A low-dose CT image reconstruction method and system based on multiple slice thickness data sets

Abstract

The present invention relates to a low-dose CT image reconstruction method and system based on multiple layer thickness data sets, belonging to the field of image processing. The method includes: pre-processing the collected multi-layer thick CT image data and transmitting it to the cloud, training and updating the model constructed in the cloud; reconstructing the input data according to the trained model, and automatically adjusting the window width and window position parameters according to the content and characteristics of the reconstructed image based on the adaptive window width and window position adjustment algorithm to improve the clarity and contrast of the lesion site in the image; adjusting the layer thickness fusion ratio of the CT image through three-dimensional reconstruction technology, and reconstructing the three-dimensional structure of the patient's body parts from different angles and levels. The present invention realizes independent feature extraction and complementary information mining of different layer thickness data by constructing parallel layer thickness specific coding branches and cross-branch dynamic feature fusion design, effectively integrating the high-resolution details of thin layers with the low-noise advantages of thick layers.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a low-dose CT image reconstruction method and system based on multiple slice thickness data sets. Background Art

[0002] The high radiation dose during traditional CT scans carries the risk of cancer, especially for patients who require multiple examinations. Therefore, reducing the radiation dose during CT scans is a key requirement for controlling radiation risks for patients.

[0003] Low-dose CT mainly reduces radiation dose by lowering tube voltage or tube current, but this method inevitably leads to a decline in image quality, which is specifically manifested as follows: 1. Noise artifacts: Dose reduction leads to a decrease in the number of photons, a decrease in the signal-to-noise ratio of the projection data, and speckle noise and streak artifacts in the reconstructed image; 2. Resolution and noise balance: Thin layer thickness (such as 1mm) images have high resolution, but more significant noise, thick layer thickness (such as 5mm) images have lower noise, but blurred details. How to fuse multiple layers of thick data is the key; 3. Clinical acceptability: The reconstruction results must meet the diagnostic standards of radiologists (such as the clarity of lung nodule boundaries and tissue contrast).

[0004] Traditional image reconstruction methods include filtered back projection (FBP), statistical iterative reconstruction, model-driven regularization, deep learning-driven reconstruction, multimodal / multi-scale data fusion, and public datasets and evaluation indicators. However, existing methods have significant defects in fusing complementary information of data of different layer thicknesses. Due to the difference in spatial resolution of images of different layer thicknesses, traditional cross-layer registration algorithms, such as interpolation methods or simple deformation registration algorithms, are prone to artifacts during the processing process, resulting in blurred image boundaries. Existing networks such as multi-scale UNet lack a mechanism for dynamically allocating weights of thin and thick layer features, making it impossible to effectively balance noise suppression and detail retention during the fusion process, resulting in information redundancy or conflict. Existing deep learning models rely heavily on training data, often limited to specific devices, scanning protocols, or patient populations (e.g., CT equipment parameter data from a single manufacturer). When applied across institutions and devices, model performance degrades significantly due to significant differences in noise patterns (e.g., quantum noise differences under different tube voltages) and hardware parameters (e.g., number of detector rows and field of view) between different devices. This significantly limits the widespread application of the technology in real-world clinical scenarios, requiring repeated model retraining to adapt to different devices, significantly increasing development and maintenance costs. Existing methods using public datasets and evaluation metrics rely heavily on subjective radiologists' scores (e.g., FROC curves) for validation, lacking objective, quantitative, and accurate metrics that accurately reflect key clinical diagnostic information. Traditional metrics such as PSNR and SSIM cannot accurately reflect image quality in critical diagnostic areas (e.g., contrast-to-noise ratio (CNR) of lung nodules). Furthermore, most datasets lack a pathological gold standard, making it difficult to accurately quantify the actual impact of reconstruction errors on clinical diagnosis. This leads to a misalignment between technology optimization and actual clinical needs, increasing the risk of technology transfer. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a low-dose CT image reconstruction method and system based on multiple layer thickness data sets, which solves the shortcomings of the prior art.

[0006] The object of the present invention is achieved by the following technical solution: a low-dose CT image reconstruction method based on multiple slice thickness data sets, the image reconstruction method comprising:

[0007] S1, pre-processing the collected multi-slice thick CT image data and transmitting it to the cloud to train and update the model built in the cloud;

[0008] S2. Reconstruct the input data according to the trained model, and automatically adjust the window width and window position parameters according to the content and characteristics of the reconstructed image using an adaptive window width and window position adjustment algorithm to improve the clarity and contrast of the lesion in the image;

[0009] S3. Use 3D reconstruction technology to adjust the layer thickness fusion ratio of CT images and reconstruct the 3D structure of the patient's body parts from different angles and levels.

[0010] The model performs image reconstruction on the input data, including:

[0011] Different encoders are used for data of different layer thicknesses. Each encoder samples the improved 3D U-Net structure to extract feature information of the corresponding layer thickness data.

[0012] A unified adaptive fusion decoder is set up, which receives feature data from different encoders and performs fusion and reconstruction processing.

[0013] The encoder specifically implements the following contents:

[0014] The improved 3D U-Net structure includes multiple downsampling and upsampling modules. , extract image features of different scales, where is the learnable convolution weight corresponding to the layer thickness t, is the ReLU activation function, is the number of channels of the input feature map, is the i-th channel of the input feature map, is the bias term of the tth layer, TH-Conv is the thickness-aware convolution kernel, and the thickness-aware convolution kernel is embedded in the convolution layer. By setting the convolution kernel parameters, it can adaptively extract effective features according to the noise distribution and characteristics of data with different layer thicknesses;

[0015] Introducing the layer thickness normalization module Perform feature alignment, where is the normalized feature, is the scaling factor used to adjust the amplitude of the feature, is the mean, is the standard deviation, is the smoothing term, is the bias term, for t Feature graph of the moment;

[0016] Constructing an attention gated fusion module Perform dynamic feature fusion, where Indicates the The attention output of each position, is the query vector, is the transpose of the key vector, is the normalization factor, represents the dimension of the key vector, represents a value vector, Represents the Softmax function;

[0017] pass Perform global average pooling on the feature maps output by each layer thickness, compress the two-dimensional feature maps into one-dimensional feature vectors, and concatenate these feature vectors, where: represents the dynamically calculated weight vector, MLP represents the multi-layer perceptron, GAP(·) represents the global average pooling, express The characteristic diagram of the moment, Represents a feature concatenation operation.

[0018] The decoder specifically implements the following contents:

[0019] Feature dimension adjustment: The feature vector output by feature fusion is adjusted to a dimension suitable for the subsequent reconstruction of the image through a series of linear transformations or convolution transposition operations;

[0020] Multi-scale feature processing: Convolution operations are performed through convolution kernels of different sizes or downsampling and upsampling operations are performed on features at different scales to obtain feature representations of different scales. These features of different scales are then fused so that the decoder can integrate information of different scales.

[0021] The adaptive window width and window level adjustment algorithm includes:

[0022]

[0023] in, is the grayscale value of the image after window width / window level adjustment, W c Indicates window level, W w Indicates the window width, Represents the original grayscale value of a pixel in the input image.

[0024] The total loss function during the model training process is:

[0025]

[0026] in, is the mean square error, which is used to represent the difference between the reconstructed image and the standard dose image. is the structural similarity loss, which is used to retain edge details. is the total variation regularization, for The hyperparameters of for The hyperparameters of for The hyperparameters of is the total loss function.

[0027] The S1 specifically includes the following contents:

[0028] According to the CT scanning protocol of different slice thicknesses, the CT scanner is controlled to emit low-dose X-rays to obtain projection data of the patient's body parts;

[0029] By simulating the reduced number of photons under low-dose conditions, the original data is downsampled and noise data is introduced to enhance the authenticity and diversity of the data. The window width and window position are adjusted according to medical imaging display standards to ensure that the image grayscale value distribution meets clinical diagnostic requirements and improve image contrast and readability. The format and parameters of data collected at different layer thicknesses are unified. A metadata file in JSON format is created to record scanning parameters, layer thickness information, and data acquisition time, and the files are mapped one-to-one with the preprocessed image data.

[0030] The pre-processed CT data packets are encrypted using medical digital imaging and communication protocols and then transmitted to the cloud. The cloud verifies the received data and uses the verified data to train the model.

[0031] A low-dose CT image reconstruction system based on multiple slice thickness data sets, the system comprising a terminal layer, an edge layer, a cloud platform layer, and an application layer;

[0032] The terminal layer is configured to be responsible for collecting low-dose multi-slice thick CT data, transmitting the collected raw data to the cloud platform layer, receiving the reconstructed images for post-processing and visualization operations, and providing doctors with remote access to the radiology information system;

[0033] The edge layer is configured as an AI acceleration box for deploying lightweight models, enabling fast inference and reconstruction of local data in single-device scenarios, and participating in local model training for federated learning in multi-terminal collaborative scenarios.

[0034] The cloud platform layer is configured to store CT data, utilize GPU server clusters for model training and optimization, employ federated learning algorithms for multi-terminal data collaborative training, and manage, deploy, and monitor the model to ensure efficient operation and real-time updates.

[0035] The application layer is configured to integrate with the hospital's existing information system through the radiology information system integration interface to achieve data sharing and business collaboration.

[0036] The cloud platform layer includes a data lake, a GPU server cluster consisting of multiple GPU servers, a cloud coordinator, and a parameter server;

[0037] The data lake stores CT data from various hospitals, including raw data, pre-processed data, and metadata.

[0038] Each GPU server acts as a training node, which contains a model for reconstructing images;

[0039] After receiving the data transmitted by the terminal layer, the cloud coordinator performs preliminary processing and classification on the data. Then, based on the resource usage and computing power of the AI ​​accelerator box at the edge layer and each training node, it distributes the training tasks to each training node. Each training node trains the model based on the federated learning algorithm. During the training process, each training node calculates the gradient of the model parameters and sends the gradient update information to the parameter server.

[0040] The parameter server receives gradient update information sent by each training node and the AI ​​accelerator box during multi-terminal collaborative training, implements aggregated updates of model parameters through summary calculations, and stores the aggregated updated model in the model warehouse. The model warehouse manages model versions and records model version information at different training stages and after optimization.

[0041] In the single-layer CT equipment scenario, the CT scanner transmits the preprocessed data to the AI ​​acceleration box deployed at the edge layer. The AI ​​acceleration box uses the built-in lightweight model to directly perform rapid inference and reconstruct the data, reducing the pressure on the cloud platform layer and quickly generating reconstructed images for use by the doctor workstation at the terminal layer.

[0042] The present invention has the following advantages: a low-dose CT image reconstruction method and system based on multiple layer thickness data sets, which realizes independent feature extraction and complementary information mining of different layer thickness data by constructing parallel layer thickness-specific coding branches and cross-branch dynamic feature fusion design. Specifically, the layer thickness-specific coding branch performs exclusive feature extraction based on the characteristics of each layer thickness data to avoid information interference; the cross-branch dynamic feature fusion design effectively integrates the high-resolution details of thin layers and the low-noise advantages of thick layers through adaptive weight distribution; with the help of the federated learning framework and the physical prior embedding strategy, the generalization performance of the model is effectively enhanced. The federated learning framework enables the model to learn a wider range of data distribution characteristics through cross-institutional and cross-device collaborative training, and the physical prior embedding integrates the physical principles of CT imaging into the model training process, so that the model output is more in line with actual physical laws. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the present application for which protection is claimed, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0045] One embodiment of the present invention relates to a low-dose CT (computed tomography) image reconstruction method based on multiple slice thickness datasets, employing a multi-input branch network architecture tailored to CT slice thickness differences. Parallel slice thickness-specific encoding branches are constructed to independently process data of different slice thicknesses, avoiding slice thickness interference. A cross-branch feature interaction module is also implemented to achieve effective feature interaction. A divide-and-conquer strategy is employed to ensure independent feature representation for each slice thickness while leveraging a dynamic feature fusion mechanism to exploit cross-slice complementary information. Jointly optimized reconstruction of dose reduction and slice thickness differences is achieved in a shared decoder. Furthermore, an innovative slice thickness adaptive normalization module is introduced to adaptively adjust the noise distribution of data from different slice thicknesses. A multi-scale attention gating mechanism is incorporated to balance the spatial resolution characteristics of data from different slice thicknesses. This solution effectively addresses the low efficiency of collaborative modeling of multi-slice thickness data, significantly improving the reconstruction consistency and detail fidelity of low-dose multi-slice CT data. It overcomes the dose-slice thickness coupling constraints of traditional methods in cross-slice reconstruction. It also offers improvements in model generalization, noise suppression, computational efficiency, and verification systems. Specifically, the following aspects are included:

[0046] like Figure 1 As shown, step 1, multi-slice thick CT data acquisition and preprocessing;

[0047] Data acquisition: Based on the CT scanner with different thicknesses (such as 1mm, 3mm and 5mm), the CT scanner is controlled to emit low-dose X-rays to obtain projection data of the patient's body parts.

[0048] Preprocessing operations include downsampling, normalization, and slice thickness matching;

[0049] Downsampling: By simulating the characteristics of reduced number of photons under low-dose conditions, the original data is downsampled and noise data that conforms to the actual situation is introduced to enhance the authenticity and diversity of the data.

[0050] Normalization processing: According to the medical image display standards, adjust the window width and window position to make the image grayscale value distribution meet the clinical diagnosis requirements and improve the image contrast and readability.

[0051] Slice thickness matching: Unify the format and parameters of data collected at different slice thicknesses to ensure compatibility of subsequent data processing.

[0052] Metadata generation: Create a metadata file in JSON format, which records key information such as scanning parameters (tube voltage kVp, tube current mAs, etc.), slice thickness information, and data acquisition time, and corresponds one-to-one with the preprocessed image data.

[0053] Step 2: After preprocessing is completed, the transmission is automatically triggered and the data is transmitted to the cloud platform layer;

[0054] Data encryption transmission: Using the DICOM (Digital Imaging and Communications in Medicine) protocol, the pre-processed CT data packet (including image matrix and slice thickness metadata) is encrypted and uploaded to the cloud server via 5G or WiFi6 network to ensure the security and integrity of the data during transmission.

[0055] Endpoint resume mechanism: To address network instability and other issues, we've introduced breakpoint resume technology. If a transmission is interrupted, it can be continued from the breakpoint, improving data transmission efficiency and avoiding duplicate transmissions.

[0056] Data Verification: The cloud receiver deploys a data verification module to verify the integrity of the received data by calculating the MD5 hash value. If the verification fails, a retransmission request is sent to the CT device terminal to ensure data accuracy.

[0057] Step 3: Build and train a multi-input and multi-branch network model;

[0058] Encoder Module: Dedicated encoders are designed for data with different slice thicknesses (1mm, 3mm, and 5mm). Each encoder utilizes a modified 3D U-Net structure to fully extract the feature information of the corresponding slice thickness. U-Net is a classic convolutional neural network with an encoder-decoder architecture, widely used in medical image segmentation tasks.

[0059] Among them, the encoder of the 3D U-Net structure can also be replaced with a Transformer-based encoder to capture long-distance dependencies; the Transformer architecture has a natural advantage in processing global information and long-distance dependencies, and can better maintain the overall structural consistency of the image. It is particularly suitable for scenarios that require attention to large-scale structural features, such as overall lung morphology analysis.

[0060] Decoder module: Set up a unified adaptive fusion decoder to receive feature data from different encoders and perform fusion and reconstruction processing.

[0061] Furthermore, in the branch encoding stage: each layer thickness branch sampling improved 3D U-Net structure includes multiple downsampling and upsampling modules, through , extract image features of different scales, where is the learnable convolution weight corresponding to the layer thickness t, is the ReLU activation function, is the number of channels of the input feature map. For CT images, the number of channels is usually 1. is the i-th channel of the input feature map, is the bias term of the tth layer, TH-Conv is the thickness-aware convolution kernel, and the thickness-aware convolution kernel is embedded in the convolution layer. By setting the convolution kernel parameters, it can adaptively extract effective features according to the noise distribution and characteristics of data with different layer thicknesses. For example, for thin layer (1mm) data, TH-Conv pays more attention to the extraction of detail features; for thick layer (5mm) data, it focuses on suppressing noise and extracting macroscopic structural features.

[0062] In the feature alignment stage: introduce the layer thickness normalization module Perform feature alignment, where is the normalized feature, is the scaling factor used to adjust the amplitude of the feature, is the mean, is the standard deviation, is the smoothing term, is the bias term.

[0063] In the dynamic feature fusion stage of cross-branch dynamic feature fusion, an attention gate fusion module is constructed Perform dynamic feature fusion, where Indicates the The attention output of each position can also be understood as the weighted feature representation or the fusion weight generation of the attention aggregation result of all key-value pairs of the current query. is the query vector, is the transpose of the key vector, is the normalization factor, represents the dimension of the key vector, represents a value vector, Represents the Softmax function;

[0064] pass Perform global average pooling on the feature maps output by each layer thickness, compress the two-dimensional feature maps into one-dimensional feature vectors, and concatenate these feature vectors, where: Represents a dynamically calculated weight vector, used to represent the current moment's attention or feature importance. MLP represents a multi-layer perceptron, used to perform nonlinear transformations on input features and extract high-level features. GAP(·) represents global average pooling, used to extract global information from high-dimensional feature maps and compress them into low-dimensional vectors. and Represents the feature graphs of different moments or different models, Represents the feature concatenation operation, which concatenates two feature vectors into a comprehensive feature vector according to their dimensions.

[0065] The adaptive fusion decoder stage of cross-branch dynamic feature fusion includes: feature dimension adjustment: the feature vector output by feature fusion is adjusted to the dimension suitable for the subsequent reconstructed image through a series of linear transformations or convolution transposition operations;

[0066] Multi-scale feature processing: Convolution operations are performed through convolution kernels of different sizes or downsampling and upsampling operations are performed on features at different scales to obtain feature representations of different scales. These features of different scales are then fused so that the decoder can integrate information of different scales.

[0067] Step 4: Joint optimization reconstruction and loss calculation;

[0068] The total loss function of the designed network model during training is:

[0069] ,

[0070] in, is the mean square error, which is used to represent the difference between the reconstructed image and the standard dose image. is the structural similarity loss, which is used to retain edge details. is the total variation regularization, are all hyperparameters.

[0071] Step 5: Post-processing and visualization of reconstructed images;

[0072] Adopting adaptive window width and window level adjustment algorithm:

[0073] ,

[0074] in, is the grayscale value of the image after window width / window level adjustment, W c Indicates the window level, which is used to control the center position of the display window and corresponds to the CT value threshold of the region of interest. wIndicates the window width, which is used to control the range of the display window and determine the grayscale stretching amplitude. Represents the original grayscale value of a pixel in the input image.

[0075] According to the content and characteristics of the reconstructed image, the window width and window position parameters are automatically adjusted to make the details of the area of ​​interest in the image (such as the lungs and lesion sites) clearer, improve the contrast and readability of the image, and facilitate doctors' observation and diagnosis.

[0076] 3D Reconstruction and Visualization: A 3D reconstruction module is integrated into the PACS (Picture Archiving and Communication System), supporting technologies such as MIP (Maximum Intensity Projection) and MPR (Multi-Planar Reconstruction). Using an interactive interface, doctors can adjust the slice thickness and fusion ratio to visualize the 3D structure of the patient's body from different angles and levels, providing more comprehensive information for diagnosis and treatment planning.

[0077] Another embodiment of the present invention relates to a low-dose CT image reconstruction system based on multiple slice thickness datasets, the system comprising a terminal layer, an edge layer, a cloud platform layer, and an application layer;

[0078] The terminal layer is responsible for collecting low-dose multi-slice thick CT data, transmitting the collected raw data to the cloud platform layer, receiving the reconstructed images for post-processing and visualization operations, and providing doctors with remote access to the radiology information system.

[0079] Edge layer: AI accelerator boxes are configured to deploy lightweight models. They enable fast inference and reconstruction of local data in single-device scenarios and participate in local model training for federated learning in multi-device collaboration scenarios. Based on the actual application requirements (such as real-time reconstruction in single-device scenarios or inference requirements in multi-device collaboration), the appropriate model version is selected from the model repository and deployed to the AI ​​accelerator boxes at the edge layer. After receiving the model, the AI ​​accelerator boxes can perform local data inference and reconstruction operations.

[0080] Cloud platform layer: This layer is configured to store CT data and utilize GPU server clusters for model training and optimization. It uses federated learning algorithms to achieve multi-terminal data collaborative training. It is also responsible for model management, deployment, and monitoring to ensure efficient model operation and real-time updates.

[0081] The application layer integrates the entire low-dose CT image reconstruction solution with the hospital's existing information systems through the PACS / RIS system integration interface. This enables data sharing and business collaboration throughout the entire process, from data acquisition, transmission, processing, to final diagnostic application. This ensures smooth information flow between various systems within the hospital, improving the efficiency and quality of medical services.

[0082] Furthermore, the terminal layer includes CT scanners, doctor workstations, and mobile terminals;

[0083] The CT scanner is responsible for collecting low-dose, multi-layer CT data (such as 1mm, 3mm, and 5mm); the doctor's workstation is used to receive reconstructed images, perform post-processing and visualization operations, and support doctors in diagnosis and analysis; the mobile terminal allows doctors to remotely access the PACS system, view and read films, and realize mobile office.

[0084] The cloud platform layer includes a data lake, a GPU server cluster consisting of multiple GPU servers, a cloud coordinator, and a parameter server;

[0085] The data lake stores CT data from various hospitals, including raw data, pre-processed data, and metadata.

[0086] Each GPU server acts as a training node and contains a model for reconstructing images. The image reconstruction process is based on a low-dose CT image reconstruction method based on multiple slice thickness datasets.

[0087] After receiving data transmitted by the terminal layer, the cloud coordinator performs preliminary processing and classification on the data. Then, based on the resource usage and computing power of the AI ​​accelerator boxes at the edge layer and each training node, it distributes the training tasks to each training node. Each training node trains the model based on the federated learning algorithm. During the training process, each training node calculates the gradient of the model parameters and sends the gradient update information to the parameter server.

[0088] The parameter server receives gradient update information from each training node and the AI ​​accelerator box during multi-terminal collaborative training. It implements aggregate updates of model parameters through summary calculations and stores the aggregated updated model in the model warehouse. The model warehouse manages model versions and records model version information at different training stages and after optimization.

[0089] In the single-layer CT equipment scenario, the CT scanner transmits the preprocessed data to the AI ​​acceleration box deployed at the edge layer. The AI ​​acceleration box uses the built-in lightweight model to directly perform rapid inference and reconstruct the data, reducing the pressure on the cloud platform layer and quickly generating reconstructed images for use by the doctor workstation at the terminal layer.

[0090] The entire system incorporates an exception handling mechanism. When data integrity checks fail, a retransmission protocol is triggered, sending a retransmission request to the data sender (such as a CT terminal or edge computing device) to ensure accurate data transmission to the target location. During model inference, if an exception such as an inference timeout occurs, a backup model instance is activated to continue providing inference services, ensuring system stability and reliability. Sandbox technology is used to strictly isolate data access rights across hospitals, preventing data leaks and unauthorized access, and safeguarding patient privacy and data security.

[0091] The foregoing description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein and should not be construed as excluding other embodiments. Rather, the present invention is capable of various other combinations, modifications, and improvements, and is capable of modifications within the scope of the concepts described herein, through the above teachings, or through techniques or knowledge in the relevant fields. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention are intended to be within the scope of the appended claims.

Claims

1. A low-dose CT image reconstruction method based on a multi-slice thickness dataset, characterized by: The image reconstruction method comprises: S1, pre-processing the collected multi-slice thick CT image data and transmitting it to the cloud to train and update the model built in the cloud; S2. Reconstruct the input data according to the trained model, and automatically adjust the window width and window position parameters according to the content and characteristics of the reconstructed image using an adaptive window width and window position adjustment algorithm to improve the clarity and contrast of the lesion in the image; S3. Adjust the slice thickness fusion ratio of the CT image through 3D reconstruction technology to reconstruct the 3D structure of the patient's body parts from different angles and levels; Among them, three-dimensional reconstruction technology includes MIP and MPR technology. Doctors use the interactive interface to adjust the layer thickness fusion ratio and observe the three-dimensional structure of the patient's body parts from different angles and levels, providing more comprehensive information for diagnosis and treatment plan formulation.

2. The low-dose CT image reconstruction method based on multiple slice thickness datasets according to claim 1, characterized in that: The model performs image reconstruction on the input data, including: Different encoders are used for data of different layer thicknesses. Each encoder samples the improved 3D U-Net structure to extract feature information of the corresponding layer thickness data. A unified adaptive fusion decoder is set up, which receives feature data from different encoders and performs fusion and reconstruction processing.

3. The low-dose CT image reconstruction method based on multiple slice thickness datasets according to claim 2, characterized in that: The encoder specifically implements the following contents: The improved 3D U-Net structure includes multiple downsampling and upsampling modules. , extract image features of different scales, where is the learnable convolution weight corresponding to the layer thickness t, is the ReLU activation function, is the number of channels of the input feature map, is the i-th channel of the input feature map, is the bias term of the tth layer, TH-Conv is the thickness-aware convolution kernel, and the thickness-aware convolution kernel is embedded in the convolution layer. By setting the convolution kernel parameters, it can adaptively extract effective features according to the noise distribution and characteristics of data with different layer thicknesses; Introducing the layer thickness normalization module Perform feature alignment, where is the normalized feature, is the scaling factor used to adjust the amplitude of the feature, is the mean, is the standard deviation, is the smoothing term, is the bias term, for t Feature graph of the moment; Constructing an attention gated fusion module Perform dynamic feature fusion, where Indicates the The attention output of each position, is the query vector, is the transpose of the key vector, is the normalization factor, represents the dimension of the key vector, represents a value vector, Represents the Softmax function; pass Perform global average pooling on the feature maps output by each layer thickness, compress the two-dimensional feature maps into one-dimensional feature vectors, and concatenate these feature vectors, where: represents the dynamically calculated weight vector, MLP represents the multi-layer perceptron, GAP(·) represents the global average pooling, express The characteristic diagram of the moment, Represents a feature concatenation operation.

4. The low-dose CT image reconstruction method based on multiple slice thickness datasets according to claim 3, characterized in that: The decoder specifically implements the following contents: Feature dimension adjustment: The feature vector output by feature fusion is adjusted to a dimension suitable for the subsequent reconstruction of the image through a series of linear transformations or convolution transposition operations; Multi-scale feature processing: Convolution operations are performed through convolution kernels of different sizes or downsampling and upsampling operations are performed on features at different scales to obtain feature representations of different scales. These features of different scales are then fused so that the decoder can integrate information of different scales.

5. The low-dose CT image reconstruction method based on multiple slice thickness data sets according to claim 1, characterized in that: The adaptive window width and window level adjustment algorithm includes: in, is the grayscale value of the image after window width / window level adjustment, W c Indicates window level, W w Indicates the window width, Represents the original grayscale value of a pixel in the input image.

6. The low-dose CT image reconstruction method based on multiple slice thickness datasets according to claim 1, characterized in that: The total loss function during the model training process is: in, is the mean square error, which is used to represent the difference between the reconstructed image and the standard dose image. is the structural similarity loss, which is used to retain edge details. is the total variation regularization, for The hyperparameters of for The hyperparameters of for The hyperparameters of is the total loss function.

7. The low-dose CT image reconstruction method based on multiple slice thickness datasets according to any one of claims 1 to 6, characterized in that: The S1 specifically includes the following contents: According to the CT scanning protocol of different slice thicknesses, the CT scanner is controlled to emit low-dose X-rays to obtain projection data of the patient's body parts; By simulating the reduced number of photons under low-dose conditions, the original data is downsampled and noise data is introduced to enhance the authenticity and diversity of the data. The window width and window position are adjusted according to medical imaging display standards to ensure that the image grayscale value distribution meets clinical diagnostic requirements and improve image contrast and readability. The format and parameters of data collected at different layer thicknesses are unified. A metadata file in JSON format is created to record scanning parameters, layer thickness information, and data acquisition time, and the files are mapped one-to-one with the preprocessed image data. The pre-processed CT data packets are encrypted using medical digital imaging and communication protocols and then transmitted to the cloud. The cloud verifies the received data and uses the verified data to train the model.

8. A low-dose CT image reconstruction system based on multiple slice thickness datasets, characterized by: The system includes a terminal layer, an edge layer, a cloud platform layer and an application layer; The terminal layer is configured to be responsible for collecting low-dose multi-slice thick CT data, transmitting the collected raw data to the cloud platform layer, receiving the reconstructed images for post-processing and visualization operations, and providing doctors with remote access to the radiology information system; The edge layer is configured as an AI acceleration box for deploying lightweight models, enabling fast inference and reconstruction of local data in single-device scenarios, and participating in local model training for federated learning in multi-terminal collaborative scenarios. The cloud platform layer is configured to store CT data, utilize GPU server clusters for model training and optimization, employ federated learning algorithms for multi-terminal data collaborative training, and manage, deploy, and monitor the model to ensure efficient operation and real-time updates. The application layer is configured to integrate with the hospital's existing information system through the radiology information system integration interface to achieve data sharing and business collaboration; The cloud platform layer includes a data lake, a GPU server cluster consisting of multiple GPU servers, a cloud coordinator, and a parameter server; The data lake stores CT data from various hospitals, including raw data, pre-processed data, and metadata. Each GPU server acts as a training node, which contains a model for reconstructing images; After receiving the data transmitted by the terminal layer, the cloud coordinator performs preliminary processing and classification on the data. Then, based on the resource usage and computing power of the AI ​​accelerator box at the edge layer and each training node, it distributes the training tasks to each training node. Each training node trains the model based on the federated learning algorithm. During the training process, each training node calculates the gradient of the model parameters and sends the gradient update information to the parameter server. The parameter server receives gradient update information sent by each training node and the AI ​​accelerator box during multi-terminal collaborative training, implements aggregated updates of model parameters through summary calculations, and stores the aggregated updated model in the model warehouse. The model warehouse manages model versions and records model version information at different training stages and after optimization.

9. The low-dose CT image reconstruction system based on multiple slice thickness datasets according to claim 8, characterized in that: In the single-layer CT equipment scenario, the CT scanner transmits the preprocessed data to the AI ​​acceleration box deployed at the edge layer. The AI ​​acceleration box uses the built-in lightweight model to directly perform rapid inference and reconstruct the data, reducing the pressure on the cloud platform layer and quickly generating reconstructed images for use by the doctor workstation at the terminal layer.

Images