A ct image assisted detection method and system for liver focal lesions

Abstract

The application discloses a CT image auxiliary detection method and system for liver focal lesions, and the method comprises the following steps: receiving and screening multi-dimensional liver CT image data, generating a standardized DICOM data set based on preset layer thickness threshold, resolution threshold and artifact filtering rules; performing two-stage lesion positioning analysis on the standardized DICOM data set, and outputting a lesion positioning atlas containing lesion coordinates, size and average CT value; starting an interactive image optimization engine based on the lesion positioning atlas, responding to the window width and window level adjustment instructions and measurement tool operations of a doctor in real time, generating a dynamically enhanced lesion visual field and synchronously updating lesion parameters; integrating the visual field data and a clinical rule base to generate a structured report, automatically filling the image findings field by checking the lesion list, and outputting a diagnosis proposal conforming to the DICOM standard through a graphic-text report module. By using the embodiment of the application, the standardization and intelligentization of auxiliary diagnosis can be realized, and the precision and efficiency of liver lesion detection are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image recognition, and particularly relates to a CT image assisted detection method and system for liver focal lesions. BACKGROUND

[0002] CT detection of liver focal lesions is a key link in clinical diagnosis, but the traditional method relies on the experience of doctors, and has the problems of high missed diagnosis rate and low efficiency. In the prior art, CT image analysis mostly uses a single model to process global images, which is easily disturbed by artifacts, and lacks dynamic interactive function, resulting in insufficient lesion positioning accuracy. In addition, report generation is disconnected from image analysis, and clinical parameters (such as lesion size and CT value) need to be manually entered, which is easy to introduce errors. Although some systems attempt to combine AI assistance, they do not solve the technical pain points of multi-dimensional data standardization, real-time interactive optimization and structured report automatic generation, and it is difficult to meet the needs of accurate diagnosis and treatment. SUMMARY

[0003] The purpose of the present application is to provide a CT image assisted detection method and system for liver focal lesions, to solve the problems in the prior art, to realize the standardization and intelligentization of assisted diagnosis, to reduce human errors, and to improve the accuracy and efficiency of liver lesion detection.

[0004] One embodiment of the present application provides a CT image assisted detection method for liver focal lesions, which comprises:

[0005] Receiving and screening multi-dimensional liver CT image data, generating standardized DICOM data sets based on preset layer thickness threshold, resolution threshold and artifact filtering rules, wherein the data sets dynamically associate patient age, scanning device parameters and lesion size threshold;

[0006] Performing two-stage lesion positioning analysis on the standardized DICOM data set, first generating a global liver segmentation mask through a cascaded deep learning model, then extracting a local ROI region based on the mask coordinates for sub-pixel level lesion detection, and outputting a lesion positioning atlas containing lesion coordinates, size and average CT value;

[0007] Starting an interactive image optimization engine based on the lesion positioning atlas, real-time responding to the window width and window level adjustment instructions and measurement tool operations of the doctor, generating a dynamically enhanced lesion view domain and synchronously updating the lesion parameters;

[0008] Generating a structured report by integrating the view domain data and the clinical rule base, automatically filling the image findings field by checking the lesion list, and outputting a diagnosis proposal book conforming to the DICOM standard through the graphic-text report module, wherein the diagnosis proposal book embeds a lesion position navigation link and a measurement error calibration.

[0009] Optionally, the receiving and screening multi-dimensional liver CT image data, based on the preset layer thickness threshold, resolution threshold and artifact filtering rules to generate standardized DICOM data set, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold, including:

[0010] Receiving the original DICOM stream from the PACS system, synchronously accessing the patient age, device model and scanning protocol metadata, and generating the original data cube with space-time label;

[0011] Input the original data cube into the adaptive frequency domain filter, dynamically select the convolution kernel size based on the device parameters, eliminate metal artifacts and motion blur, and output the purified image sequence;

[0012] Performing interlayer consistency analysis on the purified image sequence, aligning adjacent slices through non-rigid registration algorithm, and generating sub-millimeter level volume data according to the preset layer thickness threshold;

[0013] Encode the patient age and lesion size threshold as a feature vector, and input it into the graph neural network together with the sub-millimeter level volume data to generate a dynamically weighted standardized DICOM data set.

[0014] Optionally, the double-stage lesion positioning analysis is performed on the standardized DICOM data set, and the cascade deep learning model is used to first generate a global liver segmentation mask, and then based on the mask coordinates, the local ROI region is extracted for sub-pixel level lesion detection, and the lesion positioning atlas containing lesion coordinates, size and average CT value is output, including:

[0015] Input the standardized DICOM data set into the cascade U-Net model, the first stage uses 3D dilated convolution to extract liver boundary features, and outputs the liver probability heat map;

[0016] Performing topological connectivity analysis on the liver probability heat map, correcting the edge through gradient vector flow field, generating sub-voxel level liver mask and corresponding spatial coordinate tree;

[0017] Based on the spatial coordinate tree, the local region of interest (ROI) is determined, and the deformable convolution network is used to adaptively adjust the ROI size, and the lesion candidate block set is output;

[0018] Apply phase consistency edge detection in the lesion candidate block set, combine with Hessian matrix to calculate lesion center offset, and generate sub-pixel positioning coordinates;

[0019] Fuse the sub-pixel positioning coordinates and CT value histogram statistics, and generate the lesion positioning atlas through radial basis function interpolation.

[0020] Optionally, the interactive image optimization engine is started based on the lesion positioning atlas, and real-time response is given to the window width and window level adjustment instructions and measurement tool operations of the physician to generate a dynamic enhanced lesion visual field and synchronously update lesion parameters, including:

[0021] Real-time reception of the window width and window level adjustment signals and measurement tool trajectory data of the physician, analysis of the operation type through a gesture intention recognition algorithm, generation of a window level and window width parameter vector and a measurement coordinate sequence;

[0022] Based on the spatial coordinate tree of the lesion positioning atlas, the corresponding ROI body data is preloaded in the GPU display memory, the light projection algorithm is dynamically adjusted in combination with the window level and window width parameter vector, and a real-time volume rendering image is output;

[0023] Anisotropic diffusion filtering is superimposed on the real-time volume rendering image, the lesion edge contrast is adaptively enhanced according to the window width parameter, the labeled trajectory of the measurement coordinate sequence is synchronously embedded, and an enhanced labeled image is generated;

[0024] The sub-pixel positioning coordinates of the lesion positioning atlas are compared with the measurement coordinate sequence, the lesion size deviation and the CT value fluctuation are calculated, and the lesion dynamic parameter table is updated in real time;

[0025] The enhanced labeled image and the lesion dynamic parameter table are fused, and a dynamic enhanced lesion visual field with a parameter superimposed layer is generated through multi-planar reformatting technology.

[0026] Optionally, the structured report is generated by integrating the visual field data and the clinical rule base, the image findings field is automatically filled by checking the lesion list, and the diagnosis proposal sheet conforming to the DICOM standard is output through the graphic-text report module, wherein the diagnosis proposal sheet is embedded with a lesion position navigation link and a measurement error calibration, including:

[0027] The quantitative data in the lesion dynamic parameter table are analyzed, and the feature matching is performed with the clinical rule base to generate a structured lesion data unit;

[0028] The lesion list checked by the physician is mapped to the structured lesion data unit, the structured diagnosis text is output through the medical knowledge graph driven natural language generation; and

[0029] The DICOM spatial coordinate system is generated based on the spatial coordinate tree of the lesion positioning atlas, the sub-pixel positioning coordinates are converted into a DICOM-RT structure set, and a lesion spatial navigation engine is created;

[0030] In combination with the structured diagnosis text and the lesion spatial navigation engine, the error heat map of the measurement coordinate sequence is added, and the diagnosis proposal sheet with error calibration is generated.

[0031] Still another embodiment of the present application provides a CT image assisted detection system for liver focal lesions, the system comprising:

[0032] The receiving module is configured to receive and screen multi-dimensional liver CT image data, generate a standardized DICOM data set based on preset layer thickness threshold, resolution threshold and artifact filtering rules, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold;

[0033] The analysis module is configured to perform two-stage lesion positioning analysis on the standardized DICOM data set, first generate a global liver segmentation mask through a cascaded deep learning model, then extract a local ROI region based on the mask coordinates for sub-pixel level lesion detection, and output a lesion positioning atlas containing lesion coordinates, size and average CT value.

[0034] The generating module is configured to start an interactive image optimization engine based on the lesion positioning atlas, real-time respond to the window width and window level adjustment instructions and measurement tool operations of the physician, generate a dynamically enhanced lesion visual field and synchronously update the lesion parameters.

[0035] The output module is configured to integrate the visual field data and the clinical rule base to generate a structured report, automatically fill the image findings field by checking the lesion list, and output a diagnosis proposal conforming to the DICOM standard through the graphic-text report module, wherein the diagnosis proposal is embedded with a lesion position navigation link and a measurement error calibration.

[0036] Still another embodiment of the present application provides a storage medium having a computer program stored therein, wherein the computer program is configured to execute the method described in any of the above embodiments when running.

[0037] Still another embodiment of the present application provides an electronic device comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is configured to execute the computer program to perform the method described in any of the above embodiments.

[0038] Compared with the prior art, the CT image assisted detection method for liver focal lesions provided by the present application can receive and screen multi-dimensional liver CT image data, generate a standardized DICOM data set based on preset layer thickness threshold, resolution threshold and artifact filtering rules, perform two-stage lesion positioning analysis on the standardized DICOM data set, output a lesion positioning atlas containing lesion coordinates, size and average CT value, start an interactive image optimization engine based on the lesion positioning atlas, real-time respond to the window width and window level adjustment instructions and measurement tool operations of the physician, generate a dynamically enhanced lesion visual field and synchronously update the lesion parameters, integrate the visual field data and the clinical rule base to generate a structured report, and output a diagnosis proposal conforming to the DICOM standard through the graphic-text report module, so as to realize the standardization and intelligentization of assisted diagnosis, reduce human error, and improve the precision and efficiency of liver lesion detection. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A hardware structure block diagram of a computer terminal of a CT image assisted detection method for liver focal lesions provided by the embodiment of the present application is shown in the figure.

[0040] Figure 2 A flowchart of a CT image assisted detection method for liver focal lesions provided by the embodiment of the present application is shown in the figure.

[0041] Figure 3 A DICOM image diagram of a lesion contour provided by the embodiment of the present application is shown in the figure.

[0042] Figure 4 A DICOM image diagram without a lesion contour provided by the embodiment of the present application is shown in the figure.

[0043] Figure 5 A structure diagram of a CT image assisted detection system for liver focal lesions provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0044] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be explained as a limitation of the present application.

[0045] The embodiment of the present application first provides a CT image assisted detection method for liver focal lesions, which can be applied to electronic equipment, such as a computer terminal, specifically, a general computer, etc.

[0046] The following will be described in detail by taking a computer terminal as an example. Figure 1 A hardware structure block diagram of a computer terminal of a CT image assisted detection method for liver focal lesions provided by the embodiment of the present application is shown in the figure. Figure 1 As shown in the figure, the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0047] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can make the processor execute any CT image assisted detection method for liver focal lesions.

[0048] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0049] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium, which, when executed by the processor, can make the processor execute any CT image assisted detection method for liver focal lesions.

[0050] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that, Figure 1 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0051] It should be understood that the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0052] Referring to Figure 2 The embodiments of the present application provide a CT image assisted detection method for liver focal lesions, which can include the following steps:

[0053] In S201, multi-dimensional liver CT image data is received and screened, and a standardized DICOM data set is generated based on preset layer thickness threshold, resolution threshold and artifact filtering rules, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold;

[0054] Specifically, the original DICOM stream can be received from the PACS system, the patient age, device model and scanning protocol metadata are synchronously accessed, and the original data cube with space-time labels is generated;

[0055] The system first receives the original DICOM (Digital Imaging and Communications in Medicine) data stream through the standardized interface (e.g. DIMSE protocol) of the Picture Archiving and Communication System (PACS). The data stream contains the original Slice Sequence of liver CT scan, each slice is attached with DICOM Tag, such as Patient ID, Acquisition Time, Series Number, etc. Meanwhile, the system synchronously retrieves the Metadata, such as Patient Age (PA), Device Model (DM, e.g. "GE Revolution CT"), Scan Protocol (SP, e.g. "Liver Triple Phase"), from the Hospital Information System (HIS) or Radiology Information System (RIS). After the integration of these Metadata and spatial information (e.g. Slice Location, Pixel Spacing) in the DICOM stream, a 3D Spatial Index is constructed. The index takes the Slice Sequence Number as the T-Axis, and the pixel coordinate as the X / Y-Axis, forming a Raw Data Cube (RDC) with Spatio-temporal Label (STL). For example, the RDC dimension of a certain patient may be 512x512x200 (widthxheightxnumber of slices), each voxel is associated with the label {PA: 45 years old, DM: "Siemens Somatom Force", SP: "Portal Venous Phase"}.

[0056] The generation of spatiotemporal tags needs to solve the problem of multi-source data alignment. The system adopts a timestamp matching algorithm (TMA): according to the acquisition time stamp (ADT) in the DICOM stream and the study request time (SRT) in the HIS / RIS, the delta time calculation (DTC) is calculated, and if the time difference is less than the preset threshold (such as 5 minutes), it is determined as the same examination sequence. The device model and scanning protocol are mapped to the preset device parameter library (DPL) through the device serial number (DSN) in the DICOM tag, for example, matching the DSN "CT-7890" to the parameter {detector row number: 128 rows, tube voltage: 120 kV}. The patient age is stored in floating point format (such as 45.3 years old) for subsequent adaptive processing. The final generated RDC is a four-dimensional structure (width, height, slice number, metadata channel), and the metadata channel (MC) stores the encoded values of fields such as {PA, DM, SP} (such as PA normalized to 0-1 interval value).

[0057] To ensure data integrity, the system performs redundancy check (RC). If the slice is missing (such as the sequence number is not continuous), the DICOM retransmission request (RR) is automatically triggered; if the metadata conflict (such as the same examination ID corresponds to two age values), the conflict resolution rule (CRR) is started, and the DICOM tag field is preferred. The final output RDC is packaged into a special binary format (BinaryContainer Format, BCF), and the header stores the key-value pairs (KVP) of spatiotemporal tags, and the data body stores the original CT value matrix (CT Value Matrix, CVM). This step provides a standardized input with spatiotemporal consistency for subsequent processing.

[0058] The original data cube is input into the adaptive frequency domain filter, the convolution kernel size is dynamically selected based on the device parameters, the metal artifacts and motion blur are eliminated, and the purified image sequence is output;

[0059] The core of Adaptive Frequency Domain Filter (AFDF) is Dynamic Convolution Kernel (DCK) technology. The system parses the Device Model (DM) of RDC, loads the corresponding Point Spread Function (PSF) parameters from the pre-set Device Characteristic Table (DCT). For example, the device model "Siemens Somatom Force" is mapped to the PSF parameters {FWHM: 0.35 mm, cutoff frequency: 0.8 Nyquist}. Based on the PSF parameters, the initial convolution kernel size (KS) is calculated, the formula is KS = 2×ceil(PSF FWHM / pixel pitch) + 1. If the pixel pitch is 0.6 mm, then KS = 2×ceil(0.35 / 0.6) + 1 = 3×3.

[0060] Frequency domain filtering is performed in three steps:

[0061] Fast Fourier Transform (FFT): Each slice of RDC is converted from the spatial domain to the frequency domain (Frequency Domain), generating a complex spectrum (Complex Spectrum, CS).

[0062] Metal Artifact Suppression (MAS): High-frequency noise (High-frequency Noise, HN) is detected in the frequency domain, which appears as bright spots (Bright Spot, BS) on the edges of the spectrum. According to the device model, select the matching notch filter (Notch Filter, NF), for example, "GE Revolution CT" uses ring notch, while "Philips IQon" uses fan notch, dynamically adjusts the notch radius (Notch Radius, NR) to cover the artifact area.

[0063] Motion Blur Correction (MBC): This uses Wiener Filter (WF) to inversely compensate for blur caused by PSF. The filter strength (FS) is dynamically adjusted based on the patient's age (PA): elderly patients (PA > 60) use FS = 0.8 (strong filtering), while younger patients (PA < 40) use FS = 0.3 (weak filtering) to reduce over-sharpening artifacts (OSA).

[0064] The filtered spectrum is transformed back to the spatial domain using an inverse Fourier transform (IFFT). The system evaluates the purification effect through residual analysis (RA): the root mean square error (RMSE) of the images before and after filtering is calculated. If the RMSE > 15 HU (Henness units), the convolution kernel size is automatically increased (e.g., from 3×3 to 5×5) and the image is re-filtered. The final output is a purified image sequence (PIS), which reduces stripe noise in metal artifact areas (such as around the gallbladder stent) by more than 90% and improves the sharpness of liver edges by 20%-40%.

[0065] Interlayer consistency analysis was performed on the cleaned image sequence, adjacent slices were aligned using a non-rigid registration algorithm, and sub-millimeter volume data were generated by interpolation based on a preset layer thickness threshold.

[0066] Inter-slice Consistency Analysis (ISCA) aims to resolve slice misalignment caused by respiratory motion. The system uses the portal vein as a baseline anatomical landmark to automatically locate the portal vein bifurcation point (BP) in the PIS. Adjacent slices are then aligned using a non-rigid registration algorithm (NRA).

[0067] Feature Extraction: The vascular structure is enhanced using a 3D Hessian matrix to extract the portal vein centerline (CL).

[0068] Deformation Field Modeling: B-spline Free Form Deformation (BFFD) was used to generate a Control Point Grid (CPG) with a Grid Spacing (GS) of 1.5 times the slice thickness (e.g., GS = 7.5 mm for 5 mm slice).

[0069] Similarity Optimization: Mutual Information (MI) was used as the metric to optimize the control point displacement using the L-BFGS algorithm (Limited-memory Broyden-Fletcher-Goldfarb-Shanno) with a Max Iterations (MI) of 100.

[0070] After registration, the Inter-slice Offset Vector (IOV) was calculated. If the Magnitude (MAG) of the IOV > 2 mm (preset tolerance), it was determined to be a Respiratory Motion Artifact (RMA), triggering Slice Reordering (SR). For example, the nth slice needed to be translated along the vector (Δx = 1.2 mm, Δy = 0.8 mm) to match the n-1th slice. The aligned sequence generated a continuous Liver Mask (LM), which filled the Vascular Gap (VG) through Morphological Closing.

[0071] Interpolation was performed according to the Slice Thickness Threshold (STT) (e.g., 1.0 mm):

[0072] If the Original Thickness (OT) ≤ STT (e.g., 0.625 mm), the original data was directly outputted.

[0073] If OT > STT (e.g. 5.0 mm), submillimeter interpolation based on Thin Plate Spline (TPS) is used: with the registered vessel centerline as Anchor Point (AP), Spline Coefficient (SC) is calculated to generate Submillimeter Volumetric Data (SVD). For example, 5 mm slice interpolation to 1 mm slice thickness, 4 new slices are inserted between two slices, and the CT values of the new slices are calculated by weighting the adjacent slices (weight inversely proportional to distance). After interpolation, the stair-step artifact of the liver edge is eliminated, and the error of lesion volume measurement is reduced from 15% to within 5%.

[0074] Patient age and lesion size threshold are encoded as feature vectors, which are input into the graph neural network together with submillimeter volumetric data to generate dynamically weighted standardized DICOM data sets.

[0075] Feature Vector Encoding (FVE) converts non-image data into a format that can be processed by neural networks:

[0076] Patient Age (PA) is encoded into three parts:

[0077] Age Normalization (AN): PA / 100 (e.g. 45 years old → 0.45).

[0078] Age Segment One-hot (ASO): [Child: 0, Adult: 1, Elderly: 0] (45 years old belongs to "Adult").

[0079] Age-related Parameter (ARP): such as Parenchyma Attenuation (PA) set to -5 + 0.1 × PA HU (45 years old → -0.5 HU).

[0080] Lesion Size Threshold (LST) is encoded as:

[0081] Absolute Threshold (AT): such as 5 mm → 0.5 cm.

[0082] Relative Threshold (RT): AT / Max Liver Diameter (MLD).

[0083] Finally, an 8-dimensional feature vector (FV) is generated, such as [0.45, 0, 1, 0, -0.5, 0.5, 12.5, 0.04] (assuming MLD = 12.5 cm).

[0084] The construction process of a graph neural network (GNN):

[0085] Graph construction:

[0086] Node: represents a supervoxel (SV) in the body data, generated by clustering 3x3x3 voxels.

[0087] Edge: connects adjacent supervoxels, and the edge weight (EW) is determined by the CT gradient (CTG) (small gradient, high weight).

[0088] Dynamic weighted fusion (DWF):

[0089] Feature vector FV is mapped to the initial node feature (INF).

[0090] Lesion size threshold LST controls the receptive field (RF) of the convolutional layer: LST < 1 cm, RF = 3 (sensitive to small lesions), LST > 2 cm, RF = 7 (covers large lesions).

[0091] Message passing:

[0092] Neighbor feature (NF) is weighted aggregated (WA) according to the edge weight.

[0093] Patient age PA adjusts the aggregation strength: for elderly patients (PA > 60), smoothing is enhanced (noise reduction), and for young patients, detail preservation is enhanced.

[0094] The network output is the dynamic weight (DW) of each supervoxel, used for weighted fusion of the original data:

[0095] Liver parenchymal region: DW = 1.0 (retain original value);

[0096] Vessel margin: DW = 2.0 (enhanced contrast);

[0097] Suspected small lesion area (LST < 1 cm): DW = 1.5 (improve sensitivity).

[0098] The weighted volume data is reconstructed by DICOM header, and standardized tags (such as window width and window level recommended value WW / WL = 250 / 50) are added to generate a standardized DICOM dataset (SDD). For example, the edge contrast of liver cysts in elderly patients is improved by 30% in SDD, and the detection rate of 1 mm micro-nodules is increased by 25%. One of the standardized DICOM data is shown in Table 1.

[0099] Table 1

[0100]

[0101] This step converts the original CT images from different scanning devices into standardized analysis units through an intelligent data screening mechanism. The system not only considers the physical parameters of the image itself (such as layer thickness and resolution), but also combines patient individual characteristics (age) and device differences (such as noise characteristics of different CT machines) to establish a dynamic data quality control system. The artifact filtering rules use a combination of frequency domain analysis and deep learning to effectively eliminate image interference caused by metal implants and other factors, achieving standardized processing of multi-source heterogeneous image data and providing high-quality input data for subsequent analysis. By dynamically associating clinical parameters, the sensitivity and specificity of subsequent lesion detection can be adjusted adaptively according to the actual situation of the patient, avoiding misjudgment caused by device differences or patient size.

[0102] S202, performing two-stage lesion positioning analysis on the standardized DICOM dataset, first generating a global liver segmentation mask through a cascaded deep learning model, and then extracting a local ROI region based on the mask coordinates for sub-pixel level lesion detection, outputting a lesion positioning atlas containing lesion coordinates, size and average CT value;

[0103] Specifically, the standardized DICOM dataset can be input into a cascaded U-Net model, the first stage uses 3D dilated convolution to extract liver boundary features, and outputs a liver probability heat map;

[0104] The standardized DICOM dataset is used as the input source, which is essentially a three-dimensional body data (voxel resolution up to 0.625 mm x 0.625 mm x 0.5 mm) that has undergone strict quality control. This dataset is first loaded into the first stage of the cascaded U-Net model. The core of this stage is 3D dilated convolution. Unlike traditional convolution, dilated convolution introduces a dilation rate (DR) parameter (for example, DR=2 or DR=4) to insert cavities between the convolution kernel elements, thereby significantly expanding the receptive field without increasing the number of parameters. For example, a 3x3x3 convolution kernel with DR=2 actually covers a range of voxels equivalent to 7x7x7, enabling it to capture liver macro boundary features (such as liver lobe boundaries and portal vein trunk orientation) across multiple slices.

[0105] At the model architecture level, the first stage consists of 4 levels:

[0106] Feature extraction layer: 3 groups of consecutive 3D dilated convolution layers (convolution kernel size 3x3x3, DR=2) are used, each followed by batch normalization (BN) and rectified linear unit (ReLU) activation function. This design can effectively suppress the noise interference (such as low-dose scan noise) commonly found in CT images, while enhancing the contrast boundary between the liver and surrounding tissues (diaphragm, intestinal tube, and kidney).

[0107] Context fusion layer: a spatial pyramid pooling (SPP) module is introduced at the end of the encoding path. This module uses multiple dilated convolution branches with different DR values (such as DR=1, 4, 8) in parallel to extract multi-scale liver features. For example, the DR=1 branch focuses on local texture (smoothness of liver capsule), and the DR=8 branch identifies global morphology (volume proportion of right liver lobe), and finally multi-scale information is fused through feature concatenation.

[0108] Probability mapping layer: the decoding path uses transposed convolution (step size 2x2x2) to gradually upsample the feature map, and the same scale features from the encoding path are fused through jump connection. The final output layer uses a 1x1x1 convolution kernel + Sigmoid activation function to generate a liver probability heat map (Liver Probability Heatmap) with the same size as the input body data. Each voxel value in this heat map is a continuous probability value between 0 and 1 (for example, 0.93 indicates that the probability of this position belonging to the liver is 93%), which directly distinguishes between liver parenchyma (highlighted area) and non-liver tissue (dark area).

[0109] The generation process of the heat map relies on GPU parallel acceleration in real time (such as NVIDIA A100 Tensor Core GPU), and the singleton data reasoning time is controlled within 3.5 seconds. The output result is stored in the form of a floating-point matrix, providing probability basis for subsequent accurate segmentation.

[0110] Topological connectivity analysis is performed on the liver probability heat map, the edge is corrected by the gradient vector flow field, and the sub-voxel level liver mask and the corresponding spatial coordinate tree are generated;

[0111] The liver probability heat map needs to be converted into an accurate liver mask (Liver Mask), that is, a binary label (0 represents background and 1 represents liver). The traditional threshold method (such as fixed threshold 0.5) is easy to cause edge sawtooth or small area missegmentation. This step uses topological connectivity analysis (Topological Connectivity Analysis) combined with gradient vector flow field (Gradient Vector Flow, GVF) to achieve sub-voxel level accuracy:

[0112] Initial segmentation and connected component labeling:

[0113] Apply adaptive thresholding to the heat map (Adaptive Thresholding), and the threshold is dynamically calculated according to the average probability value of the liver parenchyma (such as mean value x 0.7).

[0114] Perform three-dimensional connected component labeling (3D Connected Component Labeling, CCL) on the binary result to label all independent regions. Based on prior knowledge (adult liver volume range 800-1600 cubic centimeters), filter out noise regions with a volume less than 100 cubic centimeters.

[0115] GVF edge optimization:

[0116] Calculate the three-dimensional gradient field (Gradient Field) of the heat map to generate the initial edge force (Edge Force).

[0117] Construct the gradient vector flow field (GVF): by iteratively solving partial differential equations, the initial edge force is diffused to the whole space to form a smooth and boundary-oriented vector field. GVF parameter settings include viscosity coefficient (Viscosity Coefficient, μ=0.15) and iteration count (Iteration Count, IC=100 times), to ensure that the force field can penetrate deep concave structures (such as hepatic portal vascular space).

[0118] Apply the GVF force field to the initial segmentation boundary to drive the contour line to move towards the true liver edge (such as correcting the missegmented gallbladder fossa region).

[0119] Sub-voxel mask generation:

[0120] The optimized continuous surface is converted into a triangular mesh grid using the Marching Cubes Algorithm, with a vertex precision of 0.1 voxel (i.e., sub-voxel level).

[0121] The mesh is voxelized to generate the final liver mask, and a Spatial Coordinate Tree is constructed:

[0122] Root node: stores the overall bounding box coordinates of the liver (e.g., start point [X_min, Y_min, Z_min] and end point [X_max, Y_max, Z_max]).

[0123] Child nodes: divided into 8 sub-regions according to liver segments (Couinaud segmentation method), each child node stores the coordinate range and volume proportion of the corresponding sub-mask (e.g., liver S8 segment accounts for 18%).

[0124] This mask can be directly used for three-dimensional liver volume measurement, and the spatial coordinate tree provides a spatial indexing basis for subsequent lesion positioning.

[0125] Based on the spatial coordinate tree, a local region of interest (ROI) is defined, and a deformable convolution network is used to adaptively adjust the ROI size, outputting a set of lesion candidate blocks.

[0126] The spatial coordinate tree of the liver mask serves as a navigation framework, guiding the lesion detection to focus on the liver parenchyma:

[0127] Intelligent ROI definition:

[0128] Traverse the child nodes (liver segment regions) of the spatial coordinate tree, and generate an initial local region of interest (ROI) centered on the bounding box of each child node, with a 15mm safety margin (to avoid edge lesions being truncated).

[0129] Adjust the ROI size according to the volume of the child node: for sub-regions with a volume greater than 150 cubic centimeters (e.g., right posterior lobe), set the ROI size to 120mm x 120mm x 80mm; for regions with a volume less than 80 cubic centimeters (e.g., caudate lobe), reduce the size to 60mm x 60mm x 40mm.

[0130] Deformable convolution network processing:

[0131] The volume data of each ROI region is cropped and input into a deformable convolutional network (DCN). The core of the DCN is to add an offset prediction layer (Offset Prediction Layer) before the standard convolutional layer:

[0132] This layer predicts the three-dimensional offset (Δx, Δy, Δz) of each sampling point through 1×1×1 convolution, and the offset range allows ±3 voxels (adapt to lesion morphology variation).

[0133] Example: For liver cancer lesions, the offset will focus the convolution kernel on the "fast-in fast-out" enhanced area; for cysts, it will avoid the internal homogeneous area and enhance the thin-walled features.

[0134] The network backbone uses a lightweight 3D ResNet-18 architecture, and the last layer is replaced with a region proposal network (RPN) that outputs lesion bounding boxes.

[0135] Lesion candidate block set generation:

[0136] The RPN generates about 300 candidate boxes for each ROI, and redundant boxes are filtered out through non-maximum suppression (NMS, overlap threshold 0.5).

[0137] Candidate boxes with a confidence score greater than 0.85 are retained, and their corresponding voxel blocks (such as 32×32×32 voxels) are extracted to form a lesion candidate block set.

[0138] Each block is accompanied by spatial coordinates (center point [X_c, Y_c, Z_c] in DICOM coordinate system) and size estimates (long diameter L_dia, short diameter S_dia).

[0139] Phase congruency edge detection is applied to the lesion candidate block set, and the lesion center offset is calculated using the Hessian matrix to generate sub-pixel positioning coordinates;

[0140] The candidate block needs to be further precisely positioned to the sub-pixel level:

[0141] Phase congruency edge detection (Phase Congruency Edge Detection): the transverse, sagittal, and coronal sections of each candidate block are processed separately;

[0142] Phase Congruency Value (PCV) is calculated for each pixel using Log-Gabor filter bank (4 orientations x 5 scales). PCV ∈ [0, 1], the closer to 1, the higher the probability of the point being on the real edge (not affected by brightness changes).

[0143] The pixel points with PCV > 0.7 are extracted to form the initial edge point set (such as the "mosaic sign" edge of liver cancer).

[0144] Hessian matrix analysis:

[0145] The Hessian matrix (second derivative matrix) at each edge point is calculated, with a matrix dimension of 3 x 3 (including XX, XY, XZ, YY, YZ, ZZ direction partial derivatives).

[0146] Eigenvalue decomposition is performed on the Hessian matrix to obtain eigenvalues λ1, λ2, λ3 (sorted by absolute value |λ1| ≥ |λ2| ≥ |λ3|).

[0147] Based on the combination of eigenvalues, the structure type is determined: if λ1 << 0 and λ2 ≈ λ3 ≈ 0, the point is a candidate point for the center of a spherical structure (such as a cyst); if λ1 << 0, λ2 << 0 and λ3 ≈ 0, it is an interference point for tubular structures (such as blood vessels), which needs to be excluded.

[0148] Sub-pixel center positioning:

[0149] For the screened feature points (such as points with λ1 < -0.1), a 5 x 5 x 5 voxel neighborhood is taken with the point as the center.

[0150] A quadratic surface function is fitted in the neighborhood, and the coefficients are solved by least squares method. The point where the function gradient is zero is the sub-pixel level lesion center (X_sub, Y_sub, Z_sub).

[0151] The final coordinate accuracy reaches 0.01 voxel (such as actual physical accuracy 0.006 millimeters), and is written into the sub-pixel positioning coordinate list.

[0152] Fusion of sub-pixel positioning coordinates and CT value histogram statistics, through radial basis function interpolation to generate lesion localization atlas.

[0153] Integration of geometric and density information to generate comprehensive lesion localization atlas (Lesion Localization Atlas):

[0154] CT value histogram statistics:

[0155] Take a spherical region of interest (diameter = 1.2 times the initial estimated size of the lesion) centered at the sub-pixel coordinates of the lesion.

[0156] Extract the CT values (in Hounsfield Units, HU) of all voxels within the region of interest, and generate a CT histogram.

[0157] Calculate key statistical quantities:

[0158] Mean CT: The mean of the histogram (e.g., ≈10 HU for cysts, ≈45 HU for liver cancer).

[0159] Peak CT: The value corresponding to the highest peak of the histogram (reflecting the main density).

[0160] Heterogeneity Index (HI): (75th percentile - 25th percentile) / median (e.g., metastases HI > 0.3).

[0161] Radial Basis Function Interpolation (RBFI):

[0162] Objective: Fuse discrete lesion point information into a continuous spatial distribution map.

[0163] Algorithm Core: Use the sub-pixel coordinates of each lesion as control points to define the interpolation function: F(X) = Σ [w_i· φ(||X - X_i||)]. Where φ is the radial basis function (choose Gaussian kernel φ(r) = exp(-(εr) 2 ), kernel width parameter ε = 0.7), and the weight w_i is assigned by normalizing the average CT value of the lesion.

[0164] Calculate F(X) value voxel by voxel in the liver mask space to generate the CT value distribution potential field.

[0165] Map synthesis and output:

[0166] The map contains three layers of data: Geometric layer: sub-pixel coordinates (X_sub, Y_sub, Z_sub), lesion size (long diameter L_dia, short diameter S_dia, obtained based on histogram distribution fitting ellipse);

[0167] Density layer: mean CT value, peak CT value, heterogeneity index HI;

[0168] Spatial distribution layer: CT value distribution potential field (used to visualize the density gradient of the lesion).

[0169] The output format is DICOM-SEG structured object, which supports direct access by PACS system. The atlas can automatically label the lesions (such as "L1: Liver S6 segment, 8mm x 6mm, MeanCT = 52HU"), and provide data basis for subsequent interactive diagnosis.

[0170] The detection strategy from coarse to fine is adopted, first the three-dimensional convolution network is used to accurately outline the anatomical boundary of the liver, and then the micro-lesion detection is carried out in the limited range. In particular, the sub-pixel level positioning technology can identify the micro-lesions smaller than the size of the voxel by analyzing the gradient distribution and texture features of the CT value, and improve the detection accuracy, which significantly improves the detection rate of small lesions. The multi-parameter output of the positioning atlas provides a quantitative basis for subsequent diagnosis, and overcomes the limitations of the traditional method relying on the visual estimation of the physician.

[0171] S203, based on the lesion positioning atlas, starting the interactive image optimization engine, real-time responding to the window width and window level adjustment instructions and measurement tool operation of the physician, generating dynamic enhanced lesion visual field and synchronously updating the lesion parameters;

[0172] Specifically, the window width and window level adjustment signals and the measurement tool trajectory data of the physician can be received in real time, the operation type is analyzed by gesture intention recognition algorithm, and the window level and window width parameter vector and the measurement coordinate sequence are generated;

[0173] When the physician operates in the medical image workstation, the interactive image optimization engine continuously monitors the input events through the graphical user interface (Graphical User Interface, GUI). The window width and window level adjustment signal (Window Width / Level Adjustment Signal, WWLAS) is derived from the physician's operation on the workstation slider or shortcut key. For example, when the physician drags the "window width" slider to 400 Hounsfield Unit (Hounsfield Unit, HU) and the "window level" slider to 60 HU, the system captures these two values as the original input signal. At the same time, the measurement tool trajectory data (Measurement Tool Trajectory Data, MTTD) is derived from the geometric figure drawn by the physician on the screen image using the mouse or stylus, such as the continuous screen coordinate point sequence (for example, 50 coordinate points per second) generated when drawing a straight line to measure the diameter of the lesion or drawing a polygon to outline the lesion area. These original input data are encapsulated into data packets with timestamp (Timestamp, TS) and device identifier (Device ID, DID), and transmitted to the analysis module through the high-speed bus.

[0174] The core of the analysis module is the Gesture Intent Recognition Algorithm (GIRA). This algorithm uses time-series pattern recognition techniques to analyze the input stream:

[0175] For the slider operation, GIRA monitors the change rate (CR, unit: HU / s) and pause interval (PI). If CR is greater than 50 HU / s and PI is less than 0.3 seconds, it is determined to be "fast coarse adjustment"; if CR is less than 10 HU / s and PI is greater than 1 second, it is determined to be "fine fine adjustment".

[0176] For trajectory data, GIRA extracts the curvature (CV), closure (CL), and point density (PD) of the trajectory. For example, a trajectory composed of 20 coordinate points, with a curvature less than 0.1 and a closure greater than 95%, will be classified as "polygon measurement"; a straight line trajectory defined by two endpoints with a length greater than 30 pixels will be classified as "linear measurement".

[0177] Finally, GIRA outputs structured instructions: packing the window width value (such as 400 HU) and window level value (such as 60 HU) into a window level / width parameter vector (WLWPV, format: [window width value, window level value]); serializing the screen coordinates of the measurement trajectory (such as start point [X1, Y1], end point [X2, Y2]) into a measurement coordinate sequence (MCS, format: [operation type, coordinate point 1, coordinate point 2,...]).

[0178] To ensure real-time performance, the system uses a double buffering mechanism (DBM). The instructions processed in the current processing cycle (PC, cycle: 16 milliseconds) are temporarily stored in the high-speed cache (HSC), while the WLWPV and MCS analyzed in the previous cycle are transmitted to the downstream module through direct memory access (DMA). At the same time, the operation log (OL) records the physician's operation habits (such as commonly used window width and window level combinations, measurement tool usage frequency) to optimize the decision threshold of GIRA (such as dynamically adjusting the judgment boundaries of CR and PI). This step accurately converts the physician's physical operation into digital instructions that can be processed by the machine, laying the foundation for subsequent image rendering.

[0179] Based on the spatial coordinate tree of the lesion localization atlas, the corresponding ROI volume data is preloaded in the GPU display memory, and the ray casting algorithm is dynamically adjusted combined with the window level and window width parameter vector to output real-time volume rendering images.

[0180] The system calls the spatial coordinate tree (SCT) in the lesion localization atlas (LLA). SCT is a hierarchical data structure (HDS), for example, with the liver portal vein as the root node (RN), and the leaf node (LN) storing the three-dimensional voxel coordinates of the lesion (such as [120, 85, 45]). When receiving the WLWPV (such as [400, 60]), the engine locates the local region of interest (ROI) where the target lesion is located according to the SCT. For example, if the lesion size is 15 mm, the ROI range is defined as a cube region (CR) that expands 30 mm outside the center of the lesion. At this time, the system queries the pre-established volume data index table (VDIT) to locate the storage location (such as data block offset Offset=0x5A3F) of the ROI in the standardized DICOM dataset.

[0181] Through the calculation of the compute unified device architecture (CUDA) interface, the original CT volume data corresponding to the ROI (such as a 512x512x30 voxel matrix) is transmitted asynchronously to the global memory (GM) of the graphics processing unit (GPU). At the same time, the GPU kernel (Kernel) of the ray casting algorithm (RCA) is started:

[0182] Transfer function remapping: dynamically generate the transfer function (TF) according to the WLWPV. For example, the window width of 400 HU corresponds to the gray scale range of -140 HU to 260 HU (window level 60±200 HU), which is linearly mapped to 0-255 gray levels, and the voxels outside the range are set to all black / white.

[0183] Light step optimization: Along each projection ray, sample at 0.5mm step size (SS), and calculate the CT value of sampling point (SP) by trilinear interpolation (TLI).

[0184] Color synthesis: Accumulate the color of SP by alpha blending (AB) formula, where the opacity (OP) of lesion area (CT value > 40 HU) is set to 0.8, and the OP of normal liver tissue (CT value 30-50 HU) is set to 0.3.

[0185] The whole process is implemented under the GPU parallel computing architecture, with a response time of less than 50 milliseconds, and the output resolution of real-time volume rendered image (RTVRI) is 1920x1080.

[0186] To reduce the delay, the system uses a smart preloading strategy (SPS):

[0187] When the physician's mouse hovers over a lesion marker, the ROI data of the adjacent three layers (about 90 megabytes) are automatically preloaded.

[0188] A memory priority queue (MPQ) is established, and the ROI data of the current focus area (FA) is kept in the high-speed texture memory (TM), while the non-focus data is transferred to the low-speed video memory.

[0189] This step converts static CT data into dynamic visual images, allowing the physician's window width and window level adjustments to be immediately reflected on the three-dimensional reconstructed image.

[0190] Anisotropic diffusion filtering is superimposed on the real-time volume rendered image, the lesion edge contrast is adaptively enhanced according to the window width parameters, and the labeled trajectory of the measurement coordinate sequence is synchronously embedded to generate an enhanced labeled image.

[0191] Post-processing enhancement is performed on the RTVRI:

[0192] Anisotropic Diffusion Filter (ADF): Adjust diffusion strength according to Image Gradient (IG) direction. Set initial Conduction Coefficient (CC) as 0.25, Iteration Count (IC) as 5 times. In lesion edge area (gradient value > 30), reduce diffusion strength (CC = 0.1) to preserve details; in homogeneous area (gradient value < 10), increase diffusion strength (CC = 0.4) to suppress noise.

[0193] Window Width Driven Contrast Enhancement: When window width value (e.g. 400 HU) is less than soft tissue window standard value (350 HU), start Edge Sharpening (ES) algorithm, use Laplacian Operator (LO) to enhance high frequency components (enhancement coefficient EC = 1.8); when window width is greater than 500 HU (bone window range), turn off sharpening to avoid artifacts. The lesion boundary clarity of the processed image is improved by about 40% (measured by boundary gradient modulus).

[0194] Label trajectory embedding is divided into three steps:

[0195] Coordinate conversion: Convert screen coordinates (e.g. [X1, Y1]) in MCS to three-dimensional space coordinates through Inverse Perspective Projection (IPP). For example, screen point (320, 240) corresponds to world coordinates (15.2, -28.7, 45.3).

[0196] Geometry rendering:

[0197] Linear measurement: Draw a red solid line (line width 2 pixels) in three-dimensional space, and label the distance value (e.g. "17.3mm") at the end.

[0198] Polygon measurement: Fill the semi-transparent green area (transparency Alpha = 0.4) and draw the boundary (line width 1.5 pixels).

[0199] Depth buffer processing: Enable Z-Buffering (ZB) to ensure that the annotation object and volume rendering image are correctly occluded, and automatically semi-transparent (transparency increases to 0.7) when the annotation object is located behind the organ.

[0200] When generating Enhanced Annotation Image (EAI), use Hybrid Rendering Pipeline (HRP):

[0201] The volume rendering result is stored in the color attachment 0 (CA0) of a frame buffer object (FBO).

[0202] The label rendering is to color attachment 1 (CA1) and the stencil test (ST) is enabled.

[0203] CA0 and CA1 are mixed by a fragment shader (FS) according to the depth value, and finally output to the display buffer (DB).

[0204] The image not only retains the anatomical details of the medical image, but also intuitively displays the measurement intention of the physician.

[0205] The lesion size deviation and CT value fluctuation are calculated by comparing the measurement coordinate sequence with the subpixel localization coordinates of the lesion positioning atlas, and the lesion dynamic parameter table is updated in real time.

[0206] The system starts the coordinate comparison engine (CCE):

[0207] Subpixel coordinate alignment: Extract the subpixel localization coordinates (SPLC, precision 0.1 voxel, such as [120.3, 85.7, 45.2]) of the lesion in LLA. The three-dimensional coordinates of MCS (such as the polygon vertex set measured by the physician) are registered to the SPLC coordinate system through affine transformation (AT).

[0208] Spatial consistency check: Calculate the Hausdorff distance (HD) between the physician's labeled point set and the AI labeled point set. If HD is greater than 2 mm, trigger the confidence warning flag (CWF) and prompt with a yellow flashing border on the interface.

[0209] Quantitative calculation of core indicators:

[0210] Lesion size deviation (LSD):

[0211] The absolute deviation (|18.5-17.2|=1.3mm) and relative deviation (1.3 / 17.2x100%=7.6%) of the physician's measured diameter (such as 18.5mm) and the AI measured value (such as 17.2mm).

[0212] For polygon region, calculate the Overlap Ratio (OR) = intersection area / union area x 100%.

[0213] CT Value Fluctuation (CTVF):

[0214] In the physician-labeled region, calculate the Standard Deviation (SD) of CT value. For example, the SD changes from 8.5 HU by AI analysis to 12.3 HU by physician-labeled region, the Fluctuation Rate (FR) = (12.3-8.5) / 8.5 x 100% = 44.7%.

[0215] The calculation results are packaged as Deviation Vector (DV), such as [dimension deviation value, fluctuation rate].

[0216] Dynamic parameter table updating mechanism:

[0217] The Lesion Parameter Table (LPT) is stored in the In-Memory Database (IMD), and the fields include lesion ID, coordinates, AI measured size, physician measured size, CT value mean, CT value standard deviation, etc.

[0218] When a new DV is received:

[0219] If LSD < 5% and CTVF < 15%, only update the "physician measured size" field.

[0220] If LSD > 10% or CTVF > 30%, add an Abnormal Flag (AF) in the table, and trigger the Re-detection Process (RDP) at the same time.

[0221] The updated LPT is pushed to the interface component through the Publish-Subscribe Pattern (PSP), and the physician can view the color-coded parameters in real time (deviation < 5% shows green, > 10% shows red).

[0222] Fusion of enhanced labeled images and lesion dynamic parameter table, through multi-planar reformatting technology to generate dynamic enhanced lesion visual field containing parameter superimposed layer.

[0223] Image-data fusion adopts Layered Compositing Architecture (LCA):

[0224] Base Layer (BL): EAI (volume rendering image and annotation track).

[0225] Parameter Overlay Layer (POL): Information Panel (IP) is generated from key data extracted from LPT. For example, a semi-transparent panel (transparency 0.7) is drawn at the lower right corner of the image, showing:

[0226] Lesion 07: Coordinates (120, 85, 45);

[0227] AI measurement: Diameter 17.2 mm | Average CT value 56 HU;

[0228] Physician measurement: Diameter 18.5 mm (↑ 7.6%) | Current CT-SD 12.3 HU (↑ 44.7%).

[0229] Navigation Marker Layer (NML): In the three-dimensional view, the current operating lesion position is marked with a flashing sphere (diameter 5 pixels).

[0230] Multiplanar Reformation (MPR) realizes multi-view cooperation:

[0231] Main View (MV): Three-dimensional volume rendering view, showing the fused BL+POL+NML.

[0232] Auxiliary Views (AV):

[0233] Axial View (AXV): Cross-section through the center of the lesion.

[0234] Coronal View (CV): Section along the front-to-back direction of the human body.

[0235] Sagittal View (SV): Section along the left-to-right direction of the human body.

[0236] View Linkage (VL): When the physician scrolls the slices in any view of MPR, other views are automatically synchronized to the same anatomical position. When the lesion is drawn in AXV, MV is updated in real time with three-dimensional annotations.

[0237] Finally, the dynamically enhanced lesion visualization domain (DELVD) is generated:

[0238] Dynamism: When the physician adjusts the window width to "liver window" (window width 150 HU / window level 70 HU), the system completes the following within 300 milliseconds: recalculates the transfer function; performs GPU ray casting; updates the ADF filter parameters; and refreshes the CT value statistics in the LPT.

[0239] Enhancement: After the lesion edge is sharpened by the ADF, the boundary contrast is improved to 1.5 times that of the original image (measured by the edge signal-to-noise ratio ESNR).

[0240] Interactivity: The physician clicks the "diameter deviation value" in the parameter table, and the system automatically highlights the difference between the AI and the manually annotated area in the MPR view (the AI area is blue outlined, and the manually annotated area is red outlined).

[0241] The visualization domain seamlessly integrates image data, quantitative parameters, and interactive operations to provide decision support for physicians.

[0242] The engine uses GPU-accelerated real-time rendering technology. When the physician adjusts the window width and window level, the system can immediately recalculate the volume rendering effect and highlight the lesion feature area. The interactive data of the measurement tool triggers the background parameter recalculation module, ensuring that the image display and quantitative data are always synchronized, achieving the "what you see is what you get" interactive experience in the diagnosis process. The physician can quickly verify the nature of the suspicious lesion through real-time feedback. The dynamic parameter update mechanism avoids the tedious operation of repeatedly switching measurement tools in the traditional workflow.

[0243] S204, the structured report is generated by integrating the visualization domain data and the clinical rule base, the image findings field is automatically filled by checking the lesion list, and the diagnosis proposal book meeting the DICOM standard is output through the graphic-text report module, wherein the diagnosis proposal book embeds the lesion position navigation link and the measurement error calibration.

[0244] Specifically, the quantitative data in the lesion dynamic parameter table can be parsed, and the feature matching is performed with the clinical rule base to generate a structured lesion data unit;

[0245] When the interactive image optimization engine generates the lesion dynamic parameter table (LDPT), the table records the key quantitative indicators in the physician's operation process in real time, such as the long axis diameter (LAD), the short axis diameter (SAD), the mean CT value (MCTV), the arterial enhancement value (AEV), and the portal enhancement value (PEV). The system starts the structured data generation module (SDGM), which first cleans the LDPT: removes abnormal fluctuation values (such as CT value mutation exceeding 50 HU) caused by respiratory motion, and uses the sliding window averaging (SWA) method to smooth the time series data. The cleaned parameter set is converted into a feature vector (FV), and the vector dimension includes lesion morphological features (such as circularity = 4π×area / perimeter 2 ), density features (such as the difference between arterial and portal CT values ΔCT = PEV-AEV), and dynamic change features (such as the enhancement rate Enhancement Rate = ΔCT / time difference).

[0246] The feature vector is input into the clinical rule base (CRB) for multi-level matching. The CRB uses a knowledge graph architecture (KGA) and includes three layers of logic:

[0247] The basic rule layer stores international guideline standards (such as LI-RADS classification), such as "arterial phase enhancement > 20 HU and portal phase clearance" corresponding to the malignant sign rule LR-5.

[0248] The expert experience layer integrates historical diagnosis paths in the hospital, such as "age > 60 years old + lesion calcification + capsule retraction" combination triggering biliary cell carcinoma warning.

[0249] The dynamic update layer regularly updates rare disease features (such as the "fast-in slow-out" enhancement mode of liver purpura) from multi-center data through federated learning (FL).

[0250] The matching process adopts a Graph Neural Network Inference Engine (GNN-IE): maps the feature vector to a knowledge graph node, and calculates the similarity score (SS) with the rule node through a Neighborhood Aggregation Algorithm (NAA). When the SS exceeds the threshold value of 0.85 (configurable parameter), it is determined that the feature matches the rule successfully, and the rule ID with a confidence label (CL) is output.

[0251] The matching result is packaged into a structured lesion data unit (SLDU). Each SLDU contains fixed fields:

[0252] Basic attributes: lesion ID, spatial coordinates (DICOM coordinate system XYZ value), size (LAD / SAD, unit: millimeter mm).

[0253] Image features: enhancement mode encoding (such as "arterial phase high enhancement-portal phase equal enhancement" encoded as AEH_PEE), calcification flag (CF, Boolean value).

[0254] Rule association: triggered rule ID list (such as ["LI-RADS-5", "HCC_Rule_2023"]), confidence vector (such as [0.92, 0.78]).

[0255] SLDU is stored in JSON-LD format (JavaScript Object Notation for Linked Data), ensuring semantic association with clinical terminology systems (such as SNOMED CT code "38738009" representing hepatocellular carcinoma).

[0256] Mapping the physician's checked lesion list to the structured lesion data unit, driving natural language generation through the medical knowledge graph, and outputting structured diagnosis text;

[0257] The physician checks the selected lesions list (SLL) by checkboxes on the operation interface, for example, checking the lesions numbered "L002" and "L005". The system activates the semantic mapper (SM), which has a built-in bidirectional index table (BIT): the forward index maps the lesion ID (e.g., "L002") to the corresponding SLDU database address, and the reverse index associates the anatomical location of the lesion (e.g., "liver S8 segment" corresponds to the Segment 8 code of the Couinaud partition). The mapping process performs conflict detection (CD): if the selected lesions have spatial overlap (e.g., center distance < 5 mm), an alert is triggered to prompt the physician to review. After successful mapping, the system sorts the SLDU sequence by priority (e.g., high confidence malignant signs are placed first).

[0258] The sorted SLDU sequence is input into the medical knowledge graph-driven natural language generator (MKG-NLG). The core of this engine is a three-layer architecture:

[0259] Entity linking layer: links SLDU fields to unified medical language system (UMLS) concepts, such as "high enhancement in arterial phase" linked to concept C0452539.

[0260] Template selection layer: matches predefined report template library (RTL) according to rule ID. For example, LI-RADS-5 level lesion automatically selects template "liver S{segment} seen {size} mm nodule, obvious enhancement in arterial phase, portal venous phase clearance, consistent with LR-5 class".

[0261] Parameter filling layer: dynamically inserts quantitative data using conditional random field (CRF) model. For example, "{size}" is replaced by the actual value "22.3 mm", and an error range is appended after the size (e.g., "±1.2 mm", derived from measurement error calibration).

[0262] The NLG engine outputs structured diagnostic text (SDT), which has the following characteristics:

[0263] Paragraph structuring: segmented by "imaging findings-diagnostic opinion-recommendations", consistent with ACR (American College of Radiology) reporting standards.

[0264] Semantic Standardization: RadLex terms are used compulsively (e.g. "early enhancement" is replaced by "high enhancement in arterial phase").

[0265] Dynamic modifiers: Confidence Level (CL) is added to the output, e.g. "clearly shown" for CL>0.9, "considered likely" for CL between 0.7 and 0.9. The final SDT is stored in DICOM SR (Structured Report) object format, containing both text content and metadata association pointers.

[0266] DICOM spatial coordinate system is generated based on the spatial coordinate tree of the lesion localization atlas, and sub-pixel positioning coordinates are converted into DICOM-RT structure sets to create a lesion spatial navigation engine.

[0267] The spatial coordinate tree (SCT) in the lesion localization atlas (LLA) is essentially an octree index (OI) in three-dimensional space. The system starts the DICOM spatial coordinate converter (DSCC), first reads the key parameters in the DICOM file header: image position (IP, stored as a floating-point array [IP_x, IP_y, IP_z]), pixel spacing (PS, such as [0.625, 0.625] millimeters mm), slice thickness (ST, such as 1.0 millimeters mm). The sub-pixel coordinates (such as X=152.34px, Y=98.21px, Z=45 layers) are converted into millimeter coordinates in the DICOM absolute coordinate system (DACS) through the affine transformation matrix (ATM).

[0268] The converted millimeter coordinates are input into the DICOM-RT structure set generator (DRSSG). Each lesion is modeled as a closed surface:

[0269] Contour generation: Centered on sub-pixel positioning coordinates, generate Elliptical Contour (EC) according to size parameters (LAD / SAD). For example, a lesion with major axis 22.3 mm and minor axis 18.1 mm, generate an ellipse with semi-major axis 11.15 mm and semi-minor axis 9.05 mm in XY plane.

[0270] Inter-slice interpolation: Connect adjacent slice contours along Z-axis direction using Thin Plate Spline Interpolation (TPSI) to form a continuous three-dimensional surface.

[0271] Structure set encoding: Define ROI Number (ROIN) and ROI name (e.g. "Liver_Lesion_01") according to DICOM RT Structure Set standard, and write into reference coordinate system UID (Unique Identifier).

[0272] Construction of Lesion Spatial Navigation Engine (LSNE) driven by structure set data:

[0273] Spatial index construction: Import DICOM-RT structure set into three-dimensional rendering engine (e.g. VTK or Three.js), establish spatial R-Tree Index (RTI), support fast distance query (e.g. find blood vessels within 5 mm range).

[0274] Interaction protocol encapsulation: Develop Navigation Protocol (NP) based on WebSocket, define instruction set such as "jump to lesion" (instruction code 0x01 + lesion ROIN) and "display adjacent structures" (instruction code 0x02 + distance threshold).

[0275] Visualization interface: Embed Navigation Panel (NP) in PACS (Picture Archiving and Communication System) browser, support automatic positioning to corresponding slice and highlight contour by clicking lesion name.

[0276] Combine structured diagnostic text with Lesion Spatial Navigation Engine, add error heat map of coordinate sequence measurement, generate diagnostic report with error calibration.

[0277] The system call diagnostic report synthesizer (DRS) first fuses the structured diagnostic text (SDT) and the lesion spatial navigation engine (LSNE):

[0278] Text embedding navigation link: Insert a hyperlink (Hyperlink) after the "lesion description" field of the SDT, with a custom navigation protocol (e.g., nav: / / lesion?id=L002) as its URL protocol.

[0279] Spatial data binding: Establish a mapping table (MT) between the ROIN of the DICOM-RT structure set and the lesion ID in the SDT, ensuring that clicking on the text link can trigger the spatial engine to jump.

[0280] Context-aware rendering: When the physician browses a lesion in the PACS, automatically highlight the corresponding text paragraph in the report sidebar, achieving image-text interlock (ITI).

[0281] Error heatmap (EH) generation based on historical measurement data:

[0282] Error source analysis: Collect the physician's measurement coordinate sequence (MCS, such as the manually drawn lesion boundary point set) and compare it with the algorithm-generated sub-pixel localization coordinates (SPLC). Calculate two types of errors:

[0283] Boundary error (BE): The Hausdorff distance (HD) between the MCS points and the algorithm's contour.

[0284] Size error (SE): The absolute difference between the manually measured length and the algorithm's length.

[0285] Heat model construction: Establish a two-dimensional error field (EF) centered on the lesion, and use Kriging interpolation (KI) to convert discrete error values into continuous distribution. Color according to error level: blue (BE < 1 millimeter mm), yellow (1 millimeter mm ≤ BE < 3 millimeter mm), red (BE ≥ 3 millimeter mm).

[0286] Dynamic calibration: Overlay semi-transparent error heat map on DICOM images and annotate statistical values (e.g. "Max Boundary Error: 2.4mm").

[0287] Final Error-Calibrated Diagnostic Report (ECDR) generation:

[0288] Format compliance: Comply with DICOM SR standard, encapsulate SDT text, DICOM-RT structured set, error heat map into a single DICOM file, e.g. SOP Class UID=1.2.840.10008.5.1.4.1.1.88.33.

[0289] Interactive function integration:

[0290] Click "Lesion 1" text → automatically jump to the corresponding CT slice and highlight the contour.

[0291] Hover over error heat map → display the historical error distribution box plot of this area.

[0292] Multi-modal output: Simultaneously generate PDF version of the report (with clickable navigation links) and HL7 message (trigger LIS system alert).

[0293] Report sent to PACS archive through DICOM Storage SCP service, completing the whole process closed loop.

[0294] The system will detect the results and match them with clinical guidelines (such as LI-RADS standards) to automatically generate a structured description containing key diagnostic elements. The specially designed navigation links allow direct jumping to the lesion corresponding section, while the error calibration intuitively displays the confidence interval of the measurement results, greatly improving the report writing efficiency while ensuring compliance with clinical standards. The navigation function makes the review process more efficient. Error calibration helps physicians judge the reliability of automated detection results, providing transparency and security for clinical decision-making.

[0295] It can be seen that the multi-dimensional liver CT image data is received and screened, the standardized DICOM data set is generated based on the preset layer thickness threshold, resolution threshold and artifact filtering rule, the two-stage lesion positioning analysis is performed on the standardized DICOM data set, and the lesion positioning atlas containing the lesion coordinates, size and average CT value is output; the interactive image optimization engine is started based on the lesion positioning atlas, the window width and window level adjustment instructions and the measurement tool operation of the doctor are responded in real time, the dynamic enhanced lesion visual field is generated, and the lesion parameters are updated synchronously; the structured report is generated by integrating the visual field data and the clinical rule base, the image findings field is automatically filled by checking the lesion list, and the diagnosis proposal conforming to the DICOM standard is output by the graphic-text report module, so that the standardization and intelligentization of auxiliary diagnosis can be realized, the human error is reduced, and the precision and efficiency of liver lesion detection are improved.

[0296] Another embodiment of the present application provides a CT image auxiliary detection system for liver focal lesions, referring to Figure 5 , the system can include:

[0297] The receiving module 501 is configured to receive and screen multi-dimensional liver CT image data, and generate a standardized DICOM data set based on a preset layer thickness threshold, a resolution threshold and an artifact filtering rule, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold;

[0298] The analysis module 502 is configured to perform two-stage lesion positioning analysis on the standardized DICOM data set, first generate a global liver segmentation mask through a cascaded deep learning model, then extract a local ROI region based on the mask coordinates for sub-pixel level lesion detection, and output a lesion positioning atlas containing lesion coordinates, size and average CT value;

[0299] The generation module 503 is configured to start an interactive image optimization engine based on the lesion positioning atlas, respond to the window width and window level adjustment instructions and the measurement tool operation of the doctor in real time, generate a dynamically enhanced lesion visual field and synchronously update the lesion parameters;

[0300] The output module 504 is configured to integrate the visual field data and the clinical rule base to generate a structured report, automatically fill the image findings field by checking the lesion list, and output a diagnosis proposal conforming to the DICOM standard by the graphic-text report module, wherein the diagnosis proposal is embedded with a lesion position navigation link and a measurement error calibration.

[0301] The embodiment of the present application also provides a storage medium, and the storage medium stores a computer program, wherein the computer program is set to execute the steps in any of the above method embodiments when running.

[0302] Specifically, in the embodiment, the storage medium can be configured to store a computer program for executing the following steps:

[0303] S201, receiving and screening multi-dimensional liver CT image data, generating a standardized DICOM data set based on preset layer thickness threshold, resolution threshold and artifact filtering rules, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold;

[0304] S202, performing two-stage lesion positioning analysis on the standardized DICOM data set, first generating a global liver segmentation mask through a cascaded deep learning model, then extracting a local ROI region based on the mask coordinates for sub-pixel level lesion detection, and outputting a lesion positioning atlas containing lesion coordinates, size and average CT value;

[0305] S203, starting an interactive image optimization engine based on the lesion positioning atlas, responding to the window width and window level adjustment instructions of the physician and the measurement tool operation in real time, generating a dynamically enhanced lesion visual field and synchronously updating the lesion parameters;

[0306] S204, integrating the visual field data and the clinical rule base to generate a structured report, automatically filling the image findings field by checking the lesion list, and synchronously outputting a diagnosis proposal conforming to the DICOM standard through the graphic-text report module, wherein the diagnosis proposal is embedded with a lesion position navigation link and a measurement error calibration.

[0307] The embodiment of the application also provides an electronic device, including a memory and a processor, the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the method embodiments.

[0308] Specifically, the electronic device can further include a transmission device and an input-output device, wherein the transmission device is connected to the processor, and the input-output device is connected to the processor.

[0309] Specifically, in the embodiment, the processor can be configured to execute the following steps through the computer program:

[0310] S201, receiving and screening multi-dimensional liver CT image data, generating a standardized DICOM data set based on preset layer thickness threshold, resolution threshold and artifact filtering rules, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold;

[0311] S202, performing a two-stage lesion positioning analysis on the standardized DICOM dataset, first generating a global liver segmentation mask through a cascaded deep learning model, then extracting a local ROI region based on the mask coordinates for sub-pixel level lesion detection, and outputting a lesion positioning atlas containing lesion coordinates, size and average CT value;

[0312] S203, starting an interactive image optimization engine based on the lesion positioning atlas, responding to the window width and window level adjustment instructions and measurement tool operations of the physician in real time, generating a dynamic enhanced lesion viewable field and synchronously updating the lesion parameters;

[0313] S204, integrating the viewable field data and the clinical rule base to generate a structured report, automatically filling the image findings field by checking the lesion list, and outputting a diagnostic proposal in accordance with the DICOM standard through the graphic-text report module, wherein the diagnostic proposal is embedded with a lesion position navigation link and a measurement error calibration.

[0314] The above describes the structure, features and effects of the present application in detail according to the embodiments shown in the drawings. The above description is only the preferred embodiments of the present application, but the present application is not limited by the drawings. Any changes or modifications made in accordance with the concept of the present application, or equivalent embodiments with equivalent changes, are still within the scope of protection of the present application.

Claims

1. A CT image assisted detection method for focal liver lesions, characterized in that, The method comprises: Receiving and screening multi-dimensional liver CT image data, generating standardized DICOM data set based on preset layer thickness threshold, resolution threshold and artifact filtering rules, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold; Performing two-stage lesion positioning analysis on the standardized DICOM data set, first generating a global liver segmentation mask through a cascaded deep learning model, then extracting a local ROI region based on the mask coordinates for sub-pixel level lesion detection, and outputting a lesion positioning atlas containing lesion coordinates, size and average CT value; wherein the standardized DICOM data set is input into a cascaded U-Net model, the first stage uses 3D dilated convolution to extract liver boundary features, and outputs a liver probability heat map; topological connectivity analysis is performed on the liver probability heat map, the edge is corrected by gradient vector flow field, and a sub-voxel liver mask and the corresponding spatial coordinate tree are generated; based on the spatial coordinate tree, a local region of interest (ROI) is defined, a deformable convolution network is used to adaptively adjust the size of the ROI, and a lesion candidate block set is output; phase consistency edge detection is applied to the lesion candidate block set, the lesion center offset is calculated by combining the Hessian matrix, and sub-pixel positioning coordinates are generated; the sub-pixel positioning coordinates and the CT value histogram statistics are fused, and the lesion positioning atlas is generated by radial basis function interpolation; Starting an interactive image optimization engine based on the lesion positioning atlas, real-time responding to the window width and window level adjustment instructions and measurement tool operations of the physician, generating a dynamically enhanced lesion visual field and synchronously updating the lesion parameters; Integrating visual field data and clinical rule library to generate a structured report, automatically filling in the image findings field by checking the lesion list, and outputting a diagnosis proposal in accordance with the DICOM standard through the graphic-text report module, wherein the diagnosis proposal embeds lesion location navigation links and measurement error calibration.

2. The method of claim 1, wherein, The receiving and screening multi-dimensional liver CT image data, generating standardized DICOM data set based on preset layer thickness threshold, resolution threshold and artifact filtering rules, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold, comprises: Receiving original DICOM stream from PACS system, synchronously accessing patient age, device model and scanning protocol metadata, and generating original data cube with space-time label; Inputting the original data cube into an adaptive frequency domain filter, dynamically selecting convolution kernel size based on device parameters, eliminating metal artifacts and motion blur, and outputting a purified image sequence; Performing interlayer consistency analysis on the purified image sequence, aligning adjacent slices through non-rigid registration algorithm, and interpolating to generate sub-millimeter level volume data according to the preset layer thickness threshold; Encoding patient age and lesion size threshold into feature vector, and inputting it into graph neural network together with sub-millimeter level volume data to generate dynamically weighted standardized DICOM data set.

3. The method of claim 2, wherein, The starting an interactive image optimization engine based on the lesion positioning atlas, real-time responding to the window width and window level adjustment instructions and measurement tool operations of the physician, generating a dynamically enhanced lesion visual field and synchronously updating the lesion parameters, comprises: Real-time receive physician's window width and window level adjustment signal and measurement tool trajectory data, analyze the operation type through gesture intention recognition algorithm, generate window level and window width parameter vector and measurement coordinate sequence; Based on the spatial coordinate tree of the lesion positioning atlas, pre-load the corresponding ROI body data in GPU memory, dynamically adjust the ray casting algorithm combined with the window level and window width parameter vector, and output the real-time volume rendering image; Superimpose anisotropic diffusion filtering on the real-time volume rendering image, adaptively enhance the lesion edge contrast according to the window width parameter, and synchronously embed the labeled trajectory of the measurement coordinate sequence to generate an enhanced labeled image; Compare the measurement coordinate sequence with the sub-pixel positioning coordinates of the lesion positioning atlas, calculate the lesion size deviation and CT value fluctuation, and real-time update the lesion dynamic parameter table; Fuse the enhanced labeled image and the lesion dynamic parameter table, generate a dynamic enhanced lesion visual field with a parameter superimposed layer through multi-planar reformatting technology.

4. The method of claim 3, wherein, The integrated visual field data and clinical rule base generate a structured report, automatically fill in the image findings field by checking the lesion list, and output a diagnosis proposal in accordance with the DICOM standard through the graphic-text report module, wherein the diagnosis proposal embeds lesion location navigation links and measurement error calibration, including: Analyze the quantitative data in the lesion dynamic parameter table, and perform feature matching with the clinical rule base to generate a structured lesion data unit; Map the physician's checked lesion list to the structured lesion data unit, generate structured diagnostic text by driving natural language generation through a medical knowledge graph, and output structured diagnostic text; Generate a DICOM spatial coordinate system based on the spatial coordinate tree of the lesion positioning atlas, convert the sub-pixel positioning coordinates into a DICOM-RT structure set, and create a lesion spatial navigation engine; Combine the structured diagnostic text and the lesion spatial navigation engine, add an error heat map of the measurement coordinate sequence, and generate a diagnosis proposal with error calibration.

5. A CT image-assisted detection system for focal liver lesions, characterized by, The system comprises: A receiving module for receiving and screening multi-dimensional liver CT image data, generating a standardized DICOM data set based on preset layer thickness threshold, resolution threshold and artifact filtering rules, wherein the data set dynamically associates patient age, scanning device parameters and lesion size threshold; The analysis module is used for performing a two-stage lesion positioning analysis on the standardized DICOM data set, first generating a global liver segmentation mask through a cascaded deep learning model, then extracting a local ROI region based on the mask coordinates for sub-pixel level lesion detection, and outputting a lesion positioning atlas containing lesion coordinates, size and average CT value; wherein the standardized DICOM data set is input into a cascaded U-Net model, the first stage uses 3D dilated convolution to extract liver boundary features, and outputs a liver probability heat map; topological connectivity analysis is performed on the liver probability heat map, the edge is corrected by a gradient vector flow field, and a sub-voxel liver mask and the corresponding spatial coordinate tree are generated; the local region of interest (ROI) is determined based on the spatial coordinate tree, and the deformable convolution network is used to adaptively adjust the ROI size, and the lesion candidate block set is output; the phase consistency edge detection is applied to the lesion candidate block set, the lesion center offset is calculated by combining the Hessian matrix, and the sub-pixel positioning coordinates are generated; the sub-pixel positioning coordinates and the CT value histogram statistics are fused, and the lesion positioning atlas is generated by radial basis function interpolation; The generation module is used for starting an interactive image optimization engine based on the lesion positioning atlas, responding to the window width and window level adjustment instructions and measurement tool operations of the physician in real time, generating a dynamic enhanced lesion visual field and synchronously updating the lesion parameters; The output module is used for integrating the visual field data and the clinical rule base to generate a structured report, automatically filling the image findings field by checking the lesion list, and outputting a diagnosis proposal in accordance with the DICOM standard through the graphic-text report module, wherein the diagnosis proposal is embedded with a lesion position navigation link and a measurement error calibration.

6. The system of claim 5, wherein, The receiving module is specifically configured to: receive an original DICOM stream from a PACS system, synchronously access patient age, device model and scanning protocol metadata, and generate an original data cube with space-time labels; input the original data cube into an adaptive frequency domain filter, dynamically select the convolution kernel size based on the device parameters, eliminate metal artifacts and motion blur, and output a purified image sequence; perform interlayer consistency analysis on the purified image sequence, align adjacent slices through a non-rigid registration algorithm, and interpolate sub-millimeter level volume data according to a preset layer thickness threshold; encode the patient age and lesion size threshold as a feature vector, and input the feature vector and the sub-millimeter level volume data into a graph neural network to generate a dynamically weighted standardized DICOM data set.

7. A storage medium, characterized by The storage medium has a computer program stored therein, wherein the computer program is configured to execute the method of any one of claims 1-4 when running.

8. An electronic device comprising a memory and a processor, characterized in that The memory has a computer program stored therein, and the processor is configured to execute the computer program to execute the method of any one of claims 1-4.

Images