A CT image real-time processing method and system based on edge computing

Abstract

The application provides a CT image real-time processing method and system based on edge computing, and relates to the technical field of image processing.The method comprises the following steps: step 1, obtaining original projection data stream output by a CT scanning device in real time, performing data buffering and time sequence alignment processing on an edge computing node, and obtaining a synchronous original projection data frame sequence; step 2, receiving the synchronous original projection data frame sequence, eliminating signal distortion caused by differences in tissue density and X-ray hardening, and obtaining a corrected projection data frame sequence; and step 3, obtaining an initial CT tomographic image sequence by using the corrected projection data frame sequence, combining real-time available computing power of the edge node, and adaptively configuring reconstruction iteration parameters according to the memory state. The application realizes low-delay and high-precision real-time processing of CT images.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a real-time CT image processing method and system based on edge computing. Background Technology

[0002] In current clinical applications of intraoperative CT-guided interventional surgery, existing CT image processing systems based on traditional computing architectures have revealed significant technical limitations. For example, bedside CT image processing systems, if employing a centralized cloud computing architecture, require the complete upload of the raw projection data stream output by the CT scanning device to a remote cloud computing center for full-process correction, reconstruction, and enhancement calculations. Due to network transmission latency and cloud task scheduling queuing mechanisms, end-to-end image processing is delayed, making it difficult to provide synchronous visual feedback for surgical operations. During dynamic adjustment of the puncture needle, the images are prone to being out of sync with the actual operation, posing a risk of surgical deviation. If a device-local embedded computing architecture is used, the computing power and memory resources of the embedded processor are limited, allowing only simplified filtered back projection to be deployed. Reconstruction algorithms and global image enhancement algorithms struggle to run computationally intensive iterative reconstruction algorithms and precise pathological feature extraction algorithms. This results in the difficulty of clearly presenting small lesions and blurred tissue boundaries in images, affecting the formulation and execution of surgical plans. It exposes the fundamental defects in the computing architecture design of existing CT image processing technology, namely, the failure to build a distributed processing pipeline adapted to the resource characteristics of edge computing nodes. It is difficult to achieve low latency and high accuracy in image processing under the constraint of limited computing resources at the edge. Furthermore, it lacks the ability to dynamically adjust computing strategies based on the real-time resource status of edge nodes. When there are network fluctuations or changes in device load, image stuttering or sudden quality drops are likely to occur, making it difficult to meet the stringent requirements of real-time computing and image quality, such as intraoperative CT guidance. Summary of the Invention

[0003] This invention provides a real-time CT image processing method and system based on edge computing, which enables low-latency, high-precision real-time processing of CT images.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] In a first aspect, a real-time CT image processing method based on edge computing is provided, the method comprising:

[0006] Step 1: Obtain the raw projection data stream output in real time from the CT scanning device, and perform data buffering and time series alignment processing at the edge computing node to obtain a synchronized raw projection data frame sequence.

[0007] Step 2: Receive the synchronized original projection data frame sequence, eliminate signal distortion caused by tissue density differences and X-ray hardening, and obtain the corrected projection data frame sequence.

[0008] Step 3: Using the corrected projection data frame sequence, combined with the real-time available computing power and memory status of the edge nodes, the reconstruction iteration parameters are adaptively configured to obtain the initial CT tomographic image sequence.

[0009] Step 4: Receive the initial CT tomographic image sequence, construct a voxel-level spatial potential energy distribution field, introduce a spatial field diffusion and local feature affinity calculation mechanism based on image gradient vectors, establish geometric boundary evolution constraints and mechanical equilibrium conditions, and iteratively deduce the curvature change and density potential difference of the morphological contour. Under kinematic trajectory constraints, converge to obtain the precise topological boundary of the pathologically sensitive region and obtain the corresponding non-uniform weight mask. Spatial domain weighted fusion of the initial CT tomographic image sequence and the non-uniform weight mask is performed to obtain a feature-enhanced CT tomographic image sequence.

[0010] Step 5: Perform streaming adaptive coding processing based on structural importance assessment on the feature-enhanced CT tomographic image sequence to obtain a lightweight real-time image data stream.

[0011] Furthermore, the raw projection data stream output in real time from the CT scanning device is acquired, and data buffering and time series alignment processing are performed at the edge computing node to obtain a synchronized raw projection data frame sequence, including:

[0012] Step 11: Acquire the raw projection data stream output in real time by the CT scanning device through the high-speed data communication interface, parse the frame synchronization identifier and timestamp field in the data stream protocol layer, and extract the raw projection data packet with time sequence mark;

[0013] Step 12: Receive the original projection data packets with time sequence marks, perform data buffering processing in the preset buffer area of ​​the edge computing node, and perform queue reordering and frame-breaking linear interpolation on the out-of-order data packets to obtain the projection data blocks to be aligned in a continuous arrangement according to the acquisition time sequence.

[0014] Step 13: Based on the projection data block to be aligned, perform multi-channel time series alignment processing in conjunction with the rotation phase angle of the scanning device to compensate for the timing misalignment caused by network transmission jitter and obtain a synchronized original projection data frame sequence.

[0015] Furthermore, the synchronous original projection data frame sequence is received, and signal distortion caused by tissue density differences and radiation hardening is eliminated to obtain a corrected projection data frame sequence, including:

[0016] Step 21: Based on the metadata header of the synchronized original projection data frame sequence, extract the three-dimensional coordinates of the X-ray source focus, the physical size matrix of the detector array, and the real-time projection angle sequence to construct a dynamic scanning geometric parameter set;

[0017] Step 22: Receive the dynamic scanning geometric parameter set and the synchronized original projection data frame sequence, perform source distance inverse square attenuation correction and air background subtraction operation to compensate for the physical energy loss on the ray propagation path, and obtain the attenuation-compensated projection data frame sequence.

[0018] Step 23: By using the attenuation-compensated projection data frame sequence, a scattered radiation distribution model is constructed based on the projection path integral thickness. Scattered component stripping and multi-spectral hardening artifact suppression operations are performed to eliminate signal distortion caused by tissue density differences and X-ray hardening, and the corrected projection data frame sequence is obtained.

[0019] Furthermore, by using the corrected projection data frame sequence and adaptively configuring the reconstruction iteration parameters based on the real-time available computing power and memory status of the edge nodes, an initial CT tomographic image sequence is obtained, including:

[0020] Step 31: Monitor the processor load rate, cache utilization rate and number of parallel computing threads of the edge computing nodes in real time to obtain a set of real-time available computing power and memory status indicators;

[0021] Step 32: Receive the real-time available computing power and memory status index set and the corrected projection data frame sequence, parse the spatial resolution and slice thickness parameters of the projection data frame sequence, and obtain the initial reconstruction iteration parameters based on the resource index set and data scale mapping;

[0022] Step 33: Based on the initial reconstruction iteration parameters and the corrected projection data frame sequence, perform forward projection simulation and backward error propagation operations, calculate the projection residual field and dynamically update the regularization coefficients to obtain the adaptive reconstruction iteration parameters;

[0023] Step 34: Using the adaptive reconstruction iteration parameters and the corrected projection data frame sequence, perform algebraic iterative reconstruction operation, cyclically mapping between the projection space and the image space until the residual converges, to obtain the initial CT tomographic image sequence.

[0024] Furthermore, the initial CT tomographic image sequence is received, a voxel-level spatial potential energy distribution field is constructed, and a spatial field diffusion and local feature affinity calculation mechanism based on image gradient vectors is introduced. Geometric boundary evolution constraints and mechanical equilibrium conditions are established. By iteratively deducing the curvature change and density potential difference of the morphological contour, the precise topological boundary of the pathologically sensitive region is obtained under kinematic trajectory constraints, resulting in the corresponding non-uniform weight mask, including:

[0025] Step 41: Based on the initial CT tomographic image sequence, extract the voxel density values ​​and three-dimensional spatial coordinate indices of each slice, perform nonlinear potential energy mapping operation, and construct a voxel-level spatial potential energy distribution field;

[0026] Step 42: Calculate the gray-level difference vector in the local neighborhood through the voxel-level spatial potential energy distribution field to obtain the image gradient vector field, perform isotropic diffusion filtering and cross-scale feature similarity measurement to obtain the local feature affinity calculation matrix.

[0027] Step 43: Calculate the matrix by local feature affinity, combine the gradient transition characteristics of the voxel-level spatial potential energy distribution field, construct the geometric boundary evolution constraint rules and local tension balance conditions, and initialize the discrete control point set of the morphological contour.

[0028] Step 44: Based on the discrete control point set, perform iterative evolution operation under the preset kinematic trajectory constraints, calculate the curvature change rate of the shape contour and the density potential difference of adjacent voxels in real time, drive the control points to shrink towards the high gradient potential well region, terminate the iteration when the evolution displacement is lower than the preset convergence threshold, extract the closed contour and perform grid mapping to obtain the corresponding non-uniform weight mask.

[0029] Furthermore, the initial CT tomographic image sequence is spatially weighted and fused with a non-uniform weight mask to obtain a feature-enhanced CT tomographic image sequence, including:

[0030] Step 45: Based on the initial CT tomographic image sequence and the non-uniform weight mask, perform spatial coordinate registration and voxel-level mapping alignment processing to ensure that the weight distribution in the non-uniform weight mask corresponds one-to-one with the anatomical structure of the initial CT tomographic image sequence, and obtain spatially aligned image mask pairs.

[0031] Step 46: Receive spatially aligned image mask pairs, extract the weight coefficients of each voxel position in the non-uniform weight mask, and perform point-by-point multiplication with the corresponding voxel gray values ​​of the initial CT tomographic image sequence to obtain local feature weighted projection data.

[0032] Step 47: Based on the local feature weighted projection data and the original grayscale reference of the initial CT tomographic image sequence, perform spatial domain weighted superposition and dynamic range normalization processing to suppress background noise and enhance the contrast of the target area, thereby obtaining a feature-enhanced CT tomographic image sequence.

[0033] Furthermore, through feature-enhanced CT tomographic image sequences, streaming adaptive coding based on structural importance assessment is performed to obtain a lightweight real-time image data stream, including:

[0034] Step 51: Based on the feature-enhanced CT tomographic image sequence, perform frequency domain decomposition and spatial texture feature analysis, calculate the structural complexity and marginality index of each image block, and obtain the structural importance evaluation matrix;

[0035] Step 52: Based on the structural importance assessment matrix, establish differentiated quantization mapping rules, perform coarse-grained compression and dimensionality reduction on blocks representing low-frequency background redundancy, and perform fine-grained fidelity coding on blocks representing high-frequency pathological details to obtain adaptive code rate allocation data.

[0036] Step 53: Allocate data through adaptive bitrate, perform streaming data encapsulation and entropy coding operations, remove invalid and redundant information, and reassemble the compressed image data frames to obtain a lightweight real-time image data stream.

[0037] Secondly, a real-time CT image processing system based on edge computing includes:

[0038] The acquisition module is used to acquire the raw projection data stream output in real time by the CT scanning device, and perform data buffering and time series alignment processing at the edge computing node to obtain a synchronized raw projection data frame sequence.

[0039] The correction module is used to receive the synchronous original projection data frame sequence, eliminate signal distortion caused by tissue density differences and X-ray hardening, and obtain the corrected projection data frame sequence.

[0040] The iterative module is used to obtain the initial CT tomographic image sequence by combining the corrected projection data frame sequence with the real-time available computing power and memory status of the edge nodes to adaptively configure the reconstruction iterative parameters.

[0041] The computation module receives the initial CT tomographic image sequence, constructs a voxel-level spatial potential energy distribution field, introduces a spatial field diffusion and local feature affinity calculation mechanism based on image gradient vectors, establishes geometric boundary evolution constraints and mechanical equilibrium conditions, and converges to obtain the precise topological boundary of the pathologically sensitive region under kinematic trajectory constraints by iteratively deducing the curvature change and density potential difference of the morphological contour, thus obtaining the corresponding non-uniform weight mask. The initial CT tomographic image sequence and the non-uniform weight mask are then spatially weighted and fused to obtain a feature-enhanced CT tomographic image sequence.

[0042] The processing module is used to perform streaming adaptive coding processing based on structural importance assessment on feature-enhanced CT tomographic image sequences to obtain a lightweight real-time image data stream.

[0043] Thirdly, a computing device, comprising:

[0044] One or more processors;

[0045] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0046] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0047] The above-described solution of the present invention has at least the following beneficial effects:

[0048] This distributed processing architecture, which offloads the entire CT image processing task to edge computing nodes, combines an adaptive reconstruction iteration parameter configuration mechanism based on the real-time available computing power and memory status of edge nodes, a non-uniform weight mask acquisition method based on voxel-level spatial potential energy distribution field and geometric boundary evolution, and streaming adaptive coding technology based on structural importance assessment. Therefore, it overcomes the technical problems of existing CT image processing technologies, such as the high latency of centralized cloud computing failing to meet real-time requirements, insufficient local embedded computing power leading to poor image quality, fixed parameters in traditional iterative reconstruction algorithms causing an imbalance between reconstruction speed and quality, the inability of global image enhancement methods to accurately distinguish between pathologically sensitive areas and normal tissue backgrounds and the tendency to amplify background noise, and the inability of fixed bitrate coding to balance pathological detail fidelity with lightweight transmission. This achieves millisecond-level low-latency and high-precision real-time processing of CT images, achieving a final balance between reconstruction efficiency and image quality under the constraint of limited computing resources at edge nodes. It improves the recognition accuracy of small lesions and blurred invasive boundaries, while significantly reducing the transmission bandwidth requirements of image data streams, fully meeting the stringent requirements of clinical applications such as intraoperative CT guidance, which demand both real-time computing performance and image quality. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a real-time CT image processing method based on edge computing, provided by an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of a real-time CT image processing system based on edge computing, provided by an embodiment of the present invention.

[0051] Figure 3 This is a diagram illustrating the arrival timestamp of the original data packet.

[0052] Figure 4 This is a diagram illustrating the timestamps after cache reordering.

[0053] Figure 5 This is a schematic diagram of the original projection data.

[0054] Figure 6 This is a schematic diagram of the corrected projection data. Detailed Implementation

[0055] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough understanding of the present disclosure and to fully convey its scope to those skilled in the art.

[0056] like Figure 1 As shown, an embodiment of the present invention proposes a real-time CT image processing method based on edge computing, the method comprising the following steps:

[0057] Step 1: Obtain the raw projection data stream output in real time from the CT scanning device, and perform data buffering and time series alignment processing at the edge computing node to obtain a synchronized raw projection data frame sequence.

[0058] Step 2: Receive the synchronized original projection data frame sequence, eliminate signal distortion caused by tissue density differences and X-ray hardening, and obtain the corrected projection data frame sequence.

[0059] Step 3: Using the corrected projection data frame sequence, combined with the real-time available computing power and memory status of the edge nodes, the reconstruction iteration parameters are adaptively configured to obtain the initial CT tomographic image sequence.

[0060] Step 4: Receive the initial CT tomographic image sequence, construct a voxel-level spatial potential energy distribution field, introduce a spatial field diffusion and local feature affinity calculation mechanism based on image gradient vectors, establish geometric boundary evolution constraints and mechanical equilibrium conditions, and iteratively deduce the curvature change and density potential difference of the morphological contour. Under kinematic trajectory constraints, converge to obtain the precise topological boundary of the pathologically sensitive region and obtain the corresponding non-uniform weight mask. Spatial domain weighted fusion of the initial CT tomographic image sequence and the non-uniform weight mask is performed to obtain a feature-enhanced CT tomographic image sequence.

[0061] Step 5: Perform streaming adaptive coding processing based on structural importance assessment on the feature-enhanced CT tomographic image sequence to obtain a lightweight real-time image data stream.

[0062] In this embodiment of the invention, by performing data buffering and time-series alignment processing on the raw projection data stream output in real time from the CT scanning device at the edge computing node, dynamic attenuation correction and scattering compensation calculations based on scanning geometric parameters, iterative tomographic reconstruction calculations combining the real-time available computing power and memory status of the edge node to adaptively configure reconstruction iteration parameters, non-uniform weighted masking based on voxel-level spatial potential energy distribution field and geometric boundary evolution and spatial domain weighted fusion, and streaming adaptive coding based on structural importance assessment, the technical problems of subsequent processing errors caused by temporal misalignment during the transmission of raw projection data, signal distortion caused by tissue density differences and ray hardening, imbalance between reconstruction speed and quality caused by fixed parameters of traditional iterative reconstruction algorithms that cannot adapt to dynamic resource changes of edge nodes, inability of global image enhancement methods to accurately extract the topological boundaries of pathologically sensitive areas and easy amplification of background noise, and inability of fixed bitrate coding to balance pathological detail fidelity and lightweight image data stream are achieved. Thus, low-latency, high-precision real-time processing of the entire CT image process is realized, achieving the final balance between reconstruction efficiency and image quality under the constraint of limited computing resources at the edge, and improving the recognition accuracy of small lesions and blurred infiltrative boundaries.

[0063] In a preferred embodiment of the present invention, step 1 above may include:

[0064] Step 11: Acquire the raw projection data stream output in real time from the CT scanning device via a high-speed data communication interface. Parse the frame synchronization identifier and timestamp fields in the data stream protocol layer to extract the raw projection data packets with time-series markers. Specifically, the edge computing node establishes a point-to-point direct communication link with the CT scanning device via the high-speed data communication interface, acquiring the raw projection data stream output in real time from the CT scanning device. The data stream transmission rate is set to 10Gbps. The protocol parsing unit of the edge computing node parses the raw data stream layer by layer according to the standard, locating the frame synchronization identifier field in the data stream protocol layer. The frame synchronization identifier uses a fixed 32-bit binary sequence. When the parsing unit detects two consecutive identical frame synchronization identifiers, it determines that a complete raw projection data frame has begun. The parsing unit extracts a 64-bit high-precision timestamp field from the header of each raw projection data frame. The timestamp precision is 1 microsecond. The timestamp field records the absolute moment when the projection data frame was acquired by the CT detector array. The parsing unit binds and encapsulates the extracted timestamp field with the corresponding raw projection data frame to obtain the raw projection data packets with time-series markers.

[0065] Step 12: Receive the raw projection data packets with time-series markers. Perform data buffering processing in the preset buffer of the edge computing node. Reorder out-of-order data packets and perform frame-breaking linear interpolation to obtain a sequence of projection data blocks to be aligned, arranged continuously according to the acquisition time sequence. Specifically, the edge computing node sequentially reads the raw projection data packets with time-series markers from the receiving queue and writes them into a 256MB circular preset buffer. The preset buffer uses a first-in-first-out (FIFO) read / write mechanism. When the buffer occupancy reaches 80%, a buffer cleanup mechanism is automatically triggered, deleting the earliest processed raw projection data packet. The edge computing node monitors the timestamp sequence of the raw projection data packets in the preset buffer in real time. When the timestamp difference between two adjacent raw projection data packets is greater than 1 millisecond, it is determined that out-of-order data packets exist. The edge computing node initiates a timestamp-based priority queue reordering mechanism to rearrange all unprocessed raw projection data packets in the preset buffer in ascending order of timestamp, obtaining a preliminary sequence of data packets arranged according to the acquisition time sequence. Edge computing nodes traverse the initially arranged data packet sequence. When a timestamp difference between two adjacent data packets is greater than 2 milliseconds, it is determined that a frame loss has occurred. The edge computing nodes perform linear interpolation on the frame loss location using the following formula: ;

[0066] in, The interpolated frame projection values ​​are the interpolated values. This is the projection value of the data packet before the frame was broken. This is the projection value of the data packet after the frame is broken. This represents the theoretical acquisition time corresponding to the frame break. The timestamp of the data packet before the frame was broken. This is the timestamp of the data packet following the frame break. The edge computing node inserts the interpolated frame break data packet into the corresponding time sequence position, resulting in a continuous arrangement of the projection data blocks to be aligned according to the acquisition time sequence.

[0067] Step 13: Based on the projection data block to be aligned, perform multi-channel time series alignment processing using the rotation phase angle of the scanning device to compensate for timing misalignment caused by network transmission jitter, obtaining a synchronized sequence of original projection data frames. Specifically, this includes: the edge computing node reading the rotation phase angle of the CT scanning device in real time through the rotating encoder interface. The sampling frequency of the rotation phase angle is 1000Hz, and the accuracy of the rotation phase angle is 0.01 degrees. The edge computing node establishes a one-to-one mapping relationship between the rotation phase angle and the projection angle, with each projection angle corresponding to one original projection data frame. The edge computing node extracts the timestamp of each original projection data frame in the projection data block to be aligned and calculates the theoretical rotation phase angle corresponding to each original projection data frame. The formula for calculating the theoretical rotation phase angle is: ;

[0068] in, The theoretical rotation phase angle corresponding to the original projected data frame. The rotation phase angle at the start of the scan. The constant rotational angular velocity of the CT scanning equipment, The timestamp of the original projected data frame. This is the timestamp of the scan start time. The edge computing node compares the theoretical rotation phase angle corresponding to each original projected data frame with the actual read rotation phase angle to calculate the timing misalignment of each original projected data frame. The edge computing node performs time-axis translation compensation on each original projected data frame based on the timing misalignment; the compensation amount is equal to the timing misalignment. The edge computing node performs the same compensation operation on the original projected data frames of all channels. When the timing error of the original projected data frames of all channels is less than 0.5 milliseconds, the time series alignment is considered complete, resulting in a synchronized sequence of original projected data frames.

[0069] In this embodiment of the invention, the original projection data stream output in real time by the CT scanning device is acquired through a high-speed data communication interface, and the frame synchronization identifier and timestamp field in the data stream protocol layer are parsed to extract the original projection data packets with time sequence marks. Data buffering is performed in the preset buffer area of ​​the edge computing node, and the out-of-order data packets are queued and linearly interpolated for broken frames. Multi-channel time series alignment processing is performed in combination with the phase angle rotation of the scanning device. Therefore, the technical problem of data packet out-of-order, broken frame loss and multi-channel time sequence misalignment caused by network transmission jitter during the transmission of original projection data, which leads to systematic errors in subsequent correction and reconstruction operations, is overcome. Thus, the timing deviation caused by network transmission is effectively compensated, and a strictly time-synchronized original projection data frame sequence is obtained.

[0070] In a preferred embodiment of the present invention, step 2 above may include:

[0071] Step 21: Based on the metadata header of the synchronized original projection data frame sequence, extract the 3D coordinates of the X-ray source focus, the physical size matrix of the detector array, and the real-time projection angle sequence to construct a dynamic scanning geometric parameter set. Specifically, the edge computing node reads the metadata header information corresponding to the synchronized original projection data frame sequence, extracts the 3D spatial coordinate values ​​of the X-ray source focus frame by frame through the metadata parsing module, extracts the physical size matrix parameters of the detector array, and extracts the projection angle sequence parameters obtained in real-time during the scanning process. The edge computing node integrates and encapsulates the extracted 3D coordinates of the X-ray source focus, the physical size matrix of the detector array, and the real-time projection angle sequence to construct the dynamic scanning geometric parameter set.

[0072] Step 22: Receive the dynamic scanning geometric parameter set and the synchronized original projection data frame sequence, perform source-distance inverse square attenuation correction and air background subtraction calculations to compensate for physical energy loss along the ray propagation path, and obtain the attenuation-compensated projection data frame sequence. Specifically, the edge computing node calls the dynamic scanning geometric parameter set and the synchronized original projection data frame sequence, starts the ray energy correction module to perform source-distance inverse square attenuation correction calculation, and the attenuation correction calculation formula is: ;

[0073] in, This is the projected intensity value after attenuation correction. The initial intensity value of the original projected data frame. This represents the real-time propagation distance from the X-ray source to the detector unit. The system's preset standard reference distance is set to 500 mm. After attenuation correction is completed, the edge computing nodes perform air background subtraction calculations using the following formula: ;

[0074] in, This is the projected intensity value after attenuation compensation. The baseline value of the air background is collected during pre-scanning. The edge computing nodes perform attenuation correction and air background subtraction operations pixel by pixel to compensate for all physical energy loss on the ray propagation path, resulting in an attenuation-compensated projection data frame sequence. The precision of the operation is set to 16-bit fixed-point numbers.

[0075] Step 23: Using the attenuation-compensated projection data frame sequence, a scattered radiation distribution model is constructed based on the projection path integral thickness. Scattered component stripping and multi-spectral hardening artifact suppression operations are performed to eliminate signal distortion caused by tissue density differences and X-ray hardening, resulting in a corrected projection data frame sequence. Specifically, this includes: edge computing nodes calling the attenuation-compensated projection data frame sequence and constructing a scattered radiation distribution model based on the projection path integral thickness. The formula for calculating the projection path integral thickness is: ;

[0076] in, The total integral thickness of the projection path. This represents the segment length value on the projection path. This represents the tissue voxel density value for the corresponding path segment. The calculation formula for the scattered radiation distribution model is: ;

[0077] in, This represents the numerical value of the scattered radiation intensity. To fix the scattering coefficient, its value is set to 0.12. The edge computing nodes perform scattering component stripping calculations based on the scattered radiation distribution model. The stripping formula is: ;

[0078] in, This represents the projected intensity value after stripping the scattering component. The intensity values ​​of the projected data frame sequence are used to attenuate and compensate. Subsequently, the edge computing nodes perform multi-spectral hardening artifact suppression operations, employing a third-order polynomial correction formula: ;

[0079] Among them, the correction coefficient The value is set to 0.002, and the correction factor is... The value is set to -0.05, and the correction factor is... The value is set to 1.2, and the correction factor is... The value was set to 0.03. After the edge computing nodes completed all the calculations, the signal distortion caused by tissue density differences and ray hardening was completely eliminated, the artifact suppression efficiency reached 95%, and the corrected projection data frame sequence was finally obtained.

[0080] In this embodiment of the invention, a dynamic scanning geometric parameter set is constructed by extracting the three-dimensional coordinates of the X-ray source focus, the physical size matrix of the detector array, and the real-time projection angle sequence from the metadata header of the synchronized original projection data frame sequence. The source distance inverse square attenuation correction and air background subtraction are performed in combination with the dynamic scanning geometric parameters. Furthermore, a scattered radiation distribution model is constructed based on the integral thickness of the projection path, and scattered component stripping and multi-energy spectrum hardening artifact suppression are performed. Therefore, the technical problems of traditional static geometric parameter correction being unable to adapt to the dynamic changes of geometric parameters during scanning, the lack of effective compensation for the physical energy loss of the X-ray propagation path, and the signal distortion and imaging artifacts caused by tissue density differences and X-ray hardening are overcome. This effectively eliminates systematic errors and various artifacts in the projection data, resulting in a high-quality corrected projection data frame sequence.

[0081] In a preferred embodiment of the present invention, step 3 above may include:

[0082] Step 31: Real-time monitoring of the processor load rate, cache utilization, and number of parallel computing threads of the edge computing node to obtain a set of real-time available computing power and memory status indicators. Specifically, this includes: the edge computing node initiating real-time system resource monitoring, periodically collecting data on the node's hardware operating status at a fixed frequency of 100 Hz, real-time reading of the current CPU load rate (with a collection precision of 0.1%), real-time reading of the physical memory cache utilization (with a statistical precision of one megabyte), and real-time reading of the number of parallel computing threads of the graphics processing unit (with a collection precision of 1). The edge computing node then standardizes and integrates the real-time collected processor load rate, cache utilization, and number of parallel computing threads to obtain a set of real-time available computing power and memory status indicators.

[0083] Step 32: Receive the real-time available computing power and memory status indicator set and the corrected projection data frame sequence. Analyze the spatial resolution and slice thickness parameters of the projection data frame sequence. Based on the resource indicator set and data scale mapping, obtain the initial reconstruction iteration parameters. Specifically, the edge computing node receives the real-time available computing power and memory status indicator set and the corrected projection data frame sequence. It starts the data parameter parsing module, parses the corrected projection data frame sequence frame by frame, and extracts the spatial resolution and slice thickness parameters of the projection data. The resolution accuracy is set to 0.01 mm, and the slice thickness accuracy is set to 0.1 mm. The edge computing node establishes a linear mapping relationship based on the data scale of the real-time available computing power and memory status indicator set and the projection data frame sequence. It calculates the initial reconstruction iteration parameters through mapping rules. The initial reconstruction iteration parameters include the initial number of iterations, the initial number of parallel computing threads, and the initial regularization coefficient. The initial number of iterations is set to 20, the initial number of parallel computing threads is consistent with the maximum number of available threads monitored in real time, and the initial regularization coefficient is set to 0.05. The edge computing node transmits the initial reconstruction iteration parameters to the iterative reconstruction operation module to provide basic parameter support for subsequent operations.

[0084] Step 33: Based on the initial reconstruction iteration parameters and the corrected projection data frame sequence, perform forward projection simulation and backward error propagation operations to calculate the projection residual field and dynamically update the regularization coefficients, obtaining the adaptive reconstruction iteration parameters. Specifically, the edge computing node calls the initial reconstruction iteration parameters and the corrected projection data frame sequence to start the iterative operation unit to perform the forward projection simulation operation. The calculation formula for the forward projection simulation is: ;

[0085] in, These are the simulated intensity values ​​based on forward projection. The system projection matrix, This represents the current image voxel matrix. After completing the forward projection simulation, the edge computing nodes perform backward error propagation to calculate the projection residual field. The formula for calculating the projection residual field is: ;

[0086] in, The values ​​are the projection residual fields. This is the corrected actual projected intensity value. This represents the simulated intensity value from the forward projection. The edge computing nodes dynamically update the regularization coefficients based on the projected residual field values. The regularization coefficient update formula is: ;

[0087] in, The updated regularization coefficients, The regularization coefficient is the value from the previous time step. To maintain a fixed update coefficient, the update coefficient value is set to 0.02. This represents the absolute value of the projected residual field. The edge computing nodes integrate the updated regularization coefficients with the number of iterations and the number of parallel threads to obtain the adaptive reconstruction iteration parameters.

[0088] Step 34: Using the adaptive reconstruction iteration parameters and the corrected projection data frame sequence, perform algebraic iterative reconstruction operations, iteratively mapping between the projection space and the image space until the residuals converge, to obtain the initial CT tomographic image sequence. Specifically, this includes: the edge computing node calling the adaptive reconstruction iteration parameters and the corrected projection data frame sequence to perform algebraic iterative reconstruction operations. The calculation formula for algebraic iterative reconstruction is: ;

[0089] in, For the first The image voxel matrix of the next iteration For the first The image voxel matrix of the next iteration The regularization coefficients are updated in real time. It is the transpose of the projection matrix. This is the corrected actual projected intensity value. The system projection matrix is ​​used. The edge computing node performs a cyclic mapping iterative operation between the projection space and the image space, and calculates the projection residual field after each iteration in real time. When the value of the projection residual field is less than 0.001, the iterative operation is determined to have reached the convergence condition and the iteration is terminated. The edge computing node performs format conversion and 3D slice reconstruction on the converged image voxel matrix to obtain the initial CT tomographic image sequence. The grayscale precision of the image sequence is set to 16 bits.

[0090] In this embodiment of the invention, the processor load rate, cache utilization rate, and number of parallel computing threads of the edge computing nodes are monitored in real time to obtain a set of real-time available computing power and memory status indicators. Initial reconstruction iteration parameters are obtained by combining the spatial resolution and slice thickness parameters of the projection data frame sequence. Then, the projection residual field is calculated through forward projection simulation and backward error propagation, and the regularization coefficient is dynamically updated to obtain adaptive reconstruction iteration parameters. Finally, algebraic iterative reconstruction operation is performed. Therefore, this method overcomes the technical problems of traditional iterative reconstruction algorithms, which use fixed parameter configurations and cannot adapt to the dynamically changing computing power and memory resources of edge nodes, resulting in a sharp drop in reconstruction speed under high load and resource waste under low load. Furthermore, fixed regularization coefficients cannot balance the contradiction between noise suppression and detail preservation. Thus, under the constraint of limited computing resources at the edge, a dynamic final balance between reconstruction efficiency and image quality is achieved, ensuring the stability of the reconstruction frame rate under different load conditions. At the same time, through the adaptive adjustment of the regularization coefficient driven by the residual, reconstruction noise is effectively suppressed and the fine structural features of the image are preserved.

[0091] In a preferred embodiment of the present invention, step 4 above may include:

[0092] Step 41: Based on the initial CT tomographic image sequence, extract the voxel density values ​​and 3D spatial coordinate indices of each slice, perform nonlinear potential energy mapping operations, and construct a voxel-level spatial potential energy distribution field. Specifically, the edge computing node reads the initial CT tomographic image sequence, traverses all voxel units slice by slice, extracts the grayscale value corresponding to each voxel as the voxel density value, and simultaneously obtains the 3D spatial coordinate index corresponding to each voxel. The 3D spatial coordinate index consists of the slice number, row number, and column number. The edge computing node performs nonlinear potential energy mapping operations on each voxel density value. The calculation formula for nonlinear potential energy mapping is: ;

[0093] in, coordinates The spatial potential energy of a voxel. coordinates The density value of the voxel. This represents the average density value of normal tissue, set to 1000 Henle units. The nonlinear mapping coefficient is set to 0.002. The edge computing node arranges the spatial potential energy values ​​of all voxels according to the three-dimensional spatial coordinate index to construct a voxel-level three-dimensional spatial potential energy distribution field.

[0094] Step 42: Using the voxel-level spatial potential energy distribution field, calculate the gray-level difference vector in the local neighborhood to obtain the image gradient vector field. Perform isotropic diffusion filtering and cross-scale feature similarity measurement to obtain the local feature affinity calculation matrix. Specifically, the edge calculation node calls the voxel-level spatial potential energy distribution field to perform gray-level difference operation on the 3×3×3 local neighborhood of each voxel, and calculates the image gradient vector field at each voxel location. The formula for calculating the image gradient vector field is: ;

[0095] in, Let be the gradient vector at coordinates (x, y, z). , , These are the partial derivatives of the spatial potential energy value in the x, y, and z directions, respectively. The edge computing nodes perform isotropic diffusion filtering on the image gradient vector field. The formula for isotropic diffusion filtering is: ;

[0096] in, This represents the spatial potential energy value after the (t+1)th diffusion. Let be the spatial potential energy value after the t-th diffusion. The diffusion coefficient is set to 0.1. The Laplacian operator is used. The number of iterations for isotropic diffusion filtering is set to 5. After diffusion filtering, the edge computing nodes construct a three-layer Gaussian pyramid to perform multi-scale decomposition of the spatial potential energy distribution field, calculate the feature similarity of corresponding voxel positions at different scales, and use cosine similarity as the feature similarity measure. The calculation formula is: ;

[0097] in Let be the feature similarity between voxel i and voxel j. Let i be the multi-scale feature vector of voxel i. Let be the multi-scale feature vector of voxel j. The edge computing node integrates the feature similarity of all voxel pairs to obtain the local feature affinity calculation matrix, and the dimension of the matrix is ​​consistent with the total number of voxels.

[0098] Step 43: By calculating the local feature affinity matrix and combining it with the gradient transition characteristics of the voxel-level spatial potential energy distribution field, construct the geometric boundary evolution constraint rules and local tension balance conditions, and initialize the discrete control point set of the morphological contour. Specifically, this includes: the edge calculation node calling the local feature affinity calculation matrix and the voxel-level spatial potential energy distribution field to calculate the gradient magnitude of the spatial potential energy distribution field. The formula for calculating the gradient magnitude is: ;

[0099] Edge computing nodes extract voxels with gradient magnitudes greater than 0.5 as candidate boundary points to construct an initial boundary point set. The edge computing nodes then use a local feature affinity calculation matrix to filter the initial boundary point set, retaining boundary points with feature affinity greater than 0.8 as valid boundary points. Based on the valid boundary points, the edge computing nodes construct geometric boundary evolution constraints and local tension balance conditions. The geometric boundary evolution constraints stipulate that the boundary contour must remain continuous and non-self-intersecting. The formula for calculating the local tension balance condition is as follows: ;

[0100] in, For the first The local tension at each control point This is the tension coefficient, and its value is set to 0.3. For the first The coordinates of the control points and The first The coordinates of the previous and next control points of each control point. The ideal spacing for the control points is set to 2 millimeters. Edge calculation nodes select one point every 2 millimeters from the effective boundary point set as initial control points, resulting in a discrete control point set for the morphological profile.

[0101] Step 44: Based on the discrete control point set, perform iterative evolution operations under preset kinematic trajectory constraints. Calculate the rate of curvature change of the morphological contour and the density potential difference between adjacent voxels in real time, driving the control points to shrink towards the high gradient potential well region. The iteration terminates when the evolution displacement is lower than a preset convergence threshold. Extract the closed contour and perform gridded raster mapping to obtain the corresponding non-uniform weight mask. Specifically, the edge computing node calls the discrete control point set and performs iterative evolution operations under preset kinematic trajectory constraints. The kinematic trajectory constraints stipulate that control points can only move in the opposite direction of the gradient vector. During each iteration, the edge computing node calculates the rate of curvature change of the morphological contour at each control point in real time. The formula for calculating the rate of curvature change is: ;

[0102] in, For the first The rate of curvature change at each control point. The edge computing nodes simultaneously calculate the density potential difference between each control point and its neighboring voxels, which is the difference between the spatial potential energy at the control point and the spatial potential energy of its neighboring voxels along the gradient direction. The edge computing nodes calculate the displacement of each control point based on the rate of curvature change and the density potential difference. The formula for calculating the displacement is: ;

[0103] in, For the first Displacement of each control point This is the evolution step size coefficient, and its value is set to 0.1. For the first Density potential difference at each control point For the first The gradient vectors of each control point are calculated. The edge computing node updates the coordinates of each control point based on the displacement. When the displacement of all control points in a single iteration is less than 0.01 mm, the iteration is considered to have reached convergence and the iteration terminates. The maximum number of iterations is set to 100. The edge computing node connects the converged control points sequentially to form a closed contour. A gridded raster mapping process is then performed on the closed contour to obtain a binary contour mask with the same resolution as the initial CT tomographic image sequence. Based on the binary contour mask, the edge computing node calculates the distance from each voxel to the nearest contour boundary and obtains a non-uniform weight mask based on the distance. The formula for calculating the weight value is: ;

[0104] in, coordinates The weighting values ​​of voxels, This is the distance from the voxel to the nearest contour boundary. This is the weight decay factor, and the value of the weight decay factor is set to 1.5 mm.

[0105] In this embodiment of the invention, voxel density values ​​and three-dimensional spatial coordinate indices are extracted from the initial CT tomographic image sequence, and nonlinear potential energy mapping is performed to construct a voxel-level spatial potential energy distribution field. The local neighborhood gray-level difference vector is calculated to obtain the image gradient vector field, and isotropic diffusion filtering and cross-scale feature similarity measurement are performed to obtain the local feature affinity calculation matrix. The gradient transition features of the voxel-level spatial potential energy distribution field are combined to construct geometric boundary evolution constraint rules and local tension balance conditions to initialize the discrete control point set of the morphological contour. Under the constraints of preset kinematic trajectory, the non-uniform weight mask is obtained by iteratively deducing the rate of change of curvature of the morphological contour and the density potential difference of adjacent voxels to drive the convergence of control points. Therefore, this method overcomes the technical problems of traditional image boundary extraction methods, such as the inability to accurately identify the blurred infiltrative pathological boundaries and the topological structure of small lesions, the susceptibility to background noise interference leading to boundary positioning deviation, and the inability to obtain differentiated regional weight distribution. Thus, it achieves sub-pixel level extraction of accurate topological boundaries of pathologically sensitive areas, and the obtained non-uniform weight mask accurately distinguishes pathological areas from normal tissue background.

[0106] In a preferred embodiment of the present invention, step 4 above may include:

[0107] Step 45: Based on the initial CT tomographic image sequence and the non-uniform weighted mask, perform spatial coordinate registration and voxel-level mapping alignment processing to ensure that the weight distribution in the non-uniform weighted mask corresponds one-to-one with the anatomical structures of the initial CT tomographic image sequence, obtaining spatially aligned image mask pairs. Specifically, the edge computing node uses the three-dimensional coordinate system of the initial CT tomographic image sequence as the registration reference to extract the three-dimensional spatial coordinate indices of all voxels in the initial CT tomographic image sequence, and simultaneously extracts the three-dimensional spatial coordinate indices of all voxels in the non-uniform weighted mask. The edge computing node extracts the anatomical structure boundary feature points in the initial CT tomographic image sequence, and simultaneously extracts the contour boundary feature points in the non-uniform weighted mask, constructing two sets of feature point sets. The edge computing node uses the least squares method to calculate the rigid transformation matrix between the two sets of feature point sets. The rigid transformation matrix includes translation and rotation transformation parameters. The edge computing node performs a three-dimensional spatial coordinate transformation on the non-uniform weighted mask according to the rigid transformation matrix, mapping the coordinate system of the non-uniform weighted mask to the coordinate system of the initial CT tomographic image sequence. The edge computing node performs voxel-level mapping alignment to ensure that the position of each voxel in the non-uniform weighted mask completely coincides with the corresponding voxel position in the initial CT tomographic image sequence. The edge computing node calculates the registration error, which is the Euclidean distance between corresponding feature points. When the registration error of all feature points is less than 0.01 mm, the spatial coordinate registration and voxel-level mapping alignment are considered complete, resulting in a spatially aligned image mask pair.

[0108] Step 46: Receive spatially aligned image mask pairs, extract the weight coefficients for each voxel position in the non-uniform weighted mask, and perform point-by-point multiplication with the corresponding voxel gray values ​​from the initial CT tomographic image sequence to obtain local feature weighted projection data. Specifically, this includes: the edge computing node receiving the spatially aligned image mask pairs and traversing all voxel units in the image mask pairs voxel-by-voxel. The edge computing node extracts the weight coefficients for each voxel position in the non-uniform weighted mask and simultaneously extracts the gray values ​​for the corresponding voxel positions from the initial CT tomographic image sequence. The edge computing node performs point-by-point multiplication with the weight coefficients and gray values ​​for each voxel position. The formula for the multiplication operation is: ;

[0109] in, coordinates Weighted projection values ​​of local features at the location, coordinates The grayscale value of the initial CT tomographic image. The weight coefficients are the weights of the non-uniform weight mask at coordinates (x, y, z). The edge computing node arranges the local feature weighted projection values ​​of all voxels according to the three-dimensional spatial coordinate index to generate local feature weighted projection data.

[0110] Step 47: Based on the local feature-weighted projection data and the original grayscale reference of the initial CT tomographic image sequence, perform spatial domain weighted superposition and dynamic range normalization processing to suppress background noise and enhance the contrast of the target area, obtaining a feature-enhanced CT tomographic image sequence. Specifically, the edge computing node receives the local feature-weighted projection data and the original grayscale reference of the initial CT tomographic image sequence, and performs spatial domain weighted superposition operation. The calculation formula for the weighted superposition operation is: ;

[0111] in, coordinates Enhanced grayscale value at the location, These are the original image weighting coefficients, with a value set to 0.3. The weighting coefficient for the weighted projection data is set to 0.7. After the weighted overlay operation is completed, the edge computing node performs dynamic range normalization, mapping the grayscale values ​​of the enhanced image to the standard CT value range. The formula for dynamic range normalization is: ;

[0112] in, This represents the normalized grayscale value at coordinates (x, y, z). The minimum grayscale value of the enhanced image. To maximize the grayscale value of the enhanced image. This represents the lower limit of the standard CT value range, set to -1000 Henle units. The upper limit of the standard CT value range is set to 3,000 Henle units. After the edge computing nodes complete all calculations, background noise in normal tissue is effectively suppressed, with a background noise suppression rate of 40%. At the same time, the contrast of pathologically sensitive areas is enhanced, with a contrast improvement rate of 60% in the target area, ultimately resulting in a feature-enhanced CT tomographic image sequence.

[0113] In this embodiment of the invention, spatial coordinate registration and voxel-level mapping alignment are performed on the initial CT tomographic image sequence and the non-uniform weighted mask. The weight coefficients of each voxel position in the non-uniform weighted mask are extracted and the corresponding voxel gray values ​​are multiplied point by point to obtain local feature weighted projection data. Then, combined with the original grayscale reference, spatial domain weighted superposition and dynamic range normalization are performed. Therefore, the technical problems of traditional global image enhancement methods, such as the inability to directionally enhance the target area, the significant amplification of background noise while enhancing pathological features, and the deviation of enhancement effect caused by spatial misalignment between the mask and the image, are overcome. Thus, directional and precise feature enhancement of pathologically sensitive areas is achieved. While effectively suppressing background noise of normal tissue, the contrast and clarity of the lesion area are improved, resulting in a high-quality feature-enhanced CT tomographic image sequence.

[0114] In a preferred embodiment of the present invention, step 5 above may include:

[0115] Step 51: Based on the feature-enhanced CT tomographic image sequence, perform frequency domain decomposition and spatial texture feature analysis to calculate the structural complexity and marginality index of each image block, obtaining a structural importance evaluation matrix. Specifically, this includes: the edge computing node reading the feature-enhanced CT tomographic image sequence and dividing each two-dimensional tomographic slice into non-overlapping image blocks, with each block size set to 16 pixels by 16 pixels. The edge computing node performing a two-dimensional discrete cosine transform operation on each image block to obtain the corresponding frequency domain coefficient matrix. The edge computing node calculating the structural complexity index of each image block is defined as the proportion of energy of high-frequency coefficients in the frequency domain coefficient matrix to the total energy, calculated using the following formula: ;

[0116] in, This is an indicator of the structural complexity of image blocks. This represents the sum of the energy of the high-frequency coefficients in the frequency domain coefficient matrix. The sum of the energy of all coefficients in the frequency domain coefficient matrix is ​​used. The edge computing node calculates the edge index of each image block and the gradient magnitude of each pixel within the image block. The formula for calculating the gradient magnitude is: ;

[0117] in, Let be the gradient magnitude of the pixel at coordinates (x, y). Let be the grayscale value of the pixel at coordinates (x, y), and let be the marginality index, which is the average gradient magnitude of all pixels within the image block. The edge computing node calculates the structural importance value of each image block based on the structural complexity index and the marginality index. The formula for calculating the structural importance value is: ;

[0118] in, This represents the structural importance value of an image region. As a structural complexity index, This is an edge indicator. Edge computing nodes arrange the structural importance values ​​of all image blocks according to their spatial position in the tomographic slices, resulting in a structural importance evaluation matrix.

[0119] Step 52: Based on the structural importance assessment matrix, establish differentiated quantization mapping rules. Perform coarse-grained compression and dimensionality reduction on blocks representing low-frequency background redundancy, and perform fine-grained fidelity coding on blocks representing high-frequency pathological details to obtain adaptive code rate allocation data. Specifically, the edge computing node receives the structural importance assessment matrix and establishes a three-level differentiated quantization mapping rule. For low-frequency background redundancy blocks with a structural importance value less than 0.3, coarse-grained compression and dimensionality reduction are used, with a quantization step size of 32. For normal tissue blocks with a structural importance value between 0.3 and 0.7, medium-grained quantization is used, with a quantization step size of 16. For high-frequency pathological detail blocks with a structural importance value greater than 0.7, fine-grained fidelity coding is used, with a quantization step size of 4. The edge computing node performs quantization operations on the discrete cosine transform coefficients of each image block according to the differentiated quantization mapping rule to obtain the quantized frequency domain coefficient matrix. Edge computing nodes perform adaptive bitrate allocation calculations, with the total transmission bitrate set to 8 Mbps. The bitrate allocated to each image block is proportional to the structural importance value of that block. The calculation formula for bitrate allocation is as follows: ;

[0120] in, For the first The bitrate allocated to each image block, For the total transmission bit rate, For the first Structural importance values ​​of each image block The first section of the same fault slice Structural importance values ​​of each image block This represents the total number of image blocks in a single tomographic slice. Edge computing nodes encode the quantized frequency domain coefficient matrix according to the assigned bitrate, generating adaptive bitrate allocation data.

[0121] Step 53: Data is allocated through adaptive bitrate, streaming data encapsulation and entropy coding operations are performed, invalid and redundant information is removed, and the compressed image data frames are reassembled to obtain a lightweight real-time image data stream. Specifically, this includes: edge computing nodes receiving adaptive bitrate allocated data and performing streaming data encapsulation processing. The encoded data of each tomographic slice is divided into multiple independent data units according to the spatial order of the image blocks. Each data unit contains the encoded data of the corresponding image block, spatial coordinate index, and timestamp information. The timestamp precision is set to 1 microsecond. The edge computing nodes perform context-adaptive binary arithmetic coding operations on each data unit to remove invalid and redundant information. The edge computing nodes reassemble the encoded data units according to time order, adding frame header and frame tail identifiers to each complete tomographic slice. The frame header contains the frame sequence number, slice sequence number, and total data length information, while the frame tail contains cyclic redundancy check (CRC) information. The edge computing nodes encapsulate the reassembled data into data packets conforming to the real-time transmission protocol, ultimately obtaining a lightweight real-time image data stream.

[0122] In this embodiment of the invention, a structural importance evaluation matrix is ​​obtained by performing frequency domain decomposition and spatial texture feature analysis on feature-enhanced CT tomographic image sequences and calculating the structural complexity and marginality index of each image block. Based on the structural importance evaluation matrix, a differentiated quantization mapping rule is established to perform coarse-grained compression and dimensionality reduction on low-frequency background redundant blocks and fine-grained fidelity encoding on high-frequency pathological detail blocks to obtain adaptive bitrate allocation data. Then, streaming data encapsulation and entropy encoding operations are performed on the adaptive bitrate allocation data to remove invalid redundant information and reassemble the compressed image data frames. Therefore, this method overcomes the technical problems of traditional fixed bitrate encoding technology, which uses a uniform compression ratio for all image blocks, cannot distinguish between key pathological details and background redundant information, resulting in either loss of high-frequency pathological details affecting diagnostic accuracy or insufficient compression ratio causing waste of transmission bandwidth resources and lack of streaming real-time processing capabilities. This method achieves content-aware adaptive bitrate precise allocation, which significantly reduces the transmission bandwidth requirements of image data streams while fully preserving key pathological details. At the same time, streaming encapsulation processing ensures low-latency real-time transmission of data, supporting real-time remote consultation and multi-terminal synchronous display under low-bandwidth networks.

[0123] like Figure 2 As shown, embodiments of the present invention also provide a real-time CT image processing system based on edge computing, comprising:

[0124] The acquisition module is used to acquire the raw projection data stream output in real time by the CT scanning device, and perform data buffering and time series alignment processing at the edge computing node to obtain a synchronized raw projection data frame sequence.

[0125] The correction module is used to receive the synchronous original projection data frame sequence, eliminate signal distortion caused by tissue density differences and X-ray hardening, and obtain the corrected projection data frame sequence.

[0126] The iterative module is used to obtain the initial CT tomographic image sequence by combining the corrected projection data frame sequence with the real-time available computing power and memory status of the edge nodes to adaptively configure the reconstruction iterative parameters.

[0127] The computation module receives the initial CT tomographic image sequence, constructs a voxel-level spatial potential energy distribution field, introduces a spatial field diffusion and local feature affinity calculation mechanism based on image gradient vectors, establishes geometric boundary evolution constraints and mechanical equilibrium conditions, and converges to obtain the precise topological boundary of the pathologically sensitive region under kinematic trajectory constraints by iteratively deducing the curvature change and density potential difference of the morphological contour, thus obtaining the corresponding non-uniform weight mask. The initial CT tomographic image sequence and the non-uniform weight mask are then spatially weighted and fused to obtain a feature-enhanced CT tomographic image sequence.

[0128] The processing module is used to perform streaming adaptive coding processing based on structural importance assessment on feature-enhanced CT tomographic image sequences to obtain a lightweight real-time image data stream.

[0129] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0130] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0131] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0132] Experimental example:

[0133] Step 1: Data Acquisition, Buffering, and Time Series Alignment:

[0134] The CT scanner begins scanning, pushing the raw projection data stream to the edge node in real time via a 10Gbps fiber optic cable. The data stream simulates network jitter during transmission, resulting in out-of-order data packets and a delay of up to approximately 15ms. Using the timestamp in the packet header and the rotation phase angle read from the CT scanner encoder, the packets are reordered within a 256MB circular buffer, and lost frames are compensated for using linear interpolation. Ultimately, the timing misalignment error of the multi-channel data is compensated to within <0.5ms, resulting in a strictly synchronized sequence of projection data frames.

[0135] Figure 3 The timestamps of the original arriving data packets show out-of-order and gaps. Figure 4 The data packets, after being reordered in the buffer and linearly interpolated, are reorganized into a sequence that is strictly arranged in the order of acquisition, providing an accurate time reference for subsequent processing.

[0136] Step 2, Projection data correction and artifact elimination:

[0137] Based on the metadata of the projection data, a dynamic scanning geometry set is constructed. Inverse squared source distance attenuation correction and air background subtraction are performed. A scattering model is constructed based on the projection path integral thickness, and a third-order polynomial is used for energy spectrum hardening correction.

[0138] Figure 5 The original projection data shows cup-shaped artifacts (arrows) and scattering noise caused by ray hardening. Figure 6 After correction, artifacts and noise are effectively suppressed, and data consistency is improved.

[0139] Step 3, Adaptive Iterative Reconstruction:

[0140] The system monitors metrics such as node processor load and memory usage. When the initial load is low, the system is configured with 20 iterations and a large number of parallel threads for algebraic iterative reconstruction. As the load increases, the system dynamically adjusts the regularization coefficient λ based on the real-time calculated projection residual field. In step 34, the number of iterations is reduced to prioritize reconstruction speed and prevent increased latency.

[0141] Step 4, Boundary Evolution and Feature Enhancement:

[0142] A voxel-level spatial potential field is constructed based on the initial image. Using the gradient vector field and feature affinity matrix, and under kinematic constraints, the initial contour, for example a rough circle around a suspected lesion, evolves towards a high-gradient region. After iterative convergence, a precise, non-uniform weight mask with weights decreasing from the center to the edge is obtained. Finally, this mask is fused with the original image and its dynamic range is normalized.

[0143] Step 5, Streaming Adaptive Coding and Transmission:

[0144] Frequency domain and texture analysis are performed on the feature-enhanced image blocks to calculate the structural importance S of each block. Based on the S value, differential quantization encoding is performed: background regions with S < 0.3 have a quantization step size of 32; normal tissue with S ≤ 0.3 ≤ S ≤ 0.7 has a step size of 16; and lesion / vascular detail regions with S > 0.7 have a step size of 4. The bitrate is allocated proportionally according to importance. Entropy coding and streaming encapsulation are then performed.

[0145] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A real-time CT image processing method based on edge computing, characterized in that, The method includes: Step 1: Obtain the raw projection data stream output in real time from the CT scanning device, and perform data buffering and time series alignment processing at the edge computing node to obtain a synchronized raw projection data frame sequence. Step 2: Receive the synchronized original projection data frame sequence, eliminate signal distortion caused by tissue density differences and X-ray hardening, and obtain the corrected projection data frame sequence. Step 3: Using the corrected projection data frame sequence, combined with the real-time available computing power and memory status of the edge computing nodes, the reconstruction iteration parameters are adaptively configured to obtain the initial CT tomographic image sequence. Step 4: Receive the initial CT tomographic image sequence, construct a voxel-level spatial potential energy distribution field, introduce a spatial field diffusion and local feature affinity calculation mechanism based on image gradient vectors, establish geometric boundary evolution constraints and mechanical equilibrium conditions, and iteratively deduce the curvature changes and density potential differences of the morphological contour. Under kinematic trajectory constraints, converge to obtain the precise topological boundary of the pathologically sensitive region, obtaining the corresponding non-uniform weighted mask. Perform spatial domain weighted fusion of the initial CT tomographic image sequence and the non-uniform weighted mask to obtain a feature-enhanced CT tomographic image sequence, including: Step 41: Based on the initial CT tomographic image sequence, extract the voxel density values ​​and three-dimensional spatial coordinate indices of each slice, perform nonlinear potential energy mapping operation, and construct a voxel-level spatial potential energy distribution field; Step 42: Calculate the gray-level difference vector in the local neighborhood through the voxel-level spatial potential energy distribution field to obtain the image gradient vector field, perform isotropic diffusion filtering and cross-scale feature similarity measurement to obtain the local feature affinity calculation matrix. Step 43: Calculate the matrix by local feature affinity, combine the gradient transition characteristics of the voxel-level spatial potential energy distribution field, construct the geometric boundary evolution constraint rules and local tension balance conditions, and initialize the discrete control point set of the morphological contour. Step 44: Based on the discrete control point set, perform iterative evolution operation under the preset kinematic trajectory constraints, calculate the curvature change rate of the shape contour and the density potential difference of adjacent voxels in real time, drive the control points to shrink towards the high gradient potential well region, terminate the iteration when the evolution displacement is lower than the preset convergence threshold, extract the closed contour and perform gridded raster mapping to obtain the corresponding non-uniform weight mask. Step 5: Perform streaming adaptive coding processing based on structural importance assessment on the feature-enhanced CT tomographic image sequence to obtain a lightweight real-time image data stream.

2. The real-time CT image processing method based on edge computing according to claim 1, characterized in that, Step 1: Acquire the raw projection data stream output in real time from the CT scanning device, perform data buffering and time series alignment processing at the edge computing node to obtain a synchronized raw projection data frame sequence, including: Step 11: Acquire the raw projection data stream output in real time by the CT scanning device through the high-speed data communication interface, parse the frame synchronization identifier and timestamp field in the data stream protocol layer, and extract the raw projection data packet with time sequence mark; Step 12: Receive the original projection data packets with time sequence marks, perform data buffering processing in the preset buffer area of ​​the edge computing node, and perform queue reordering and frame-breaking linear interpolation on the out-of-order data packets to obtain the projection data blocks to be aligned in a continuous arrangement according to the acquisition time sequence. Step 13: Based on the projection data block to be aligned, perform multi-channel time series alignment processing in conjunction with the rotation phase angle of the scanning device to compensate for the timing misalignment caused by network transmission jitter and obtain a synchronized original projection data frame sequence.

3. The real-time CT image processing method based on edge computing according to claim 2, characterized in that, Step 2: Receive the synchronized original projection data frame sequence, eliminate signal distortion caused by tissue density differences and X-ray hardening, and obtain the corrected projection data frame sequence, including: Step 21: Based on the metadata header of the synchronized original projection data frame sequence, extract the three-dimensional coordinates of the X-ray source focus, the physical size matrix of the detector array, and the real-time projection angle sequence to construct a dynamic scanning geometric parameter set; Step 22: Receive the dynamic scanning geometric parameter set and the synchronized original projection data frame sequence, perform source distance inverse square attenuation correction and air background subtraction operation to compensate for the physical energy loss on the ray propagation path, and obtain the attenuation-compensated projection data frame sequence. Step 23: By using the attenuation-compensated projection data frame sequence, a scattered radiation distribution model is constructed based on the projection path integral thickness. Scattered component stripping and multi-spectral hardening artifact suppression operations are performed to eliminate signal distortion caused by tissue density differences and X-ray hardening, and the corrected projection data frame sequence is obtained.

4. The real-time CT image processing method based on edge computing according to claim 3, characterized in that, Step 3: Using the corrected projection data frame sequence, and combining the real-time available computing power and memory status of the edge computing nodes to adaptively configure the reconstruction iteration parameters, an initial CT tomographic image sequence is obtained, including: Step 31: Monitor the processor load rate, cache utilization rate and number of parallel computing threads of the edge computing nodes in real time to obtain a set of real-time available computing power and memory status indicators; Step 32: Receive the real-time available computing power and memory status index set and the corrected projection data frame sequence, parse the spatial resolution and slice thickness parameters of the projection data frame sequence, and obtain the initial reconstruction iteration parameters based on the resource index set and data scale mapping; Step 33: Based on the initial reconstruction iteration parameters and the corrected projection data frame sequence, perform forward projection simulation and backward error propagation operations, calculate the projection residual field and dynamically update the regularization coefficients to obtain the adaptive reconstruction iteration parameters; Step 34: Using the adaptive reconstruction iteration parameters and the corrected projection data frame sequence, perform algebraic iterative reconstruction operation, cyclically mapping between the projection space and the image space until the residual converges, to obtain the initial CT tomographic image sequence.

5. The real-time CT image processing method based on edge computing according to claim 4, characterized in that, The initial CT tomographic image sequence is spatially weighted and fused with a non-uniform weight mask to obtain a feature-enhanced CT tomographic image sequence, including: Step 45: Based on the initial CT tomographic image sequence and the non-uniform weight mask, perform spatial coordinate registration and voxel-level mapping alignment processing to ensure that the weight distribution in the non-uniform weight mask corresponds one-to-one with the anatomical structure of the initial CT tomographic image sequence, and obtain spatially aligned image mask pairs. Step 46: Receive spatially aligned image mask pairs, extract the weight coefficients of each voxel position in the non-uniform weight mask, and perform point-by-point multiplication with the corresponding voxel gray values ​​of the initial CT tomographic image sequence to obtain local feature weighted projection data. Step 47: Based on the local feature weighted projection data and the original grayscale reference of the initial CT tomographic image sequence, perform spatial domain weighted superposition and dynamic range normalization processing to suppress background noise and enhance the contrast of the target area, thereby obtaining a feature-enhanced CT tomographic image sequence.

6. The real-time CT image processing method based on edge computing according to claim 5, characterized in that, Step 5: Using the feature-enhanced CT tomographic image sequence, perform streaming adaptive coding processing based on structural importance assessment to obtain a lightweight real-time image data stream, including: Step 51: Based on the feature-enhanced CT tomographic image sequence, perform frequency domain decomposition and spatial texture feature analysis, calculate the structural complexity and marginality index of each image block, and obtain the structural importance evaluation matrix; Step 52: Based on the structural importance assessment matrix, establish differentiated quantization mapping rules, perform coarse-grained compression and dimensionality reduction on blocks representing low-frequency background redundancy, and perform fine-grained fidelity coding on blocks representing high-frequency pathological details to obtain adaptive code rate allocation data. Step 53: Allocate data through adaptive bitrate, perform streaming data encapsulation and entropy coding operations, remove invalid and redundant information, and reassemble the compressed image data frames to obtain a lightweight real-time image data stream.

7. A real-time CT image processing system based on edge computing, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The acquisition module is used to acquire the raw projection data stream output in real time by the CT scanning device, and perform data buffering and time series alignment processing at the edge computing node to obtain a synchronized raw projection data frame sequence. The correction module is used to receive the synchronous original projection data frame sequence, eliminate signal distortion caused by tissue density differences and X-ray hardening, and obtain the corrected projection data frame sequence. The iteration module is used to obtain an initial CT tomographic image sequence by combining the corrected projection data frame sequence with the real-time available computing power and memory status of the edge computing nodes to adaptively configure the reconstruction iteration parameters. The computation module receives the initial CT tomographic image sequence, constructs a voxel-level spatial potential energy distribution field, introduces a spatial field diffusion and local feature affinity calculation mechanism based on image gradient vectors, establishes geometric boundary evolution constraints and mechanical equilibrium conditions, and converges to obtain the precise topological boundary of the pathologically sensitive region under kinematic trajectory constraints by iteratively deducing the curvature change and density potential difference of the morphological contour, thus obtaining the corresponding non-uniform weight mask. The initial CT tomographic image sequence and the non-uniform weight mask are then spatially weighted and fused to obtain a feature-enhanced CT tomographic image sequence. The processing module is used to perform streaming adaptive coding processing based on structural importance assessment on feature-enhanced CT tomographic image sequences to obtain a lightweight real-time image data stream.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images