT 3D volumetric data slices, enabling efficient slice enhancement before 3D reconstruction on computationally limited clinical workstations or embedded devices. It also exhibits stronger generalization ability under small-sample medical 3D data conditions, effectively reducing the risk of overfitting. In terms of residual learning strategy and 3D reconstruction assurance, this invention employs a dual residual mechanism of global noise residual prediction (output = input − noise) and local feature residual connection. This simplifies the network's learning objective while ensuring effective gradient propagation, improving training stability and convergence speed. The enhanced slices can be directly used for 3D reconstruction, better preserving anatomical structure edges and texture details compared to traditional methods, significantly improving the clinical usability of the 3D reconstruction results. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0050] Figure 1 This is a flowchart illustrating the overall method of a preferred embodiment of the present invention;

[0051] Figure 2 This is a diagram of the overall architecture of the TriAD-Net network according to a preferred embodiment of the present invention;

[0052] Figure 3 This is a comparison of the denoising results of TriAD-Net, a preferred embodiment of the present invention, with traditional methods and U-Net.

[0053] Figure 4 This is a basic discretization schematic diagram of a preferred embodiment of the present invention;

[0054] Figure 5 This is a bar chart comparing the PSNR / SSIM indices of various methods in this invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0056] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0057] Example 1:

[0058] like Figure 1-5 As shown, this invention provides a method for three-dimensional reconstruction and enhancement of low-dose CT images based on a lightweight TriAD-Net network, comprising the following steps:

[0059] S1. Preprocess the 3D volumetric data provided by the low-dose Mayo Clinic LDCT Dataset public dataset by pairing full-dose / low-dose CT slices to obtain training and testing datasets; S2. Construct a TriAD-Net network model, including a shared feature extraction head, three-way parallel multi-scale dilated convolution branches, and a scale attention fusion module; S3. Train the TriAD-Net network model obtained in step S2 on the training dataset obtained in step S1; S4. Input the low-dose CT 3D volumetric data slices to be enhanced into the trained TriAD-Net network model, output the enhanced CT slices through model inference, complete 3D reconstruction based on the enhanced slices, and compare the performance with the 3D reconstruction results of traditional methods and large network models;

[0060] Step S1 includes the following steps:

[0061] S1.1 Obtain the Mayo Clinic LDCTDataset public dataset containing paired full-dose and low-dose CT scans of 3D volumetric data, and extract continuous CT slices from the dataset; perform grayscale normalization on each slice, uniformly scaling pixel values ​​to the 0,1 range, and uniformly setting the image resolution to 512×512 pixels; retain the 3D inter-slice index information of the slices to ensure the sequence integrity of subsequent 3D reconstruction; S1.2 Randomly divide the patients in the dataset according to random seeds, using an 8:2 ratio to divide them into training and test sets, ensuring that all 3D volumetric data slices of the same patient appear only in the same set, avoiding data leakage and not destroying the correlation of 3D structures; S1.3 Perform data augmentation processing on the training set slices, including random 128×128 cropping, random horizontal flipping, random vertical flipping, and random 90° rotation, to increase the diversity of training samples;

[0062] Step S2 involves constructing the TriAD-Net network model, specifically as follows:

[0063] S2.1, Construct a shared feature extraction head, consisting of a single-layer 3×3 convolution and a ReLU activation function, to map a single-channel CT image into a multi-channel feature map, expressed as the formula:

[0064]

[0065] in For the input low-dose CT images, To share the initial feature map;

[0066] S2.2, construct three parallel dilated convolution branches, each branch consisting of 5 layers of "3×3 dilated convolution-batch normalization-ReLU" stacked together. The three branches adopt different dilation rates. , , To obtain effective receptive fields of different sizes without increasing the number of parameters;

[0067] The principle of the three-way parallel dilated convolution branch is as follows:

[0068] For the Road branch ( Its void ratio is The feature extraction process can be expressed as a formula:

[0069]

[0070] in , Indicates the void ratio 3×3 convolution operation, For batch normalization operations;

[0071] The outputs of the three branches are as follows: Corresponding to local detail features (effective receptive field approximately 35 pixels). Corresponding to mid-range contextual features (effective receptive field approximately 69 pixels). Corresponding remote context features (effective receptive field approximately 137 pixels);

[0072] S2.3, construct a scale attention fusion module, concatenate the outputs of the three branches along the channel dimension, compress the channels through 1×1 convolution, and then adaptively weight them through the channel attention mechanism (SE, Squeeze-and-Excitation);

[0073] The principle of the scale attention fusion module is as follows:

[0074] First, the three branch outputs are concatenated along the channel dimension and compressed using a 1×1 convolution, as shown in the formula:

[0075]

[0076] in This indicates a channel-level concatenation operation;

[0077] Then, the importance weight of each channel is calculated using the SE channel attention mechanism, expressed as the formula:

[0078]

[0079] in This is a global average pooling operation. and These are the learnable weight matrices for compressing and activating the fully connected layers, respectively. Use the Sigmoid activation function;

[0080] The final fused features are weighted by attention and superimposed with local residual connections, expressed as the formula:

[0081]

[0082] in This indicates a channel-by-channel multiplication operation. For the shared initial features in step S2.1, residual connections ensure that the network can learn identity mappings;

[0083] S2.4, construct the noise prediction tail and residual subtraction output; the tail consists of a 3×3 convolution layer, which maps the fused features to a single-channel noise estimation map, and the final output obtains the enhanced image through residual subtraction, expressed as the formula:

[0084]

[0085]

[0086] in The noise component estimated for the network. The image is a denoised and enhanced CT image.

[0087] Preferably, in S3, the loss function used in the network training process is a weighted combination of L1 loss and structural similarity loss, expressed as the formula:

[0088]

[0089] in For the corresponding full-dose reference image, For structural similarity loss weights, The structural similarity index is used; the AdamW optimizer and cosine annealing learning rate scheduling strategy are employed for training.

[0090] Preferably, in S3 and S4, the image enhancement effect evaluation metrics include peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and root mean square error (RMSE), and the calculation formulas are as follows:

[0091]

[0092]

[0093]

[0094] Example 2:

[0095] The computer-readable storage medium of this embodiment stores a computer program that, when executed by a processor, implements the steps in the low-dose CT image three-dimensional reconstruction and enhancement method based on a lightweight TriAD-Net network in Embodiment 1.

[0096] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0097] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0098] Example 3:

[0099] The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the low-dose CT image three-dimensional reconstruction and enhancement method based on a lightweight TriAD-Net network in Embodiment 1.

[0100] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0101] Those skilled in the art will understand that the content disclosed in the embodiments can be provided as a method, system, or computer program product. Therefore, this solution can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this solution can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage) containing computer-usable program code.

[0102] This solution is described with reference to flowchart illustrations and / or block diagrams of methods and computer program products according to embodiments of this solution. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0106] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the specific embodiments described above. The specific embodiments and descriptions in the specification are merely for further illustrating the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the claims and their equivalents.

Claims

1. A method for three-dimensional reconstruction and enhancement of low-dose CT images based on a lightweight TriAD-Net network, characterized in that, Includes the following steps: S1. Preprocess the 3D volumetric data provided by the low-dose Mayo Clinic LDCT Dataset public dataset by pairing full-dose / low-dose CT slices to obtain training and testing datasets; S2. Construct a TriAD-Net network model, including a shared feature extraction head, three-way parallel multi-scale dilated convolution branches, and a scale attention fusion module; S3. Train the TriAD-Net network model obtained in step S2 on the training dataset obtained in step S1; S4. Input the low-dose CT 3D volumetric data slices to be enhanced into the trained TriAD-Net network model, output the enhanced CT slices through model inference, and complete the 3D reconstruction based on the enhanced slices.

2. The method as described in claim 1, characterized in that, Step S1 includes the following steps: S1.1 Obtain the Mayo Clinic LDCT Dataset public dataset containing paired full-dose and low-dose CT scans with 3D volumetric data, and extract continuous CT slices from the dataset; perform grayscale normalization on each slice, uniformly scaling pixel values ​​to the 0,1 range, and uniformly setting the image resolution to 512×512 pixels; retain the 3D inter-slice index information of the slices; S1.2 Randomly divide the patients in the dataset according to random seeds, using an 8:2 ratio to divide them into training and test sets; S1.3 Perform data augmentation processing on the training set slices, including random 128×128 cropping, random horizontal flipping, random vertical flipping, and random 90° rotation.

3. The method as described in claim 2, characterized in that, Step S2 involves constructing the TriAD-Net network model, specifically as follows: S2.1, Construct a shared feature extraction head, consisting of a single-layer 3×3 convolution and a ReLU activation function, to map a single-channel CT image into a multi-channel feature map, expressed as the formula: in For the input low-dose CT images, To share the initial feature map; S2.2, construct three parallel dilated convolution branches, each branch consisting of 5 layers of "3×3 dilated convolution-batch normalization-ReLU" stacked together. The three branches adopt different dilation rates. , , To obtain effective receptive fields of different sizes without increasing the number of parameters; The principle of three-way parallel dilated convolution branches is as follows: For the Road branch ( Its void ratio is The feature extraction process can be expressed as a formula: in , Indicates the void ratio The 3×3 convolution operation, For batch normalization operations; The outputs of the three branches are as follows: Corresponding to local detail features, (corresponding to mid-range context features) Corresponding remote context features; S2.3, construct a scale attention fusion module, concatenate the outputs of the three branches along the channel dimension, compress the channels through 1×1 convolution, and then adaptively weight them through the channel attention mechanism; The principle of the scale attention fusion module is as follows: First, the three branch outputs are concatenated along the channel dimension and compressed using a 1×1 convolution, as shown in the formula: in This indicates a channel-level concatenation operation; Then, the importance weight of each channel is calculated using the SE channel attention mechanism, expressed as the formula: in This is a global average pooling operation. and These are the learnable weight matrices for compressing and activating the fully connected layers, respectively. Use the Sigmoid activation function; The final fused features are weighted by attention and superimposed with local residual connections, expressed as the formula: in This indicates a channel-by-channel multiplication operation. For the shared initial features in step S2.1, residual connections ensure that the network can learn identity mappings; S2.4, construct the noise prediction tail and residual subtraction output; the tail consists of a 3×3 convolution layer, which maps the fused features to a single-channel noise estimation map, and the final output obtains the enhanced image through residual subtraction, expressed as the formula: in The noise component estimated for the network. This is a CT image after noise reduction and enhancement.

4. The method as described in claim 3, characterized in that, In S3, the loss function used in the network training process is a weighted combination of L1 loss and structural similarity loss, expressed as the formula: in For the corresponding full-dose reference image, For structural similarity loss weights, The structure similarity index is used; the AdamW optimizer and cosine annealing learning rate scheduling strategy are employed for training.

5. The method as described in claim 4, characterized in that: In S3 and S4, the image enhancement effect evaluation metrics include: Peak signal-to-noise ratio, or PSNR; Structural similarity index, or SSIM; The root mean square error, or RMSE; The calculation formulas are as follows: 。 6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps in the method for three-dimensional reconstruction and enhancement of low-dose CT images based on a lightweight TriAD-Net network as described in any one of claims 1 to 5.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for three-dimensional reconstruction and enhancement of low-dose CT images based on a lightweight TriAD-Net network as described in any one of claims 1 to 5.

Images