Tomoscan control method and device, tomoscan equipment and storage medium

Abstract

The application discloses a fault scanning control method and device, a fault scanning equipment and a storage medium. The method comprises the following steps: identifying a key anatomical structure of an initial projection image based on a pre-trained trajectory planning model, generating a probability distribution map, determining a sparse scanning angle sequence and corresponding first scanning parameters, densely sampling and appropriately increasing a radiation dose in a high-weight area, and sparsely sampling and reducing a dose in a low-weight area; controlling the fault scanning equipment to perform scanning, obtaining first projection data and reconstructing the first projection data into a sparse reconstruction image; performing multi-modal fusion and scoring on the sparse reconstruction image and the current scanning parameters through an image quality evaluation model, adaptively optimizing scanning instructions according to the probability distribution map, a fusion feature vector, an image quality score, an equipment state and historical action data, generating second scanning parameters, making the scanning trajectory more continuous and smooth through an artificial intelligence technology, and realizing regional adaptive regulation and control of angles, doses and paths, so that the imaging resolution can be improved.

Description

Technical Field

[0001] This invention relates to the field of computed tomography (CT) technology, and in particular to a CT control method, apparatus, CT equipment, and storage medium. Background Technology

[0002] Computed tomography (CT) technology, especially cone-beam computed tomography (CBCT), has been widely used in human clinical practice, veterinary medicine, and pathology. Currently, the mainstream structures of CBCT equipment are C-type and O-type. Some CBCT equipment uses dual robotic arms as the drive structure, but the high difficulty of controlling dual robotic arms results in image quality that cannot meet the requirements of high-resolution imaging. With the development of AI-optimized operating systems, AI middleware, function libraries, computer vision and audiovisual software, biometric recognition software, and other application software development technologies, CBCT equipment is expected to achieve further technological advancements in motion and imaging control. Summary of the Invention

[0003] The present invention aims to at least solve one of the technical problems existing in the prior art. To this end, the present invention proposes a tomographic scanning control method, apparatus, tomographic scanning device, and storage medium, which can improve the imaging resolution of tomographic scanning.

[0004] In a first aspect, embodiments of the present invention provide a tomographic scanning control method, applied to a tomographic scanning device with dual robotic arms, comprising:

[0005] Acquire initial projection image data and initial scanning parameters of the computed tomography device;

[0006] Based on a pre-trained trajectory planning model, key anatomical structures are identified in the initial projected image data, a probability distribution map representing the importance of each anatomical region is generated, and a sparse scanning angle sequence and the first scanning parameter corresponding to each scanning angle are determined according to the probability distribution map and the initial scanning parameters.

[0007] Based on the sparse scanning angle sequence and the first scanning parameters, the tomography device is controlled to perform a tomography operation to obtain first projection image data.

[0008] Three-dimensional image reconstruction is performed based on the first projected image data to generate a sparse reconstructed image;

[0009] Using a pre-trained image quality assessment model, multimodal feature fusion is performed on the sparse reconstructed image and the current scanning parameters of the tomographic scanning device, and the image quality score and corresponding fusion feature vector are output based on the fusion result.

[0010] Based on the probability distribution map, the fusion feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device, the scanning action command is adaptively optimized to obtain the optimized second scanning parameters.

[0011] According to some embodiments of the present invention, determining the sparse scanning angle sequence and the first scanning parameter corresponding to each scanning angle based on the probability distribution map and the initial scanning parameters includes:

[0012] Based on the probability distribution map, a forward projection simulation is performed to calculate the sum of voxel probabilities along the ray path at each candidate scanning angle;

[0013] The sum of voxel probabilities for each candidate scanning angle is normalized to obtain an angle weight sequence. Each weight in the angle weight sequence is used to characterize the contribution of the corresponding projection angle to the reconstruction quality of key anatomical structures.

[0014] Based on the Poisson disk sampling algorithm, the angle weight sequence is used as the local density control factor to generate dense sampling points in high-weight regions and sparse sampling points in low-weight regions, thus obtaining a sparse scanning angle sequence.

[0015] The initial scanning parameters are dynamically modulated according to the angle weight sequence to generate first scanning parameters corresponding to each angle in the sparse scanning angle sequence.

[0016] According to some embodiments of the present invention, the initial scanning parameters include tube current and scanning rotation speed, and the step of dynamically modulating the initial scanning parameters according to the angle weight sequence to generate first scanning parameters corresponding to each angle in the sparse scanning angle sequence includes:

[0017] Based on the angle weight sequence, the tube current and the scanning rotation speed in the initial scanning parameters are dynamically modulated to generate first scanning parameters corresponding to each angle in the sparse scanning angle sequence.

[0018] According to some embodiments of the present invention, the step of performing three-dimensional image reconstruction based on the first projected image data to generate a sparse reconstructed image includes:

[0019] Using a physical reconstruction module based on iterative coordinate descent, the first projected image data is iteratively optimized. The reconstructed image is updated on a voxel-by-voxel basis to minimize the projection domain error. Combined with projection data consistency constraints and total variation regularization terms, a preliminary reconstructed image that satisfies physical measurement constraints is generated.

[0020] The multi-scale feature enhancement module based on 3D Swin Transformer is used to perform multi-scale feature enhancement processing on the preliminary reconstructed image. In this process, the contextual information from local texture to global long-range anatomical structure is extracted layer by layer through self-attention calculation within the local window and the cross-window shifting window mechanism, thereby improving the structural details and perceptual quality of the image.

[0021] At each feature scale, the scanning parameters of the tomography device are encoded into conditional vectors by an adaptive feature gating unit based on a geometric perception attention mechanism. The feature fusion weights of the physical reconstruction module and the multi-scale feature enhancement module are dynamically modulated according to the conditional vectors to fuse the features output by the physical reconstruction module and the multi-scale feature enhancement module.

[0022] According to some embodiments of the present invention, the step of using a pre-trained image quality assessment model to perform multimodal feature fusion on the sparse reconstructed image and the current scanning parameters of the tomographic scanning device, and outputting an image quality score and a corresponding fusion feature vector based on the fusion result, includes:

[0023] The 3D convolutional module of the pre-trained image quality assessment model is used to extract features from the sparse reconstructed image to generate a first image feature representing the dynamic structure information between frames.

[0024] The first image features are input into a residual network with an embedded channel attention module for feature enhancement to obtain the second image features;

[0025] The current scanning parameters of the tomography device are nonlinearly mapped by a multilayer perceptron to generate scanning parameter features that match the second image feature dimension.

[0026] Based on the cross-attention mechanism, attention weights are determined according to the second image features and the scanning parameter features, and the second image features are weighted and modulated to obtain preliminary fused features;

[0027] The preliminary fusion features are input into the pyramid pooling module to generate and output the fusion feature vector;

[0028] An image quality score is obtained by activating the fused feature vector using an activation function.

[0029] According to some embodiments of the present invention, the step of adaptively optimizing the scanning action command based on the probability distribution map, the fused feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device to obtain optimized second scanning parameters includes:

[0030] A fusion state vector is generated based on the probability distribution map, the fusion feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device;

[0031] Based on the SAC algorithm, a scanning action command is generated according to the fusion state vector, and the scanning action command is adaptively optimized by maximizing the multi-objective reward function to obtain the optimized second scanning parameters.

[0032] According to some embodiments of the present invention, the step of adaptively optimizing the scanning action command by maximizing a multi-objective reward function to obtain optimized second scanning parameters includes:

[0033] The multi-objective reward function is determined by weighted summation based on the current image quality score, the reciprocal of the scan dose, and the continuous motion stability index.

[0034] The scanning action command is adaptively optimized by maximizing the multi-objective reward function to obtain the optimized second scanning parameters.

[0035] Secondly, embodiments of the present invention provide a tomographic scanning control device, applied to a tomographic scanning equipment with dual robotic arms, comprising:

[0036] The data acquisition module is used to acquire initial projection image data and initial scanning parameters of the computed tomography device;

[0037] The trajectory planning module is used to identify key anatomical structures in the initial projected image data based on a pre-trained trajectory planning model, generate a probability distribution map representing the importance of each anatomical region, and determine a sparse scanning angle sequence and a first scanning parameter corresponding to each scanning angle based on the probability distribution map and the initial scanning parameters.

[0038] The scanning control module is used to control the tomography scanning device to perform tomography scanning operation according to the sparse scanning angle sequence and the first scanning parameters, so as to obtain the first projection image data;

[0039] The image reconstruction module is used to perform three-dimensional image reconstruction based on the first projected image data to generate a sparse reconstructed image;

[0040] The quality assessment module is used to perform multimodal feature fusion on the sparse reconstructed image and the current scanning parameters of the tomographic scanning device using a pre-trained image quality assessment model, and output an image quality score and corresponding fusion feature vector based on the fusion result;

[0041] The parameter optimization module is used to adaptively optimize the scanning action command based on the probability distribution map, the fusion feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device, to obtain the optimized second scanning parameters.

[0042] Thirdly, embodiments of the present invention provide a computed tomography (CT) scanning device, including a processor and a memory, wherein the memory stores a computer program, and the processor runs the computer program to implement the aforementioned computed tomography control method.

[0043] Fourthly, embodiments of the present invention provide a storage medium storing a computer program, which, when run, implements the above-described tomographic scanning control method.

[0044] The embodiments of the present invention have at least the following beneficial effects:

[0045] A pre-trained trajectory planning model is used to identify key anatomical structures in the initial projection image, generating a probability distribution map representing the importance of each region. Based on this, a sparse scanning angle sequence and corresponding first scanning parameters are determined. Dense sampling and appropriately increased radiation dose are used in high-weight regions, while sparse sampling and reduced dose are used in low-weight regions. The tomographic scanning device is controlled to perform scanning, obtaining the first projection data and reconstructing it into a sparse reconstructed image. Then, a multimodal fusion and scoring of the sparse reconstructed image and the current scanning parameters are performed by an image quality assessment model. Based on the probability distribution map, fused feature vector, image quality score, device status, and historical action data, the scanning command is adaptively optimized to generate second scanning parameters. This makes the scanning trajectory more continuous and smooth, and enables regional adaptive control of angle, dose, and path, which is beneficial to improving imaging resolution.

[0046] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0047] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0048] Figure 1 This is a schematic diagram of the structure of a tomography scanning device according to an embodiment of the present invention;

[0049] Figure 2 This is a flowchart illustrating the steps of the tomographic scanning control method according to an embodiment of the present invention.

[0050] Figure 3 This is a schematic block diagram of the tomographic scanning control device according to an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of the tomographic scanning device according to an embodiment of the present invention.

[0052] Figure label:

[0053] First robotic arm 110, second robotic arm 120, X-ray source 130, detector 140, rotating stage 150, data acquisition module 210, trajectory planning module 220, scanning control module 230, image reconstruction module 240, quality assessment module 250, parameter optimization module 260, processor 310, and memory 320. Detailed Implementation

[0054] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0055] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0056] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," "exceeding," etc. are understood to exclude the stated number, and "above," "below," "within," etc. are understood to include the stated number. If "first," "second," etc. are used in the description, they are only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance or implicitly indicating the number of indicated technical features or the order of the indicated technical features.

[0057] In the description of this invention, unless otherwise explicitly defined, terms such as "set", "install", and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0058] Tomographic scanning is an imaging technique that generates two-dimensional or three-dimensional views by acquiring cross-sectional images of the interior of an object. The main types include: computed tomography (CT), magnetic resonance imaging (MRI), optical coherence tomography (OCT), positron emission tomography (PET), and thermal tomography (TTM). Computed tomography uses an X-ray beam to perform a tomographic scan of the target object and then uses a computer to process the data to generate detailed images of the body's internal structures. Cone-beam computed tomography (CBCT), with its significant advantages such as high spatial resolution, low radiation dose, rapid scanning, and three-dimensional imaging, has been widely used in human clinical practice, veterinary medicine, and pathological examination.

[0059] Please refer to Figure 1 This embodiment discloses a tomographic scanning device, including a first robotic arm 110, a second robotic arm 120, an X-ray source 130, a detector 140, and a rotating stage 150. The X-ray source 130 is installed at the execution end of the first robotic arm 110, and the detector 140 is installed at the execution end of the second robotic arm 120. The X-ray source 130 and the detector 140 are matched with each other, and a scanning channel is formed between them. The rotating stage 150 is used to place the target object to be scanned so that the target object is located within the scanning channel. Both the first robotic arm 110 and the second robotic arm 120 can be six-degree-of-freedom industrial robotic arms. The first robotic arm 110 and the second robotic arm 120 form a dual-robotic arm scanning control structure, thereby driving the X-ray source 130 and the detector 140 to adjust their positions respectively. The rotating stage 150 can drive the target object to rotate, thereby performing a 360° full-circumferential scan of the target object.

[0060] Please refer to Figure 2 This embodiment discloses a tomographic scanning control method applied to a tomographic scanning device with dual robotic arms, including steps S100 to S600. It should be noted that the numbering of the steps in this embodiment is only for ease of review and understanding, and not to limit the execution order of the steps. The details of each step are described below:

[0061] S100: Acquire initial projection image data and initial scanning parameters of the tomography device;

[0062] For example, projection image data at a preset initial scanning angle is acquired for key anatomical structure identification. The preset initial scanning angle can be a custom angle such as 5°, 10°, or 12°. In a specific application example, the initial scanning angle is 5°, and 120 initial projection images are provided to obtain the initial projection image data. During this step, the initial scanning parameters of the computed tomography (CT) device are simultaneously acquired, such as the tube current of the X-ray source 130 and the rotation speed of the rotating stage 150. The tube current of the X-ray source 130 is used to control the X-ray radiation dose, and the rotation speed of the rotating stage 150 is used to control the rate of change of the scanning angle of the target object, affecting imaging quality and efficiency.

[0063] S200: Based on a pre-trained trajectory planning model, key anatomical structures are identified from the initial projected image data, a probability distribution map representing the importance of each anatomical region is generated, and a sparse scanning angle sequence and the first scanning parameter corresponding to each scanning angle are determined based on the probability distribution map and the initial scanning parameters.

[0064] For example, the quality of 3D reconstruction of tomographic images is directly affected by the choice of sampling angle and radiation dose. Theoretically, increasing the sampling angle density and increasing the radiation dose can help increase the number and quality of projected images, thereby enhancing the overall quality of the 3D reconstructed image. However, in the modern tomographic technology trend of pursuing efficient and low-dose scanning, simply relying on increasing sampling density and dose is not an ideal solution.

[0065] Therefore, this embodiment uses a pre-trained trajectory planning model to determine sparse angle scanning sequences and corresponding first scanning parameters for each scanning angle, achieving high-quality, low-dose imaging with limited projection data. The trajectory planning model is a pre-trained model based on a large amount of sample data, such as a 3D U-Net++ network combined with a non-local attention module. It has the ability to identify key anatomical structures and can perform key anatomical structure identification on initial projection image data. For example, the input initial projection image data is preprocessed by the 3D U-Net++ network to form a 3D projection volume. After multi-scale feature extraction and long-distance association enhancement using non-local attention, a probability distribution map representing the importance of each anatomical region is generated in the form of a three-dimensional voxel grid. In this distribution map, each voxel is assigned a value between 0 and 1, representing the probability that the three-dimensional voxel spatial location belongs to a key anatomical structure.

[0066] By analyzing the information in the probability distribution map, the importance of different regions can be quantified, thus guiding subsequent sampling strategies: denser sampling and a moderately increased radiation dose are applied to high-probability critical areas; while sparser sampling and a correspondingly reduced radiation dose are used for low-probability non-critical areas. This not only ensures the imaging quality of critical parts but also effectively reduces the overall radiation dose.

[0067] Based on the aforementioned probability distribution map and the initially set scanning parameters, a sparse scanning angle sequence and corresponding first scanning parameters for each scanning angle are determined. This aims to balance image quality and radiation dose, providing an optimized execution path for subsequent tomographic scanning. The first scanning parameters are optimized based on the initial scanning parameters. For example, for critical regions, the first scanning parameters can be adjusted by appropriately increasing the tube current (i.e., increasing the radiation dose) and reducing the rotation speed of the target object (e.g., by reducing the rotation speed of the rotating stage 150 to achieve dense sampling).

[0068] S300: Based on the sparse scanning angle sequence and the first scanning parameters, control the tomography scanning device to perform tomography scanning operation and obtain the first projection image data;

[0069] For example, the control system of the dual robotic arms constructs a dynamic model based on Denavit-Hartenberg (DH) parameters and uses an adaptive particle swarm optimization algorithm for trajectory planning to ensure that the relative pose error between the X-ray source 130 and the detector 140 is controlled within a preset range under complex scanning paths. The control system can automatically plan the motion trajectory of the dual robotic arms according to the sparse scanning angle sequence, which is prior art and will not be described in detail in this embodiment. During the actual scanning process, according to the sparse scanning angle sequence and the corresponding first scanning parameters, the dual robotic arms, X-ray source 130, detector 140, and rotating stage 150 are coordinated and controlled to operate synchronously to perform tomographic scanning operations, thereby acquiring first projection image data; wherein, the first projection image data includes multiple frames of projection images corresponding to each scanning angle.

[0070] S400: Perform three-dimensional image reconstruction based on the first projected image data to generate a sparse reconstructed image;

[0071] For example, given that the first projected image data consists of a series of independent two-dimensional projected images, it is difficult to directly provide an intuitive view of the internal structure of the target object. Therefore, a three-dimensional image reconstruction technique is used to process these two-dimensional projected images to generate a sparse reconstructed image. This process first uses a set of two-dimensional projected images at various scanning angles as input, and then uses a reconstruction algorithm (such as iterative reconstruction, filtered back projection, or deep learning methods) to accurately calculate the three-dimensional representation of the sample's internal structure. The final output sparse reconstructed image not only clearly displays the details of the sample's internal structure but also helps in further analysis and understanding of its features.

[0072] S500: Using a pre-trained image quality assessment model, multimodal feature fusion is performed on sparse reconstructed images and the current scanning parameters of the tomographic scanning device, and the image quality score and corresponding fusion feature vector are output based on the fusion result.

[0073] For example, to ensure the clarity of the 3D reconstructed image, this embodiment uses a pre-trained image quality assessment model to perform multimodal feature fusion between the sparse reconstructed image and the current scanning parameters of the tomographic scanning device, and outputs an image quality score and a corresponding fusion feature vector based on the fusion result. The image quality score is a value between 0 and 1, used to quantify the current imaging quality and achieve a precise mapping between scanning parameters and image quality; the fusion feature vector characterizes the key factors affecting imaging quality and serves as one of the important bases for subsequent adaptive optimization of scanning parameters.

[0074] S600: Based on the probability distribution map, fused feature vector, image quality score, current device status data and historical action data of the tomographic scanning device, the scanning action command is adaptively optimized to obtain the optimized second scanning parameters.

[0075] For example, based on probability distribution maps, fused feature vectors, image quality scores, current device status data, and historical action data of the tomographic scanning equipment, a comprehensive analysis is performed from multiple dimensions to adaptively optimize the scanning action commands, generating optimized second scanning parameters. Based on these second scanning parameters, the tomographic scanning equipment is controlled to perform the scanning operation, acquire second projection data, and use this second projection data for 3D image reconstruction. This closed-loop feedback mechanism continuously optimizes the parameters throughout the entire tomographic scanning and image reconstruction process, ensuring that each scan is adjusted based on the latest image quality and device status information, thereby effectively improving imaging resolution.

[0076] Thus, based on a pre-trained trajectory planning model, key anatomical structures are identified in the initial projected image, generating a probability distribution map representing the importance of each region. Based on this, a sparse scanning angle sequence and corresponding first scanning parameters are determined. Dense sampling and appropriately increased radiation dose are used in high-weight regions, while sparse sampling and reduced dose are used in low-weight regions. The tomographic scanning device is controlled to perform the scan, obtaining the first projection data and reconstructing it into a sparse reconstructed image. Then, an image quality assessment model performs multimodal fusion and scoring of the sparse reconstructed image and the current scanning parameters. Based on the probability distribution map, fused feature vector, image quality score, device status, and historical action data, the scanning command is adaptively optimized to generate second scanning parameters. This makes the scanning trajectory more continuous and smooth, achieving regional adaptive control of angle, dose, and path, which is beneficial for improving imaging resolution.

[0077] In step S200, based on the probability distribution diagram and the initial scanning parameters, the sparse scanning angle sequence and the first scanning parameters corresponding to each scanning angle are determined, including:

[0078] Forward projection simulation is performed based on the probability distribution map to calculate the sum of voxel probabilities along the ray path at each candidate scanning angle;

[0079] The sum of voxel probabilities for each candidate scanning angle is normalized to obtain an angle weight sequence. Each weight in the angle weight sequence is used to characterize the contribution of the corresponding projection angle to the reconstruction quality of key anatomical structures.

[0080] Based on the Poisson disk sampling algorithm, using the angle weight sequence as the local density control factor, dense sampling points are generated in the high-weight region and sparse sampling points are generated in the low-weight region to obtain a sparse scanning angle sequence.

[0081] The initial scanning parameters are dynamically modulated based on the angle weight sequence to generate the first scanning parameters corresponding to each angle in the sparse scanning angle sequence.

[0082] For example, each voxel value in the probability distribution map represents the probability that the spatial location belongs to a key anatomical structure. The X-ray forward projection process is simulated based on the probability distribution map. Specifically, for multiple candidate scanning angles (e.g., covering 0° to 360° with a step size of 1°), the voxel probability values ​​are summed along all ray paths at each angle to obtain the total voxel probability for each candidate angle. Subsequently, the total voxel probability of all candidate angles is normalized (e.g., using min-max normalization or Softmax normalization) to generate an angle weight sequence. Each weight value in the angle weight sequence belongs to the interval [0, 1] and is used to quantify the contribution of the corresponding projection angle to the reconstruction quality of the key anatomical structure. The higher the weight, the richer the key structural information contained in the corresponding angle, and the greater the impact on the final image quality. Next, the Poisson disk sampling algorithm is used, with the angle weight sequence as the local sampling density control factor, to generate non-uniformly distributed sampling points in the angle domain: in high-weight regions, a smaller minimum sampling interval is allowed, thus forming dense sampling; in low-weight regions, a larger minimum interval is used to achieve sparse sampling. The resulting sparse scanning angle sequence retains the information integrity of key viewpoints while significantly reducing the total number of projections. Furthermore, the initial scanning parameters are dynamically modulated according to the angle weight sequence to generate first scanning parameters that correspond one-to-one with each angle in the sparse scanning angle sequence.

[0083] The initial scan parameters include tube current and scan rotation speed. These initial scan parameters are dynamically modulated according to an angle weight sequence to generate first scan parameters corresponding to each angle in the sparse scan angle sequence, including:

[0084] Based on the angle weight sequence, the tube current and scanning rotation speed in the initial scanning parameters are dynamically modulated to generate the first scanning parameters corresponding to each angle in the sparse scanning angle sequence.

[0085] For example, the initial scanning parameters include the tube current of the X-ray source 130 and the rotation speed of the rotating stage 150. Corresponding to the high-weight angle range (i.e., the viewpoint dominated by critical anatomical regions), the tube current of the X-ray source 130 is increased to approximately 150 mA, and the rotation speed of the rotating stage 150 is reduced to approximately 10° / s to improve the signal-to-noise ratio and enhance the clarity of structural details. Corresponding to the low-weight angle range (i.e., non-critical regions), the tube current is reduced to approximately 20 mA, and the rotation speed is increased to approximately 30° / s, thereby effectively reducing the radiation dose received by the target object and shortening the overall scanning time. Through the aforementioned adaptive angle selection and parameter modulation mechanism, synergistic optimization of imaging quality and scanning efficiency can be achieved with limited projection data, ensuring high-fidelity reconstruction of critical anatomical structures while minimizing unnecessary radiation exposure and scanning time.

[0086] Step S400: Perform 3D image reconstruction based on the first projected image data to generate a sparse reconstructed image, including:

[0087] Using a physical reconstruction module based on iterative coordinate descent, the first projected image data is iteratively optimized. The reconstructed image is updated on a voxel-by-voxel basis to minimize the projection domain error. Combined with projection data consistency constraints and total variation regularization terms, a preliminary reconstructed image that satisfies physical measurement constraints is generated.

[0088] The multi-scale feature enhancement module based on 3D Swin Transformer is used to perform multi-scale feature enhancement processing on the initially reconstructed image. In this process, the contextual information from local texture to global long-range anatomical structure is extracted layer by layer through self-attention calculation within the local window and the cross-window shifting window mechanism, thereby improving the structural details and perceptual quality of the image.

[0089] At each feature scale, the scanning parameters of the tomography device are encoded into conditional vectors by an adaptive feature gating unit based on a geometric perception attention mechanism. The feature fusion weights of the physical reconstruction module and the multi-scale feature enhancement module are dynamically modulated according to the conditional vectors to fuse the features output by the physical reconstruction module and the multi-scale feature enhancement module.

[0090] For example, the 3D image reconstruction in this embodiment employs a dual-path processing mechanism. The first processing path is a physical reconstruction module, and the second processing path is a multi-scale feature enhancement module. In the first processing path, a physical reconstruction module based on Iterative Coordinate Descent (ICD) is constructed. This module updates the image pixel-by-pixel or voxel-by-voxel, adjusting the value of only one pixel at a time to minimize projection domain error. It maintains the well-posedness of the solution through projection domain data consistency constraints and TV regularization, iterating through all pixels multiple times until convergence, thereby generating a stable preliminary reconstruction result that conforms to the measurement data. In the second processing path, a cascaded multi-scale feature enhancement module based on the 3D Swin Transformer architecture is designed. This module divides the volume data into local windows for self-attention computation and implements a shift-window attention mechanism at the next layer to achieve cross-window information interaction through window offset. This cascading layer-by-layer approach captures multi-scale features from local texture to global long-range anatomical structures, thereby enhancing the structural details and perceptual quality of the reconstructed image.

[0091] The two paths interact across multiple scales via adaptive feature gating units. A geometrically aware attention mechanism (GAA) encodes scanning parameters (such as the distance and angle between X-ray source 130 and detector 140) into conditional vectors, modulating the attention distribution during feature fusion. This serves as prior conditions guiding the network to focus on physically relevant features. Specifically, the GAA dynamically adjusts the weight allocation of the output features of the first and second processing paths based on the scanning parameters, enabling more comprehensive feature representation of key structures under different viewpoints or projection conditions. This achieves the synergistic fusion of physical constraints and data-driven information.

[0092] The network training employs a three-stage strategy. First, the physical reconstruction module is pre-trained on simulation data, enabling it to learn preliminary reconstruction capabilities based on projection domain data consistency and regularization constraints. The simulation dataset primarily includes projection data and corresponding reconstructed images generated by the physical imaging model, covering different noise levels, scanning parameters, and variations in various anatomical structures to ensure the physical reconstruction module can master reconstruction methods consistent with the imaging system's rules. Next, end-to-end training is performed on clinical data to adjust the second processing path and dual-path fusion, allowing the network to adapt to real-world scanning noise and complex anatomical structures. The training data used in this stage comes from projection data and reconstructed images of real patients, ensuring the network can recover detailed and structurally accurate images under actual clinical conditions. Finally, adversarial training further optimizes perceptual quality. The adversarial training dataset includes real clinical images as real samples for the discriminator and reconstructed images generated by the network as samples for the generator. Through adversarial optimization between the generator and the discriminator, the network improves the visual perceptual quality of the images while maintaining physical consistency.

[0093] The loss function integrates projection domain L1 loss to ensure physical consistency, image domain multi-scale SSIM loss to preserve structural features, and task-specific anatomical structure preservation loss. The weights of key regions such as blood vessels and nerves are dynamically adjusted through learnable parameters, enabling the network to focus on optimizing the recovery of these important structures.

[0094] Step S500: Using a pre-trained image quality assessment model, multimodal feature fusion is performed on the sparse reconstructed image and the current scanning parameters of the tomographic scanning device. Based on the fusion result, an image quality score and the corresponding fused feature vector are output, including:

[0095] The 3D convolutional module of the pre-trained image quality assessment model is used to extract features from sparse reconstructed images to generate first image features that represent inter-frame dynamic structure information.

[0096] The first image features are input into a residual network with an embedded channel attention module for feature enhancement to obtain the second image features;

[0097] The current scanning parameters of the tomography device are nonlinearly mapped by a multilayer perceptron to generate scanning parameter features that match the second image feature dimension.

[0098] Based on the cross-attention mechanism, attention weights are determined according to the second image features and scanning parameter features, and the second image features are weighted and modulated to obtain preliminary fused features;

[0099] The initial fused features are input into the pyramid pooling module to generate and output the fused feature vector;

[0100] An image quality score is obtained by activating the fused feature vector using an activation function.

[0101] For example, the image quality assessment model is implemented based on the ResNet50 network architecture. It employs a dual-branch input structure to process the sparsely reconstructed image and the current scanning parameters of the tomographic scanning device, respectively, to output an image quality score belonging to the interval [0, 1] and the corresponding fused feature vector. Specifically, the sparsely reconstructed image is processed by the 3D convolutional module of the image quality assessment model to extract spatiotemporal related features, thereby modeling inter-frame dynamic structural information and generating the first image feature. Subsequently, feature enhancement is performed through a residual network (such as a ResNet50 network) embedded with a channel attention module. The channel attention mechanism enhances the responsiveness to key features through adaptive channel weight allocation, increasing the network's attention to quality-sensitive regions (such as edges, noise, and texture), resulting in the second image feature. The scanning parameter branch uses a multilayer perceptron (MLP) to nonlinearly map the current scanning parameters of the tomographic scanning device to a high-dimensional feature representation consistent with the image feature space, thus obtaining the scanning parameter features. Then, based on the cross-attention mechanism, the two branches perform cross-attention feature fusion, using the second image feature as the main query and the scanning parameter feature as the key / value pair. The influence of scanning parameters on image quality features is adaptively modeled through attention weights, achieving a deep correlation mapping between scanning conditions and imaging quality characteristics, resulting in preliminary fused features. Next, the preliminary fused features are integrated into a multi-scale context through a pyramid pooling module (PPM) to capture global information under different receptive fields, enabling the model to simultaneously possess local detail sensitivity and global structure awareness, generating and outputting a fused feature vector. The output layer uses a sigmoid activation function to output an image quality score within the [0, 1] interval, achieving accurate quantitative evaluation of the reconstructed image quality under different scanning conditions. The overall model is jointly optimized through transfer learning and a hybrid loss function, taking into account spatial resolution, contrast, and visual structure consistency.

[0102] Step S600: Based on the probability distribution map, fused feature vector, image quality score, current device status data of the tomographic scanning device, and historical action data, adaptively optimize the scanning action command to obtain the optimized second scanning parameters, including:

[0103] A fusion state vector is generated based on the probability distribution map, fusion feature vector, image quality score, current device status data and historical action data of the tomographic scanning device;

[0104] Based on the SAC algorithm, scanning action commands are generated according to the fused state vector, and the scanning action commands are adaptively optimized by maximizing the multi-objective reward function to obtain the optimized second scanning parameters.

[0105] For example, multi-source information such as probability distribution maps, fused feature vectors, image quality scores, current device status data of the computed tomography (CT) scanner, and historical action data are normalized and stitched together to construct a unified fused state vector (e.g., 28-dimensional). The probability distribution map reflects the importance of each anatomical region in the target sample; the fused feature vector and image quality score characterize the structural integrity and imaging quality of the current sparse reconstructed image; the current device status data includes operating parameters such as the real-time pose of the dual robotic arms, the tube voltage or current of the X-ray source 130, the gain of the detector 140, and the angle of the rotating stage 150; and the historical action data records the scanning angles, parameter adjustment trajectories, and feedback results executed in previous scanning cycles. The fused state vector provides multi-dimensional perception of the current scanning scene, which helps enhance the accuracy and robustness of subsequent decisions.

[0106] Subsequently, the Soft Actor-Critic (SAC) reinforcement learning algorithm was employed, using the fused state vector as input to generate the scanning action commands for the next stage. The SAC algorithm is an off-policy deep reinforcement learning method based on a maximum entropy framework, specifically designed for continuous action spaces. It combines high sample efficiency, strong exploration capabilities, and training stability, making it particularly suitable for real-time control optimization of complex electromechanical systems such as tomographic scanning equipment with dual robotic arms. To further guide the strategy towards clinical needs, a multi-objective reward function was defined, and the scanning action commands were adaptively optimized by maximizing this function.

[0107] Thus, the generated second scanning parameters (including tube current, rotation speed, exposure time, etc. corresponding to each scanning angle) can achieve regional adaptive adjustment: in high-weight critical anatomical regions, the sampling density is automatically increased, the tube current is appropriately increased (e.g., by about 20 mA), and the rotation speed of the rotating stage 150 is reduced (e.g., dynamically adjusted to within ±15° / s) to enhance the signal-to-noise ratio and structural clarity; while in low-weight non-critical regions, a sparse sampling strategy is adopted, reducing the tube current and increasing the rotation speed, effectively reducing the radiation dose received by the subject and shortening the scanning time. Ultimately, the optimized scanning parameters not only make the motion trajectory of the dual robotic arms more continuous and smooth, but also intelligently allocate imaging resources according to the importance of anatomical structures, thereby achieving synergistic optimization of imaging quality, radiation safety, and equipment performance under limited projection conditions.

[0108] The above steps involve adaptively optimizing the scanning action command by maximizing the multi-objective reward function to obtain the optimized second scanning parameters, including:

[0109] The multi-objective reward function is determined by weighted summation based on the current image quality score, the reciprocal of the scan dose, and the continuous motion stability index.

[0110] The scanning action command is adaptively optimized by maximizing the multi-objective reward function to obtain the optimized second scanning parameters.

[0111] For example, in the reinforcement learning optimization algorithm of the present invention, a multi-objective reward function is defined to adaptively adjust the scanning parameters. The expression of the multi-objective reward function is: R = α·Q-score + β·(1 / Dose) + γ·Temporal_smoothness. Here, R represents the total reward value of the current decision step, and the reinforcement learning strategy maximizes the accumulated expected reward by selecting scanning actions; Q-score is the current image quality score, ranging from 0 to 1, output by the image quality assessment model mentioned above, used to quantify the structural fidelity and perceptual quality of the reconstructed image; Dose represents the estimated radiation dose corresponding to the current scanning operation, using the reciprocal 1 / Dose as the reward term to encourage minimizing radiation exposure while meeting imaging requirements; Temporal_smoothness characterizes the smoothness of changes between consecutive scanning actions, such as the difference in angular velocity or acceleration of the robotic arm between adjacent angles, used to suppress abrupt motion changes and ensure the stability and safety of the dual robotic arm operation; α, β, and γ are non-negative weight coefficients used to adjust the priority among image quality, dose control, and motion smoothness. In actual scanning, the weighting coefficients can be dynamically adjusted according to the scanning stage. Through the multi-objective reward function, the reinforcement learning strategy can effectively reduce the overall radiation dose and maintain the continuous smoothness of the scanning trajectory while ensuring the imaging quality of key anatomical structures.

[0112] Please refer to Figure 3 This embodiment provides a tomographic scanning control device, applied to a tomographic scanning equipment with dual robotic arms, comprising:

[0113] Data acquisition module 210 is used to acquire initial projection image data and initial scanning parameters of the tomographic scanning device;

[0114] The trajectory planning module 220 is used to identify key anatomical structures in the initial projected image data based on a pre-trained trajectory planning model, generate a probability distribution map representing the importance of each anatomical region, and determine the sparse scanning angle sequence and the first scanning parameter corresponding to each scanning angle based on the probability distribution map and the initial scanning parameters.

[0115] The scanning control module 230 is used to control the tomographic scanning device to perform tomographic scanning operation according to the sparse scanning angle sequence and the first scanning parameters, so as to obtain the first projection image data.

[0116] Image reconstruction module 240 is used to perform three-dimensional image reconstruction based on the first projected image data to generate a sparse reconstructed image;

[0117] The quality assessment module 250 is used to perform multimodal feature fusion on sparse reconstructed images and the current scanning parameters of the tomographic scanning device using a pre-trained image quality assessment model, and outputs an image quality score and the corresponding fusion feature vector based on the fusion result.

[0118] The parameter optimization module 260 is used to adaptively optimize the scanning action command based on the probability distribution map, fused feature vector, image quality score, current device status data and historical action data of the tomographic scanning device, to obtain the optimized second scanning parameters.

[0119] The inventive concept of this tomographic scanning control device embodiment is the same as that of the tomographic scanning control method embodiment described above. Content not covered in this tomographic scanning control device embodiment can be referred to in the tomographic scanning control method embodiment described above, and will not be repeated here. Based on a pre-trained trajectory planning model, key anatomical structures are identified in the initial projection image, generating a probability distribution map representing the importance of each region. Based on this, a sparse scanning angle sequence and corresponding first scanning parameters are determined. Dense sampling and appropriately increased radiation dose are used in high-weight regions, while sparse sampling and reduced dose are used in low-weight regions. The tomographic scanning device is controlled to perform scanning, obtaining the first projection data and reconstructing it into a sparse reconstructed image. Then, a multimodal fusion and scoring of the sparse reconstructed image and current scanning parameters are performed using an image quality assessment model. Based on the probability distribution map, fused feature vector, image quality score, device status, and historical action data, the scanning command is adaptively optimized to generate second scanning parameters. This makes the scanning trajectory more continuous and smooth, achieving regional adaptive control of angle, dose, and path, which is beneficial for improving imaging resolution.

[0120] Please refer to Figure 4 This embodiment provides a tomographic scanning device, including a processor 310 and a memory 320. The memory 320 stores a computer program, and the processor 310 executes the computer program to implement the tomographic scanning control method described above. The tomographic scanning control method is detailed above and will not be repeated here. Based on a pre-trained trajectory planning model, key anatomical structures are identified in the initial projection image, generating a probability distribution map representing the importance of each region. Based on this, a sparse scanning angle sequence and corresponding first scanning parameters are determined. Dense sampling and appropriately increased radiation dose are used in high-weight regions, while sparse sampling and reduced dose are used in low-weight regions. The tomographic scanning device is controlled to perform scanning, obtaining the first projection data and reconstructing it into a sparse reconstructed image. Then, a multimodal fusion and scoring of the sparse reconstructed image and the current scanning parameters are performed using an image quality assessment model. Based on the probability distribution map, fused feature vector, image quality score, device status, and historical action data, the scanning instructions are adaptively optimized to generate second scanning parameters. This makes the scanning trajectory more continuous and smooth, achieving regional adaptive control of angle, dose, and path, which is beneficial for improving imaging resolution.

[0121] This embodiment provides a storage medium storing a computer program. When the computer program is run, it implements the aforementioned tomographic scanning control method. The tomographic scanning control method is detailed above and will not be repeated here. Based on a pre-trained trajectory planning model, key anatomical structures are identified in the initial projection image, generating a probability distribution map representing the importance of each region. Based on this, a sparse scanning angle sequence and corresponding first scanning parameters are determined. Dense sampling and appropriately increased radiation dose are used in high-weight regions, while sparse sampling and reduced dose are used in low-weight regions. The tomographic scanning device is controlled to perform scanning, obtaining the first projection data and reconstructing it into a sparse reconstructed image. Then, a multimodal fusion and scoring of the sparse reconstructed image and current scanning parameters are performed using an image quality assessment model. Based on the probability distribution map, fused feature vector, image quality score, device status, and historical action data, the scanning instructions are adaptively optimized to generate second scanning parameters. This makes the scanning trajectory more continuous and smooth, achieving regional adaptive control of angle, dose, and path, which is beneficial for improving imaging resolution.

[0122] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A tomographic scanning control method, applied to a tomographic scanning device with dual robotic arms, characterized in that, include: Acquire initial projection image data and initial scanning parameters of the computed tomography device; Based on a pre-trained trajectory planning model, key anatomical structures are identified in the initial projected image data, a probability distribution map representing the importance of each anatomical region is generated, and a sparse scanning angle sequence and the first scanning parameter corresponding to each scanning angle are determined according to the probability distribution map and the initial scanning parameters. Based on the sparse scanning angle sequence and the first scanning parameters, the tomography device is controlled to perform a tomography operation to obtain first projection image data. Three-dimensional image reconstruction is performed based on the first projected image data to generate a sparse reconstructed image; Using a pre-trained image quality assessment model, multimodal feature fusion is performed on the sparse reconstructed image and the current scanning parameters of the tomographic scanning device, and the image quality score and corresponding fusion feature vector are output based on the fusion result. Based on the probability distribution map, the fusion feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device, the scanning action command is adaptively optimized to obtain the optimized second scanning parameters; The step of adaptively optimizing the scanning action command based on the probability distribution map, the fused feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device to obtain the optimized second scanning parameters includes: A fusion state vector is generated based on the probability distribution map, the fusion feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device; Based on the SAC algorithm, a scanning action command is generated according to the fusion state vector, and the scanning action command is adaptively optimized by maximizing the multi-objective reward function to obtain the optimized second scanning parameters.

2. The tomographic scanning control method according to claim 1, characterized in that, The step of determining the sparse scanning angle sequence and the first scanning parameter corresponding to each scanning angle based on the probability distribution map and the initial scanning parameters includes: Based on the probability distribution map, a forward projection simulation is performed to calculate the sum of voxel probabilities along the ray path at each candidate scanning angle; The sum of voxel probabilities for each candidate scanning angle is normalized to obtain an angle weight sequence. Each weight in the angle weight sequence is used to characterize the contribution of the corresponding projection angle to the reconstruction quality of key anatomical structures. Based on the Poisson disk sampling algorithm, the angle weight sequence is used as the local density control factor to generate dense sampling points in high-weight regions and sparse sampling points in low-weight regions, thus obtaining a sparse scanning angle sequence. The initial scanning parameters are dynamically modulated according to the angle weight sequence to generate first scanning parameters corresponding to each angle in the sparse scanning angle sequence.

3. The tomographic scanning control method according to claim 2, characterized in that, The initial scanning parameters include tube current and scanning rotation speed. The step of dynamically modulating the initial scanning parameters according to the angle weight sequence to generate first scanning parameters corresponding to each angle in the sparse scanning angle sequence includes: Based on the angle weight sequence, the tube current and the scanning rotation speed in the initial scanning parameters are dynamically modulated to generate first scanning parameters corresponding to each angle in the sparse scanning angle sequence.

4. The tomographic scanning control method according to claim 1, characterized in that, The step of performing three-dimensional image reconstruction based on the first projected image data to generate a sparse reconstructed image includes: Using a physical reconstruction module based on iterative coordinate descent, the first projected image data is iteratively optimized. The reconstructed image is updated on a voxel-by-voxel basis to minimize the projection domain error. Combined with projection data consistency constraints and total variation regularization terms, a preliminary reconstructed image that satisfies physical measurement constraints is generated. The multi-scale feature enhancement module based on 3D Swin Transformer is used to perform multi-scale feature enhancement processing on the preliminary reconstructed image. In this process, the contextual information from local texture to global long-range anatomical structure is extracted layer by layer through self-attention calculation within the local window and the cross-window shifting window mechanism, thereby improving the structural details and perceptual quality of the image. At each feature scale, the scanning parameters of the tomography device are encoded into conditional vectors by an adaptive feature gating unit based on a geometric perception attention mechanism. The feature fusion weights of the physical reconstruction module and the multi-scale feature enhancement module are dynamically modulated according to the conditional vectors to fuse the features output by the physical reconstruction module and the multi-scale feature enhancement module.

5. The tomographic scanning control method according to claim 1, characterized in that, The process of using a pre-trained image quality assessment model to perform multimodal feature fusion on the sparse reconstructed image and the current scanning parameters of the tomographic scanning device, and outputting an image quality score and corresponding fused feature vector based on the fusion result, includes: The 3D convolutional module of the pre-trained image quality assessment model is used to extract features from the sparse reconstructed image to generate a first image feature representing the dynamic structure information between frames. The first image features are input into a residual network with an embedded channel attention module for feature enhancement to obtain the second image features; The current scanning parameters of the tomography device are nonlinearly mapped by a multilayer perceptron to generate scanning parameter features that match the second image feature dimension. Based on the cross-attention mechanism, attention weights are determined according to the second image features and the scanning parameter features, and the second image features are weighted and modulated to obtain preliminary fused features; The preliminary fusion features are input into the pyramid pooling module to generate and output the fusion feature vector; An image quality score is obtained by activating the fused feature vector using an activation function.

6. The tomographic scanning control method according to claim 1, characterized in that, The step of adaptively optimizing the scanning action command by maximizing the multi-objective reward function to obtain the optimized second scanning parameters includes: The multi-objective reward function is determined by weighted summation based on the current image quality score, the reciprocal of the scan dose, and the continuous motion stability index. The scanning action command is adaptively optimized by maximizing the multi-objective reward function to obtain the optimized second scanning parameters.

7. A tomographic scanning control device, applied to a tomographic scanning equipment with dual robotic arms, characterized in that, include: The data acquisition module is used to acquire initial projection image data and initial scanning parameters of the computed tomography device; The trajectory planning module is used to identify key anatomical structures in the initial projected image data based on a pre-trained trajectory planning model, generate a probability distribution map representing the importance of each anatomical region, and determine a sparse scanning angle sequence and a first scanning parameter corresponding to each scanning angle based on the probability distribution map and the initial scanning parameters. The scanning control module is used to control the tomography scanning device to perform tomography scanning operation according to the sparse scanning angle sequence and the first scanning parameters, so as to obtain the first projection image data; The image reconstruction module is used to perform three-dimensional image reconstruction based on the first projected image data to generate a sparse reconstructed image; The quality assessment module is used to perform multimodal feature fusion on the sparse reconstructed image and the current scanning parameters of the tomographic scanning device using a pre-trained image quality assessment model, and output an image quality score and corresponding fusion feature vector based on the fusion result; The parameter optimization module is used to adaptively optimize the scanning action command based on the probability distribution map, the fusion feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device, to obtain the optimized second scanning parameters. The step of adaptively optimizing the scanning action command based on the probability distribution map, the fused feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device to obtain the optimized second scanning parameters includes: A fusion state vector is generated based on the probability distribution map, the fusion feature vector, the image quality score, the current device status data and historical action data of the tomographic scanning device; Based on the SAC algorithm, a scanning action command is generated according to the fusion state vector, and the scanning action command is adaptively optimized by maximizing the multi-objective reward function to obtain the optimized second scanning parameters.

8. A computed tomography (CT) scanner, comprising a processor and a memory, wherein the memory stores a computer program, characterized in that, When the processor runs the computer program, it is used to implement the computed tomography control method as described in any one of claims 1 to 6.

9. A storage medium storing a computer program, characterized in that, When the computer program is run, it implements the tomographic scanning control method as described in any one of claims 1 to 6.

Images