CT image auxiliary detection method and system for liver focal lesion

Abstract

The invention discloses a CT image auxiliary detection method and system for liver focal lesions, and the method comprises the steps: receiving and screening multi-dimensional liver CT image data, and generating a standardized DICOM data set based on a preset layer thickness threshold value, a resolution threshold value and an artifact filtering rule; performing two-stage focus positioning analysis on the standardized DICOM data set, and outputting a focus positioning map containing focus coordinates, sizes and average CT values; starting an interactive image optimization engine based on the focus positioning map, responding to a window width and window level adjustment instruction and measurement tool operation of a physician in real time, generating a dynamically enhanced focus visible range and synchronously updating focus parameters; and integrating the visual field data and the clinical rule base to generate a structured report, automatically filling fields seen by the image by checking a focus list, and outputting a diagnosis proposal meeting the DICOM standard in linkage with an image-text report module. By utilizing the embodiment of the invention, standardization and intelligence of auxiliary diagnosis can be realized, and the accuracy and efficiency of liver lesion detection are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular to a CT image-assisted detection method and system for focal liver lesions. Background Art

[0002] CT detection of focal liver lesions is a key step in clinical diagnosis, but traditional methods rely on physician experience, resulting in high rates of missed diagnoses and low efficiency. Existing technologies often use a single model to process global images for CT image analysis, which is susceptible to artifacts and lacks dynamic interactive functionality, resulting in insufficient lesion localization accuracy. Furthermore, report generation is disconnected from image analysis, and clinical parameters (such as lesion size and CT value) must be manually entered, which can easily introduce errors. Although some systems have attempted to incorporate AI assistance, they have not addressed the technical pain points of multi-dimensional data standardization, real-time interactive optimization, and automatic generation of structured reports, making it difficult to meet the needs of precise diagnosis and treatment. Summary of the Invention

[0003] The purpose of the present invention is to provide a CT image-assisted detection method and system for focal liver lesions to address the deficiencies in the existing technology, achieve standardization and intelligence of auxiliary diagnosis, reduce human errors, and improve the accuracy and efficiency of liver lesion detection.

[0004] One embodiment of the present application provides a CT image-assisted detection method for focal liver lesions, the method comprising: Receive and filter multi-dimensional liver CT image data and generate a standardized DICOM dataset based on preset slice thickness thresholds, resolution thresholds, and artifact filtering rules, where the dataset is dynamically associated with patient age, scanning device parameters, and lesion size thresholds; A two-stage lesion localization analysis is performed on standardized DICOM datasets. A cascaded deep learning model is used to first generate a global liver segmentation mask. Then, based on the mask coordinates, local ROI regions are extracted for sub-pixel lesion detection. The resulting lesion localization map includes lesion coordinates, size, and average CT value. An interactive image optimization engine is activated based on the lesion localization map, responding in real time to the physician's window width and window position adjustment instructions and measurement tool operations, generating a dynamically enhanced lesion visual field and synchronously updating lesion parameters. Integrate visual domain data with the clinical rule base to generate structured reports. Automatically fill in the image findings field by checking the lesion list, and link the graphic report module to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation embeds a lesion location navigation link and measurement error calibration.

[0005] Optionally, the receiving and screening of multi-dimensional liver CT image data, and generating a standardized DICOM dataset based on a preset slice thickness threshold, resolution threshold, and artifact filtering rules, wherein the dataset is dynamically associated with patient age, scanning device parameters, and lesion size threshold, includes: Receive the original DICOM stream from the PACS system, synchronously access the patient age, device model and scan protocol metadata, and generate a raw data cube with time and space labels; The raw data cube is input into an adaptive frequency domain filter, and the convolution kernel size is dynamically selected based on the device parameters to remove metal artifacts and motion blur, and output a purified image sequence; Inter-slice consistency analysis is performed on the purified image sequence, adjacent slices are aligned using a non-rigid registration algorithm, and sub-millimeter volume data are generated by interpolation based on a preset slice thickness threshold; The patient's age and lesion size threshold are encoded as feature vectors and input into the graph neural network together with the submillimeter volume data to generate a dynamically weighted standardized DICOM dataset.

[0006] Optionally, a two-stage lesion localization analysis is performed 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 lesion detection, and outputting a lesion localization map containing lesion coordinates, size, and average CT value, including: The standardized DICOM dataset was input into the cascade U-Net model. In the first stage, 3D dilated convolution was used to extract liver boundary features and output a liver probability heat map. Perform topological connectivity analysis on the liver probabilistic heat map, correct the edges using the gradient vector flow field, and generate a subvoxel-level liver mask and the corresponding spatial coordinate tree; Based on the spatial coordinate tree, the local region of interest (ROI) is delineated, and the deformable convolutional network is used to adaptively adjust the ROI size and output a set of candidate lesion blocks. Phase consistency edge detection is applied to the lesion candidate block set, and the lesion center offset is calculated in combination with the Hessian matrix to generate sub-pixel positioning coordinates; Sub-pixel positioning coordinates and CT value histogram statistics are fused and radial basis function interpolation is used to generate a lesion localization map.

[0007] Optionally, the interactive image optimization engine is started based on the lesion localization map, responding in real time to the physician's window width and window position adjustment instructions and measurement tool operations, generating a dynamically enhanced lesion visual field and synchronously updating lesion parameters, including: Receive the physician's window width and position adjustment signals and measurement tool trajectory data in real time, analyze the operation type through gesture intention recognition algorithm, and generate window position and width parameter vectors and measurement coordinate sequences; Based on the spatial coordinate tree of the lesion localization map, the corresponding ROI volume data is preloaded into the GPU memory, and the ray casting algorithm is dynamically adjusted in combination with the window position and width parameter vectors to output a real-time volume rendering image. Anisotropic diffusion filtering is superimposed on the real-time volume rendering image to adaptively enhance the lesion edge contrast according to the window width parameter, and the annotation trajectory of the measurement coordinate sequence is simultaneously embedded to generate an enhanced annotation image; Compare the measured coordinate sequence with the sub-pixel positioning coordinates of the lesion localization map, calculate the lesion size deviation and CT value fluctuation, and update the lesion dynamic parameter table in real time; The enhanced annotated image is fused with the lesion dynamic parameter table, and a dynamic enhanced lesion visual field with parameter overlay is generated through multi-plane reconstruction technology.

[0008] Optionally, the integration of visual field data and clinical rule base generates a structured report, automatically populates the image findings field by checking the lesion list, and links the graphic report module to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation embeds a lesion location navigation link and measurement error calibration, including: Analyze the quantitative data in the lesion dynamic parameter table, perform feature matching with the clinical rule base, and generate structured lesion data units; Map the list of lesions selected by the physician to structured lesion data units, drive natural language generation through the medical knowledge graph, and output structured diagnostic text; Generate a DICOM spatial coordinate system based on the spatial coordinate tree of the lesion localization atlas, convert the sub-pixel localization coordinates into a DICOM-RT structure set, and create a lesion spatial navigation engine; Combining structured diagnostic text with the lesion spatial navigation engine, adding the error heat map of the measurement coordinate sequence, and generating a diagnostic recommendation with error calibration.

[0009] Another embodiment of the present application provides a CT image-assisted detection system for focal liver lesions, the system comprising: a receiving module, configured to receive and filter multi-dimensional liver CT image data and generate a standardized DICOM dataset based on preset slice thickness thresholds, resolution thresholds, and artifact filtering rules, wherein the dataset is dynamically associated with patient age, scanning device parameters, and lesion size thresholds; The analysis module performs a two-stage lesion localization analysis on standardized DICOM datasets. It first generates a global liver segmentation mask through a cascaded deep learning model, then extracts local ROI regions based on the mask coordinates for sub-pixel lesion detection, and outputs a lesion localization map containing lesion coordinates, size, and average CT value. A generation module is used to start the interactive image optimization engine based on the lesion localization map, respond to the physician's window width and window position adjustment instructions and measurement tool operations in real time, generate a dynamically enhanced lesion visual field, and synchronously update the lesion parameters; The output module is used to integrate visual domain data with the clinical rule base to generate structured reports. By checking the lesion list, the image fields are automatically filled in, and the graphic report module is linked to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation is embedded with a lesion location navigation link and measurement error calibration.

[0010] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0011] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0012] Compared with the existing technology, the present invention provides a CT image-assisted detection method for focal liver lesions, which receives and filters multi-dimensional liver CT image data, generates a standardized DICOM data set based on preset layer thickness thresholds, resolution thresholds and artifact filtering rules; performs two-stage lesion localization analysis on the standardized DICOM data set, and outputs a lesion localization map containing lesion coordinates, size and average CT value; activates an interactive image optimization engine based on the lesion localization map, responds in real time to the physician's window width and window position adjustment instructions and measurement tool operations, generates a dynamically enhanced lesion visual field and synchronously updates lesion parameters; integrates the visual field data with the clinical rule library to generate a structured report, automatically fills in the image view fields by checking the lesion list, and links the graphic report module to output a diagnosis recommendation that complies with the DICOM standard, thereby achieving standardization and intelligence of auxiliary diagnosis, reducing human errors, and improving the accuracy and efficiency of liver lesion detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A hardware structure block diagram of a computer terminal for a CT image-assisted detection method of focal liver lesions provided by an embodiment of the present invention; Figure 2 A schematic flow chart of a CT image-assisted detection method for focal liver lesions provided in an embodiment of the present invention; Figure 3 A schematic diagram of a DICOM image with a lesion outline circled provided by an embodiment of the present invention; Figure 4 A schematic diagram of a DICOM image without a lesion outline provided by an embodiment of the present invention; Figure 5 A structural schematic diagram of a CT image assisted detection system for liver focal lesions provided by an embodiment of the present application. DETAILED DESCRIPTION

[0014] The embodiments described below with reference to the drawings are exemplary and are only used to explain the present application and cannot be explained as a limitation of the present application.

[0015] The present application provides a CT image assisted detection method for liver focal lesions, which can be applied to an electronic device, such as a computer terminal, specifically, a general computer, etc.

[0016] The following will be described in detail by taking a computer terminal as an example. Figure 1 A hardware structural block diagram of a computer terminal of a CT image assisted detection method for liver focal lesions provided by an embodiment of the present application is shown in FIG. 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.

[0017] 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.

[0018] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0019] The internal memory provides an environment for the running 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.

[0020] 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 less components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0021] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0022] See also Figure 2 The embodiment of the present invention provides a CT image-assisted detection method for focal liver lesions, which may include the following steps: S201, receiving and screening multi-dimensional liver CT image data, generating a standardized DICOM dataset based on preset slice thickness thresholds, resolution thresholds, and artifact filtering rules, wherein the dataset is dynamically associated with patient age, scanning device parameters, and lesion size thresholds; Specifically, the original DICOM stream can be received from the PACS system, and the patient age, device model, and scan protocol metadata can be accessed synchronously to generate a raw data cube with spatiotemporal labels; The system first receives a raw DICOM (Digital Imaging and Communications in Medicine) data stream through a standardized interface (e.g., the DIMSE protocol) from the Picture Archiving and Communication System (PACS). This data stream contains the raw slice sequence of a liver CT scan, with each slice accompanied by DICOM tags such as patient ID, acquisition time, and series number. Simultaneously, the system retrieves metadata such as patient age (PA), CT device model (DM, such as "GE Revolution CT"), and scan protocol (SP, such as "Liver Triple Phase") from the hospital information system (HIS) or radiology information system (RIS). This metadata is integrated with spatial information (e.g., slice location and pixel spacing) in the DICOM stream to construct a 3D spatial index. This index uses the slice sequence number as the time axis (T-Axis) and the pixel coordinates as the spatial axes (X / Y-Axis), forming a raw data cube (RDC) with spatio-temporal labels (STL). For example, the RDC dimensions for a patient might be 512 × 512 × 200 (width × height × number of slices), with each voxel associated with the label {PA: 45 years old, DM: "Siemens Somatom Force", SP: "Portal Venous Phase"}.

[0023] The generation of spatiotemporal tags requires addressing the issue of multi-source data alignment. The system utilizes a timestamp matching algorithm (TMA): Delta Time Calculation (DTC) is calculated between the acquisition datetime (ADT) in the DICOM stream and the study request time (SRT) in the HIS / RIS. If the time difference is less than a preset threshold (e.g., 5 minutes), the examinations are considered to be from the same series. The device model and scan protocol are mapped to a preset device parameter library (DPL) using the device serial number (DSN) in the DICOM tag. For example, the DSN "CT-7890" is mapped to the parameters {number of detector rows: 128, tube voltage: 120 kV}. Patient age is stored in floating-point format (e.g., 45.3 years) for subsequent adaptive processing. The final generated RDC is a four-dimensional structure (width, height, number of slices, metadata channel). The metadata channel (MC) stores the encoded values ​​of fields such as {PA, DM, SP} (for example, PA is normalized to a value in the range 0-1).

[0024] To ensure data integrity, the system performs a redundancy check (RC). If a missing slice is detected (e.g., a discontinuous sequence number), a DICOM retransmission request (RR) is automatically triggered. If metadata conflicts occur (e.g., two age values ​​are associated with the same exam ID), the conflict resolution rule (CRR) is activated, prioritizing the fields within the DICOM tag. The final RDC output is encapsulated in a proprietary binary container format (BCF), with the header storing key-value pairs (KVPs) of the spatiotemporal tags and the data body storing the raw CT value matrix (CVM). This step provides standardized, temporally and spatially consistent input for subsequent processing.

[0025] The raw data cube is input into an adaptive frequency domain filter, and the convolution kernel size is dynamically selected based on the device parameters to remove metal artifacts and motion blur, and output a purified image sequence; The core of the Adaptive Frequency Domain Filter (AFDF) is the Dynamic Convolution Kernel (DCK) technology. The system parses the RDC's device model metadata (DM) and 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 using the formula KS = 2×ceil(PSF FWHM / pixel pitch) + 1. For a pixel pitch of 0.6 mm, KS = 2×ceil(0.35 / 0.6) + 1 = 3×3.

[0026] Frequency domain filtering is performed in three steps: Fast Fourier Transform (FFT): Converts each slice of the RDC from the spatial domain to the frequency domain (Frequency Domain) to generate a complex spectrum (Complex Spectrum, CS).

[0027] Metal Artifact Suppression (MAS): Detects high-frequency noise (HN) in the frequency domain, which appears as bright spots (BS) at the edges of the spectrum. A matching notch filter (NF) is selected based on the device model. For example, the GE Revolution CT uses a ring notch, while the Philips IQon uses a fan notch. The notch radius (NR) is dynamically adjusted to cover the artifact area.

[0028] Motion Blur Correction (MBC): A Wiener filter (WF) is used to inversely compensate for the blur caused by the PSF. The filter strength (FS) is dynamically adjusted based on patient 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).

[0029] The filtered spectrum is converted back to the spatial domain via an inverse Fourier transform (IFFT). The system evaluates the purification effect through residual analysis (RA): the root mean square error (RMSE) of the pre- and post-filtered images is calculated. If the RMSE exceeds 15 HU (Hounsfield units), the convolution kernel size is automatically increased (for example, from 3×3 to 5×5) and the filter is re-filtered. The resulting purified image sequence (PIS) reduces streak noise in areas with metal artifacts (such as those around gallbladder stents) by over 90% and improves liver edge sharpness by 20%-40%.

[0030] Inter-slice consistency analysis is performed on the purified image sequence, adjacent slices are aligned using a non-rigid registration algorithm, and sub-millimeter volume data are generated by interpolation based on a preset slice thickness threshold; Inter-slice Consistency Analysis (ISCA) aims to address slice misalignment caused by respiratory motion. The system uses the portal vein as a reference anatomical landmark and automatically locates the portal vein bifurcation point (BP) in the PIS. Adjacent slices are aligned using a non-rigid registration algorithm (NRA): Feature Extraction: Use the 3D Hessian Matrix to enhance the vascular structure and extract the portal vein centerline (CL).

[0031] Deformation Field Modeling: B-spline Free Form Deformation (BFFD) was used to generate the control point grid (CPG), and the grid spacing (GS) was set to 1.5 times the slice thickness (e.g., a 5 mm slice corresponds to GS = 7.5 mm).

[0032] Similarity Optimization: The control point displacements are optimized using the Limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) algorithm, using mutual information (MI) as the metric. The maximum number of MI iterations is set to 100.

[0033] After registration, the inter-slice offset vector (IOV) is calculated. If the IOV magnitude (MAG) is greater than 2 mm (preset tolerance), it is considered a respiratory motion artifact (RMA), triggering slice reordering (SR). For example, the nth slice must be translated along the vector (Δx = 1.2 mm, Δy = 0.8 mm) to match the n-1th slice. A continuous liver mask (LM) is generated from the aligned sequence, and morphological closing is used to fill the intrahepatic vascular gap (VG).

[0034] Interpolation is performed based on a preset slice thickness threshold (STT) (e.g., 1.0 mm): If the original thickness (OT) ≤ STT (e.g. 0.625 mm), the original data is directly output.

[0035] If OT > STT (e.g., 5.0 mm), subvoxel interpolation based on thin plate splines (TPS) is employed: Using the registered vessel centerline as the anchor point (AP), the spline coefficient (SC) is calculated to generate submillimeter volumetric data (SVD). For example, when interpolating a 5 mm slice to a 1 mm slice thickness, four new slices are inserted between two slices. The CT values ​​of these new slices are weighted by the adjacent slices (weights are inversely proportional to distance). This interpolation eliminates stair-step artifacts at the liver margin and reduces the error in lesion volume measurement from 15% to less than 5%.

[0036] The patient's age and lesion size threshold are encoded as feature vectors and input into the graph neural network together with the submillimeter volume data to generate a dynamically weighted standardized DICOM dataset.

[0037] Feature Vector Encoding (FVE) converts non-image data into a format that can be processed by neural networks: Patient age (PA) is coded in three parts: Age Normalization (AN): PA / 100 (e.g., 45 years old → 0.45).

[0038] Age Segment One-hot (ASO): [Child: 0, Adult: 1, Elderly: 0] (45 years old is "adult").

[0039] Age-related parameters (ARP): For example, the parenchyma attenuation coefficient (PA) is set to -5 + 0.1×PA HU (45 years old → -0.5 HU).

[0040] Lesion Size Threshold (LST) is coded as: Absolute Threshold (AT): such as 5 mm → 0.5 cm.

[0041] Relative Threshold (RT): AT / Max Liver Diameter (MLD).

[0042] Finally, an 8-dimensional feature vector (FV) is generated, for example, [0.45, 0, 1, 0, -0.5, 0.5, 12.5, 0.04] (assuming MLD = 12.5 cm).

[0043] The construction process of Graph Neural Network (GNN): Graph Construction: Node: represents a supervoxel (SV) in the volume data, which is generated by 3×3×3 voxel clustering.

[0044] Edge: connects adjacent supervoxels. The edge weight (EW) is determined by the CT value gradient (CT Gradient, CTG) (the smaller the gradient, the higher the weight).

[0045] Dynamic Weighted Fusion (DWF): The feature vector FV is mapped to the initial node feature (INF).

[0046] The lesion size threshold LST controls the receptive field (RF) of the convolutional layer: when LST < 1 cm, RF = 3 (sensitive to small lesions), and when LST > 2 cm, RF = 7 (covering large lesions).

[0047] Message Passing: Neighbor Feature (NF) is weighted aggregated (WA) according to edge weight.

[0048] The patient's age and PA modulate the aggregation strength: elderly patients (PA>60) enhance smoothing (reduce noise), while young patients enhance detail preservation.

[0049] The network output is the dynamic weight (DW) of each supervoxel, which is used to weightedly fuse the original data: Liver parenchyma: DW = 1.0 (retain original value); Vessel edge: DW = 2.0 (enhanced contrast); Suspected small lesion area (LST < 1 cm): DW = 1.5 (increased sensitivity).

[0050] The weighted volume data were reconstructed using the DICOM header file (Header Reconstruction), with standardized labels added (e.g., recommended window width / level values ​​of WW / WL = 250 / 50), to generate a standardized DICOM dataset (SDD). For example, the edge contrast of liver cysts in elderly patients was improved by 30% in the SDD, and the detection rate of small nodules with a diameter of 1 mm was increased by 25%. One example of standardized DICOM data is shown in Table 1.

[0051] Table 1

[0052] This step uses an intelligent data screening mechanism to convert raw CT images from different scanning devices into standardized analysis units. The system not only considers the physical parameters of the image itself (such as slice thickness and resolution), but also integrates individual patient characteristics (age) and device differences (such as the noise characteristics of different CT machines) to establish a dynamic data quality control system. Artifact filtering rules utilize a combination of frequency domain analysis and deep learning to effectively eliminate image interference caused by metal implants, achieving standardized processing of multi-source heterogeneous image data and providing high-quality input data for subsequent analysis. By dynamically correlating clinical parameters, the sensitivity and specificity of subsequent lesion detection can be adaptively adjusted based on the patient's actual situation, avoiding misjudgments due to device differences or patient size.

[0053] S202, performing a two-stage lesion localization analysis on the standardized DICOM dataset. A cascaded deep learning model is used to first generate a global liver segmentation mask. Then, based on the mask coordinates, local ROI regions are extracted for sub-pixel lesion detection. The lesion localization map containing lesion coordinates, size, and average CT value is output. Specifically, the standardized DICOM dataset can be input into the cascade U-Net model. In the first stage, 3D dilated convolution is used to extract liver boundary features and output a liver probability heat map; The input source is a standardized DICOM dataset, essentially three-dimensional volume data (with a voxel resolution of up to 0.625 mm × 0.625 mm × 0.5 mm) that has undergone rigorous quality control. This dataset is first loaded into the first stage of the cascaded U-Net model. The core of this stage is the 3D dilated convolution. Unlike traditional convolution, dilated convolution introduces a dilation rate (DR) parameter (e.g., DR=2 or DR=4) to insert cavities between convolution kernel elements, significantly expanding the receptive field without increasing the number of parameters. For example, a 3×3×3 convolution kernel with DR=2 actually covers a voxel range equivalent to 7×7×7, enabling it to capture macroscopic liver boundary features (such as liver lobe boundaries and the direction of the portal vein) across multiple slices.

[0054] At the model architecture level, the first stage consists of four layers: Feature extraction layer: Three consecutive 3D dilated convolutional layers (kernel size 3×3×3, DR=2) are used, each followed by batch normalization (BN) and rectified linear unit (ReLU) activation. This design effectively suppresses common noise interference in CT images (such as low-dose scanning noise) while enhancing the contrast boundary between the liver and surrounding tissues (diaphragm, intestine, and kidney).

[0055] Contextual Fusion Layer: A spatial pyramid pooling (SPP) module is introduced at the end of the encoding path. This module extracts multi-scale liver features by using multiple dilated convolution branches with different DR values ​​(e.g., DR = 1, 4, and 8) in parallel. For example, a DR = 1 branch focuses on local texture (such as the smoothness of the liver capsule), while a DR = 8 branch identifies global morphology (such as the volume percentage of the right lobe of the liver). Finally, multi-scale information is fused through feature concatenation.

[0056] Probability Mapping Layer: The decoding path uses transposed convolution (stride 2×2×2) to progressively upsample the feature map and fuse it with the same-scale features from the encoding path via skip connections. The final output layer uses a 1×1×1 convolution kernel with a sigmoid activation function to generate a liver probability heatmap of the same size as the input volume. Each voxel in this heatmap represents a continuous probability value between 0 and 1 (e.g., 0.93 indicates a 93% probability that the location belongs to the liver), enabling intuitive distinction between liver parenchyma (highlighted areas) and non-liver tissue (dark areas).

[0057] The heat map generation process relies on real-time GPU parallel acceleration (such as the NVIDIA A100 Tensor Core GPU), keeping single-case data inference time within 3.5 seconds. The output is stored as a floating-point matrix, providing a probabilistic basis for subsequent precise segmentation.

[0058] Perform topological connectivity analysis on the liver probabilistic heat map, correct the edges using the gradient vector flow field, and generate a subvoxel-level liver mask and the corresponding spatial coordinate tree; The liver probability heat map needs to be converted into an accurate liver mask, i.e., a binary label (0 for background, 1 for liver). Traditional thresholding methods (such as a fixed threshold of 0.5) can easily lead to jagged edges or mis-segmentation of small areas. This step uses topological connectivity analysis combined with gradient vector flow (GVF) to achieve subvoxel-level accuracy: Initial segmentation and connected domain labeling: Adaptive thresholding was applied to the heat map, and the threshold was dynamically calculated based on the average probability value of the liver parenchyma (e.g., mean × 0.7).

[0059] 3D Connected Component Labeling (CCL) was performed on the binarized results to label all independent regions. Based on prior knowledge (adult liver volume ranges from 800 to 1600 cubic centimeters), noisy regions with a volume smaller than 100 cubic centimeters were filtered out.

[0060] GVF edge optimization: Calculate the three-dimensional gradient field of the heat map and generate the initial edge force.

[0061] Constructing a gradient vector flow (GVF): By iteratively solving partial differential equations, the initial edge force is diffused throughout space, forming a smooth vector field directed toward the boundary. GVF parameter settings include the viscosity coefficient (μ = 0.15) and the iteration count (IC = 100) to ensure that the force field can penetrate deep concave structures (such as the vascular space in the portal area).

[0062] The GVF force field is applied to the initial segmentation boundary to drive the contour line to move toward the real liver edge (such as correcting the mis-segmented gallbladder fossa area).

[0063] Subvoxel mask generation: The optimized continuous surface is converted into a triangular mesh using the Marching Cubes Algorithm, with a vertex accuracy of 0.1 voxel (i.e., sub-voxel level).

[0064] Voxelize the mesh to generate the final liver mask and construct the spatial coordinate tree: Root node: stores the coordinates of the liver's overall bounding box (such as the starting point [X_min, Y_min, Z_min], the end point [X_max, Y_max, Z_max]).

[0065] Subnodes: Divide the liver into 8 subregions according to the Couinaud segmentation method. Each subnode stores the coordinate range and volume proportion of the corresponding submask (e.g., liver S8 segment accounts for 18%).

[0066] This mask can be directly used for three-dimensional liver volume measurement, and the spatial coordinate tree provides a spatial index basis for subsequent lesion localization.

[0067] Based on the spatial coordinate tree, the local region of interest (ROI) is delineated, and the deformable convolutional network is used to adaptively adjust the ROI size and output a set of candidate lesion blocks. The spatial coordinate tree of the liver mask serves as a navigation framework to guide lesion detection to focus on the liver parenchyma: Intelligent ROI delineation: The child nodes (liver segment regions) of the spatial coordinate tree were traversed, and the initial local region of interest (ROI) was generated by expanding the bounding box of each child node with a 15 mm safety margin (to avoid truncation of edge lesions).

[0068] The ROI size is dynamically adjusted based on the volume of the subnodes: for subregions with a volume greater than 150 cubic centimeters (such as the right posterior lobe), the ROI size is set to 120 mm × 120 mm × 80 mm; for regions with a volume less than 80 cubic centimeters (such as the caudate lobe), the ROI size is reduced to 60 mm × 60 mm × 40 mm.

[0069] Deformable Convolutional Network Processing: The volume data of each ROI region is cropped and input into the Deformable Convolutional Network (DCN). The core of DCN is to add an offset prediction layer before the standard convolution layer: This layer predicts the three-dimensional offset (Δx, Δy, Δz) of each sampling point through a 1×1×1 convolution, and the offset range allows ±3 voxels (to adapt to lesion morphological variations).

[0070] For example, for liver cancer lesions, the offset will cause the convolution kernel to focus on the "fast in and fast out" enhancement area; for cysts, it will avoid the internal homogeneous area and enhance the thin wall features.

[0071] The network backbone adopts a lightweight 3D ResNet-18 architecture, and the last layer is replaced by a Region Proposal Network (RPN) to output a bounding box.

[0072] Generation of lesion candidate block set: RPN generates about 300 candidate boxes for each ROI and filters out redundant boxes through non-maximum suppression (NMS, overlap threshold 0.5).

[0073] The candidate boxes with a confidence score greater than 0.85 are retained, and their corresponding voxel blocks (e.g., 32×32×32 voxels) are intercepted to form a lesion candidate block set.

[0074] Each block is accompanied by spatial coordinates (center point [X_c, Y_c, Z_c] in the DICOM coordinate system) and initial size estimation (long diameter L_dia, short diameter S_dia).

[0075] Phase consistency edge detection is applied to the lesion candidate block set, and the lesion center offset is calculated in combination with the Hessian matrix to generate sub-pixel positioning coordinates; Candidate blocks need to be further accurately located at the sub-pixel level: Phase Congruency Edge Detection: Process the transverse, sagittal, and coronal planes of each candidate block separately; Use a Log-Gabor filter bank (4 directions × 5 scales) to calculate the phase congruency value (PCV) of each pixel. PCV∈[0,1], where values ​​closer to 1 indicate a higher probability that the point is located at a true edge (not affected by brightness changes). Pixels with PCV>0.7 are extracted to form the initial edge point set (such as the "mosaic sign" edge of liver cancer).

[0076] Hessian matrix analysis: Calculate the Hessian matrix (second-order derivative matrix) at each edge point. The matrix dimension is 3×3 (including partial derivatives in the XX, XY, XZ, YY, YZ, and ZZ directions). Perform eigenvalue decomposition on the Hessian matrix to obtain eigenvalues ​​λ1, λ2, and λ3 (sorted by absolute value: |λ1|≥|λ2|≥|λ3|). The structure type is determined based on the combination of eigenvalues: if λ1<<0 and λ2≈λ3≈0, then the point is a candidate center point of a spherical structure (such as a cyst); if λ1<<0, λ2<<0 and λ3≈0, then it is an interference point of a tubular structure (such as a blood vessel) and needs to be excluded.

[0077] Sub-pixel center positioning: For the filtered feature points (such as points with λ1<-0.1), a 5×5×5 voxel neighborhood is taken with the point as the center.

[0078] A quadratic surface function is fitted in the neighborhood, and the coefficients are solved by the least squares method. The point where the function gradient is zero is the sub-pixel lesion center (X_sub, Y_sub, Z_sub).

[0079] The final coordinate accuracy reaches 0.01 voxel (such as the actual physical accuracy of 0.006 mm) and is written into the sub-pixel positioning coordinate list.

[0080] Sub-pixel positioning coordinates and CT value histogram statistics are fused and radial basis function interpolation is used to generate a lesion localization map.

[0081] Integrate geometric and density information to generate a comprehensive lesion localization atlas (Lesion Localization Atlas): CT value histogram statistics: A spherical region of interest (diameter = 1.2 times the initial estimated size of the lesion) was taken as the center with the sub-pixel coordinates.

[0082] The CT values ​​(unit: Hounsfield units HU) of all voxels in the region were extracted to generate a CT value histogram (CT Histogram).

[0083] Calculate key statistics: Mean CT value (Mean CT): histogram mean (e.g., cyst ≈ 10 HU, liver cancer ≈ 45 HU).

[0084] Peak CT value (Peak CT): The value corresponding to the highest peak in the histogram (reflecting the subject density).

[0085] Heterogeneity Index (HI): (75th percentile - 25th percentile) / median (e.g., HI of metastatic tumor > 0.3).

[0086] Radial Basis Function Interpolation (RBFI): Objective: To fuse discrete lesion point information into a continuous spatial distribution map.

[0087] The core of the algorithm is to use the sub-pixel coordinates of each lesion as the control point and define the interpolation function: F(X) = Σ [wi· φ(||X - X_i||)]. Where φ is the radial basis function (using the Gaussian kernel φ(r) = exp(-(εr) 2 ), kernel width parameter ε=0.7), and weights w_i are assigned after normalization by the average CT value of the lesion.

[0088] The F(X) value is calculated voxel by voxel in the liver mask space to generate the CT value distribution potential field (CT Value PotentialField).

[0089] Graph synthesis and output: The atlas 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 by fitting an ellipse based on the histogram distribution); Density layer: average CT value, peak CT value, heterogeneity index HI; Spatial distribution layer: CT value distribution potential field (used to visualize lesion density gradient).

[0090] The output format is DICOM-SEG structured objects, which can be directly accessed by PACS systems. The atlas can automatically annotate lesions (e.g., "L1: liver segment S6, 8mm×6mm, MeanCT=52HU"), providing a data foundation for subsequent interactive diagnosis.

[0091] Using a coarse-to-fine detection strategy, the system first accurately delineates the liver's anatomical boundaries using a three-dimensional convolutional network, and then detects microlesions within a limited range. In particular, sub-pixel localization technology, by analyzing the gradient distribution and texture characteristics of CT values, can identify tiny lesions smaller than the voxel size, improving detection accuracy and significantly increasing the detection rate of small lesions. The multi-parameter output of the localization map provides a quantitative basis for subsequent diagnosis, overcoming the limitations of traditional methods that rely on physicians' visual estimation.

[0092] S203, starting an interactive image optimization engine based on the lesion localization map, responding in real time to the physician's window width and window position adjustment instructions and measurement tool operations, generating a dynamically enhanced lesion visual field and synchronously updating lesion parameters; Specifically, it can receive the physician's window width and window position adjustment signals and measurement tool trajectory data in real time, analyze the operation type through the gesture intention recognition algorithm, and generate the window position and window width parameter vector and measurement coordinate sequence; As physicians operate on the medical imaging workstation, the interactive image optimization engine continuously monitors input events through the graphical user interface (GUI). The Window Width / Level Adjustment Signal (WWLAS) is generated by the physician's manipulation of workstation sliders or shortcut keys. For example, when the physician drags the "Window Width" slider to 400 Hounsfield Units (HU) and the "Window Level" slider to 60 HU, the system captures these two values ​​as raw input signals. Furthermore, Measurement Tool Trajectory Data (MTTD) is generated by the physician's use of a mouse or stylus to draw geometric shapes on the screen image. This includes drawing a straight line to measure a lesion's diameter or a polygon to outline the lesion area, generating a continuous sequence of screen coordinate points (e.g., 50 coordinate points per second). This raw input data is encapsulated into data packets with a timestamp (TS) and an operator device identifier (DID) and transmitted to the analysis module via a high-speed bus.

[0093] The core of the parsing module is the Gesture Intent Recognition Algorithm (GIRA). This algorithm uses temporal pattern recognition technology to analyze the input stream: For slider operations, GIRA monitors the value change rate (CR, in HU / s) and the pause interval (PI). If the CR is greater than 50 HU / s and the PI is less than 0.3 seconds, it is considered a "rapid coarse adjustment." If the CR is less than 10 HU / s and the PI is greater than 1 second, it is considered a "fine adjustment."

[0094] For trajectory data, GIRA extracts the trajectory's curvature (CV), closure (CL), and point density (PD). For example, a trajectory consisting of 20 coordinate points with a curvature less than 0.1 and a closure greater than 95% is classified as a "polygonal measurement"; a straight trajectory defined by two endpoints and longer than 30 pixels is classified as a "linear measurement."

[0095] Finally, GIRA outputs structured instructions: the window width value (such as 400 HU) and the window level value (such as 60 HU) are packaged into a window level / width parameter vector (Window Level / Width Parameter Vector, WLWPV, in the format of [window width value, window level value]); the screen coordinates of the measurement trajectory (such as the starting point [X1, Y1], the end point [X2, Y2]) are serialized into a measurement coordinate sequence (Measurement Coordinate Sequence, MCS, in the format of [operation type, coordinate point 1, coordinate point 2,...]).

[0096] To ensure real-time performance, the system employs a double buffering mechanism (DBM). Instructions processed during the current processing cycle (PC, 16 milliseconds) are temporarily stored in the high-speed cache (HSC), while the WLWPV and MCS parsed during the previous cycle are transferred to downstream modules via direct memory access (DMA). Furthermore, an operation log (OL) records physician operating habits (such as commonly used window width and window level combinations and frequency of measurement tool use) to optimize GIRA's decision thresholds (e.g., dynamically adjusting the decision boundaries for CR and PI). This step accurately translates the physician's physical operations into machine-processable digital instructions, laying the foundation for subsequent image rendering.

[0097] Based on the spatial coordinate tree of the lesion localization map, the corresponding ROI volume data is preloaded into the GPU memory, and the ray casting algorithm is dynamically adjusted in combination with the window position and width parameter vectors to output a real-time volume rendering image. The system uses the Spatial Coordinate Tree (SCT) in the Lesion Localization Atlas (LLA). The SCT is a hierarchical data structure (HDS), with the portal vein as the root node (RN) and leaf nodes (LN) storing the lesion's 3D voxel coordinates (e.g., [120, 85, 45]). Upon receiving a WLWPV (e.g., [400, 60]), the engine uses the SCT to locate the local region of interest (ROI) where the target lesion is located. For example, if the lesion is 15 mm in size, the ROI is defined as a cube region (CR) extending 30 mm from the lesion's center. The system then queries the pre-established Volume Data Index Table (VDIT) to locate the ROI's storage location in the standardized DICOM dataset (e.g., data block offset = 0x5A3F).

[0098] The raw CT volume data corresponding to the ROI (e.g., a 512×512×30 voxel matrix) is asynchronously transferred to the global memory (GM) of the graphics processing unit (GPU) via the Compute Unified Device Architecture (CUDA) interface. Simultaneously, the GPU kernel for the Ray Casting Algorithm (RCA) is started: Transfer function remapping: A transfer function (TF) is dynamically generated based on the WLWPV. For example, a window width of 400 HU corresponds to a grayscale range of -140 HU to 260 HU (window level 60 ± 200 HU). This range is linearly mapped to a grayscale range of 0-255, and voxels outside the range are set to completely black / white.

[0099] Light step optimization: Along each ray, the CT value of sampling points (SP) is calculated by trilinear interpolation (TLI) with a step size (SS) of 0.5 mm.

[0100] Color synthesis: The color of SP is accumulated 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.

[0101] The whole process is implemented under the GPU parallel computing architecture, with a response time of less than 50 milliseconds and an output resolution of 1920x1080 real-time volume rendered images (RTVRI).

[0102] To reduce the delay, the system uses a smart preloading strategy (SPS): When the physician's mouse hovers over a lesion marker, the ROI data of the adjacent three layers (about 90 megabytes) are automatically preloaded.

[0103] A memory priority queue (MPQ) is established to keep the ROI data of the current focus area (FA) in the high-speed texture memory (TM) and transfer the non-focus data to the low-speed video memory.

[0104] 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 images.

[0105] Anisotropic diffusion filtering is superimposed on the real-time volume rendered image to adaptively enhance the lesion edge contrast according to the window width parameters, and the labeled trajectory of the measurement coordinate sequence is synchronously embedded to generate an enhanced labeled image. Post-processing enhancement is performed on the RTVRI: Anisotropic Diffusion Filter (ADF): Adjusts the diffusion intensity based on the image gradient (IG) direction. Set the initial conduction coefficient (CC) to 0.25 and the iteration count (IC) to 5. Reduce the diffusion intensity (CC=0.1) to preserve details at the edge of the lesion (gradient value >30); increase the diffusion intensity (CC=0.4) to suppress noise in homogeneous areas (gradient value <10).

[0106] Window-width-driven contrast enhancement: When the window width (e.g., 400 HU) is less than the standard soft tissue window value (350 HU), the edge sharpening (ES) algorithm is activated, using the Laplacian operator (LO) to enhance high-frequency components (enhancement factor EC = 1.8). When the window width exceeds 500 HU (bone window range), sharpening is disabled to avoid artifacts. The processed image shows an approximately 40% improvement in lesion boundary clarity (measured by the boundary gradient modulus).

[0107] Annotation trajectory embedding is divided into three steps: Coordinate conversion: Convert screen coordinates (e.g., [X1, Y1]) in the MCS to 3D coordinates using inverse perspective projection (IPP). For example, the screen point (320, 240) corresponds to the world coordinates (15.2, -28.7, 45.3).

[0108] Geometry drawing: Linear measurement: Draw a solid red line (2 pixels thick) in 3D space and mark the distance value at the end (such as "17.3mm").

[0109] Polygon measurement: fill the semi-transparent green area (transparency Alpha=0.4), and stroke the border (line width 1.5 pixels).

[0110] Depth buffer processing: Enable Z-buffering (ZB) to ensure that the annotations are correctly occluded with the volume rendering image, and automatically make the annotations semi-transparent (transparency increased to 0.7) when they are behind the organs.

[0111] The Hybrid Rendering Pipeline (HRP) is used to generate Enhanced Annotation Image (EAI): The volume rendering result is stored in the color attachment 0 (CA0) of the framebuffer object (FBO).

[0112] Annotations are rendered to Color Attachment 1 (CA1) with Stencil Test (ST) enabled.

[0113] CA0 and CA1 are mixed according to the depth value through the fragment shader (FS) and finally output to the display buffer (DB).

[0114] The image not only preserves the anatomical details of the medical image, but also intuitively displays the physician's measurement intentions.

[0115] Compare the measured coordinate sequence with the sub-pixel positioning coordinates of the lesion localization map, calculate the lesion size deviation and CT value fluctuation, and update the lesion dynamic parameter table in real time; The system starts the Coordinate Comparison Engine (CCE): Subpixel coordinate alignment: Extract the subpixel localization coordinates (SPLC) of the lesion in the LLA (with a precision of 0.1 voxel, such as [120.3, 85.7, 45.2]). Align the 3D coordinates of the MCS (e.g., the set of polygon vertices measured by the physician) to the SPLC coordinate system using an affine transformation (AT).

[0116] Spatial consistency check: Calculates the Hausdorff distance (HD) between the physician's annotation point set and the AI's annotation point set. If the HD is greater than 2 mm, a confidence warning flag (CWF) is triggered, indicated by a flashing yellow border on the interface.

[0117] Core indicators of quantitative calculation: Lesion Size Deviation (LSD): The absolute deviation (|18.5-17.2|=1.3mm) and relative deviation (1.3 / 17.2×100%=7.6%) between the physician-measured diameter (e.g., 18.5mm) and the AI-measured value (e.g., 17.2mm) were significant.

[0118] For polygonal areas, the calculated area overlap ratio (OR) is intersection area / union area × 100%.

[0119] CT Value Fluctuation (CTVF): The standard deviation (SD) of the CT values ​​within the physician-annotated area is calculated. For example, if the SD changes from 8.5 HU in the AI ​​analysis to 12.3 HU in the physician-annotated area, the fluctuation rate (FR) = (12.3-8.5) / 8.5 × 100% = 44.7%.

[0120] The calculation results are encapsulated as a deviation vector (DV), such as [size deviation value, volatility].

[0121] Dynamic parameter table update mechanism: The lesion parameter table (LPT) is stored in the in-memory database (IMD). The fields include lesion ID, coordinates, AI-measured size, physician-measured size, CT value mean, CT value standard deviation, etc.

[0122] When receiving a new DV: If LSD is less than 5% and CTVF is less than 15%, only the "Physician Measured Size" field is updated.

[0123] If LSD>10% or CTVF>30%, an abnormal flag (AF) is added to the table and the lesion re-detection process (RDP) is triggered.

[0124] The updated LPT is pushed to the interface component via the Publish-Subscribe Pattern (PSP), allowing physicians to view color-coded parameters in real time (deviations <5% are displayed in green, >10% are displayed in red).

[0125] The enhanced annotated image is fused with the lesion dynamic parameter table, and a dynamic enhanced lesion visual field with parameter overlay is generated through multi-plane reconstruction technology.

[0126] Image-data fusion uses a layered compositing architecture (LCA): Base Layer (BL): EAI (including volume rendering images and annotation trajectories).

[0127] Parameter Overlay Layer (POL): Extracts key data from the LPT to generate an Information Panel (IP). For example, a semi-transparent panel (0.7 transparency) is drawn in the lower right corner of the image to display: Lesion 07: coordinates (120,85,45); AI measurement: diameter 17.2 mm | average CT value 56 HU; Physician measurements: Diameter 18.5 mm (↑7.6%) | Current CT-SD 12.3 HU (↑44.7%).

[0128] Navigation Marker Layer (NML): Marks the location of the lesion being operated on with a flashing sphere (5 pixels in diameter) in the 3D view.

[0129] Multiplanar Reformation (MPR) technology enables multi-perspective collaboration: Main View (MV): 3D volume rendering view showing the fused BL+POL+NML.

[0130] Auxiliary Views (AV): Axial View (AXV): A cross-sectional view through the center of the lesion.

[0131] Coronal View (CV): A section along the front-to-back direction of the human body.

[0132] Sagittal View (SV): A section along the left-right direction of the human body.

[0133] View Linkage (VL): When the physician scrolls through a slice in any MPR view, other views automatically synchronize to the same anatomical location. When the physician outlines the lesion in AXV, the MV updates the 3D annotation in real time.

[0134] Finally, the dynamically enhanced lesion visualization domain (DELVD) is generated: Dynamic: When the physician adjusts the window width to the "liver window" (window width 150 HU / window level 70 HU), the system completes the following tasks within 300 milliseconds: recalculating the transfer function; executing GPU ray casting; updating the ADF filter parameters; and refreshing the CT value statistics in the LPT.

[0135] Enhancement: After the lesion edge is sharpened by ADF, the boundary contrast is increased to 1.5 times that of the original image (measured by edge signal-to-noise ratio ESNR).

[0136] Interactivity: When the physician clicks on the "Diameter Deviation Value" in the parameter table, the system automatically highlights the difference areas between AI and manual annotations in the MPR view (AI areas are outlined in blue, and manual areas are outlined in red).

[0137] This visual domain seamlessly integrates imaging data, quantitative parameters, and interactive operations to provide decision support for physicians.

[0138] This engine utilizes GPU-accelerated real-time rendering technology. As the physician adjusts the window width and window level, the system instantly recalculates the volume rendering and highlights the characteristic areas of the lesion. Interactive data from measurement tools triggers a backend parameter recalculation module, ensuring that the image display and quantitative data remain synchronized. This provides a "what you see is what you get" interactive experience during the diagnostic process, allowing physicians to quickly verify the nature of suspected lesions through real-time feedback. This dynamic parameter update mechanism eliminates the tedious process of repeatedly switching measurement tools in traditional workflows.

[0139] S204 integrates the visual domain data and the clinical rule base to generate a structured report. By checking the lesion list, the image findings field is automatically filled in, and the graphic report module is linked to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation is embedded with a lesion location navigation link and measurement error calibration.

[0140] Specifically, the quantitative data in the lesion dynamic parameter table can be parsed, feature matched with the clinical rule base, and structured lesion data units can be generated; Once the interactive image optimization engine generates the Lesion Dynamic Parameter Table (LDPT), it records key quantitative indicators during the physician's operation in real time, including the lesion's Long Axis Diameter (LAD), Short Axis Diameter (SAD), Mean CT Value (MCTV), Arterial Enhancement Value (AEV), and Portal Enhancement Value (PEV). The system then activates the Structured Data Generation Module (SDGM), which first cleans the LDPT data: abnormal fluctuations caused by respiratory motion (e.g., sudden changes in CT values ​​exceeding 50 Hounsfield Units (HU)) are removed, and the time series data is smoothed using Sliding Window Averaging (SWA). The cleaned parameter set is converted into a feature vector (FV), the dimensions of which include lesion morphological features (e.g., circularity = 4π × area / perimeter). 2 ), density characteristics (such as the difference in CT values ​​between the arterial phase and the portal venous phase ΔCT = PEV-AEV) and dynamic change characteristics (such as enhancement rate = ΔCT / time difference).

[0141] The feature vector is input into the Clinical Rule Base (CRB) for multi-level matching. CRB uses the Knowledge Graph Architecture (KGA) and contains three layers of logic: Basic rule layer: stores international guideline standards (such as LI-RADS classification). For example, "arterial phase enhancement >20HU and portal venous phase washout" corresponds to the malignant sign rule LR-5.

[0142] Expert experience layer: Integrates the hospital's historical diagnostic pathways, such as the combination of "age > 60 years old + lesion calcification + capsule retraction" triggering a cholangiocarcinoma warning.

[0143] Dynamic update layer: Federated Learning (FL) is used to regularly update rare disease features from multi-center data (such as the "fast in, slow out" reinforcement model of vitiligo hepatitis).

[0144] 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.

[0145] The matching result is encapsulated as a Structured Lesion Data Unit (SLDU). Each SLDU contains fixed fields: Basic attributes: lesion ID, spatial coordinates (DICOM coordinate system XYZ value), size (LAD / SAD, unit: millimeter mm).

[0146] Image features: enhancement mode encoding (such as "arterial phase high enhancement-portal phase equal enhancement" encoded as AEH_PEE), calcification flag (Calcification Flag, CF, Boolean value).

[0147] Rule association: triggered rule ID list (such as ["LI-RADS-5", "HCC_Rule_2023"]), confidence vector (such as [0.92, 0.78]).

[0148] 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).

[0149] Mapping the physician's checked lesion list to the structured lesion data unit, generating natural language through the medical knowledge graph, and outputting structured diagnosis text; Physicians select lesions from the selected lesion list (SLL) using checkboxes on the user interface, for example, selecting lesions numbered "L002" and "L005." The system activates the semantic mapper (SM), which includes a built-in bidirectional index table (BIT). The forward index maps lesion IDs (e.g., "L002") to their corresponding SLDU database addresses, while the reverse index associates the lesions' anatomical locations (e.g., "liver segment S8" corresponds to Segment 8 in the Couinaud partitioning). The mapping process performs conflict detection (CD). If selected lesions overlap spatially (e.g., center-to-center distance <5 mm), an alarm is triggered, prompting the physician to review. After successful mapping, the system prioritizes the SLDU sequence (e.g., those with higher confidence levels of malignancy).

[0150] The sorted SLDU sequence is fed into the Medical Knowledge Graph-driven Natural Language Generator (MKG-NLG). The core of this engine is a three-layer architecture: Entity linking layer: Links SLDU fields to UMLS (Unified Medical Language System) concepts, such as "arterial phase hyperenhancement" to concept C0452539.

[0151] Template selection layer: This layer matches predefined report templates from the Report Template Library (RTL) based on rule IDs. For example, for LI-RADS-5 lesions, the template automatically selected is "liver S {segment} see {size} mm nodule, with significant enhancement during the arterial phase and washout during the portal venous phase, meeting LR-5 category."

[0152] Parameter filling layer: Uses a Conditional Random Field (CRF) model to dynamically insert quantitative data. For example, "{size}" is replaced with the actual value "22.3mm" and an error range (such as "±1.2mm" derived from measurement error calibration) is appended after the size.

[0153] The NLG engine outputs structured diagnostic text (SDT), which has the following characteristics: Paragraph structuring: divided into sections according to "imaging findings - diagnostic opinions - recommendations", in line with ACR (American College of Radiology) reporting standards.

[0154] Semantic standardization: enforced use of RadLex terminology (e.g., replacing the non-standard term "early highlighting" with "arterial phase hyperenhancement").

[0155] Dynamic qualifiers: Qualifiers are added based on the confidence level (CL). For example, "clearly indicated" is output when CL is > 0.9, and "tends to consider" is output when CL is between 0.7 and 0.9. The final SDT is temporarily stored in the DICOM SR (Structured Report) object format, containing text content and metadata pointers.

[0156] Generate a DICOM spatial coordinate system based on the spatial coordinate tree of the lesion localization atlas, convert the sub-pixel localization coordinates into a DICOM-RT structure set, and create a lesion spatial navigation engine; The Spatial Coordinate Tree (SCT) in the Lesion Localization Atlas (LLA) is essentially an Octree Index (OI) in three-dimensional space. The system initiates the DICOM Spatial Coordinate Converter (DSCC), which first reads key parameters from 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] mm), and Slice Thickness (ST, such as 1.0 mm). The Affine Transformation Matrix (ATM) is then used to convert sub-pixel coordinates (e.g., X = 152.34 px, Y = 98.21 px, Z = 45 slices) into millimeter coordinates in the DICOM Absolute Coordinate System (DAS).

[0157] The converted millimeter coordinates are input into the DICOM-RT Structure Set Generator (DRSSG). Each lesion is modeled as a closed surface: Contour Generation: Using the sub-pixel location coordinates as the center, an elliptical contour (EC) is generated based on the size parameters (LAD / SAD). For example, for a lesion with a major axis of 22.3 mm and a minor axis of 18.1 mm, an ellipse with a semi-major axis of 11.15 mm and a semi-minor axis of 9.05 mm is generated in the XY plane.

[0158] Interlayer interpolation: Thin plate spline interpolation (TPSI) is used along the Z-axis to connect the contours of adjacent slices to form a continuous three-dimensional surface.

[0159] Structure set encoding: Define the ROI number (ROI Number, ROIN) and ROI name (such as "Liver_Lesion_01") according to the DICOM RT Structure Set standard, and write the reference coordinate system UID (Unique Identifier).

[0160] The construction of the Lesion Spatial Navigation Engine (LSNE) driven by structure set data: Spatial index construction: Import the DICOM-RT structure set into a 3D rendering engine (such as VTK or Three.js) to create a spatial R-Tree Index (RTI) to support fast distance queries (such as searching for blood vessels within a 5 mm range).

[0161] Interaction protocol encapsulation: Develop a WebSocket-based navigation protocol (NP), defining instruction sets such as "jump to lesion" (instruction code 0x01 + lesion ROIN) and "show adjacent structures" (instruction code 0x02 + distance threshold).

[0162] Visualization interface: A navigation panel (NP) is embedded in the PACS (Picture Archiving and Communication System) browser, which allows users to click on the lesion name to automatically locate the corresponding slice and highlight the outline.

[0163] Combining structured diagnostic text with the lesion spatial navigation engine, adding the error heat map of the measurement coordinate sequence, and generating a diagnostic recommendation with error calibration.

[0164] The system calls the Diagnostic Report Synthesizer (DRS), which first fuses the Structured Diagnostic Text (SDT) with the Lesion Space Navigation Engine (LSNE): Text-embedded navigation link: Insert a hyperlink after the "lesion description" field in the SDT. The URL protocol is a custom navigation protocol (such as nav: / / lesion?id=L002).

[0165] Spatial data binding: Create a mapping table (MappingTable, MT) between the ROIN of the DICOM-RT structure set and the lesion ID in the SDT to ensure that clicking a text link can trigger the spatial engine jump.

[0166] Context-aware rendering: When a physician browses a lesion in PACS, the corresponding text paragraph is automatically highlighted in the report sidebar, achieving image-text interlock (ITI).

[0167] The generation of the Error Heatmap (EH) is based on historical measurement data: Error source analysis: The physician's measurement coordinate sequence (MCS), such as a manually drawn lesion boundary point set, is collected and compared with the algorithm-generated sub-pixel localization coordinate (SPLC). Two types of errors are calculated: Boundary Error (BE): Hausdorff Distance (HD) between the MCS point and the algorithm contour.

[0168] Size Error (SE): The absolute difference between the manually measured long diameter and the algorithmic long diameter.

[0169] Thermal model construction: A two-dimensional error field (EF) was established with the lesion as the center. Kriging interpolation (KI) was used to convert discrete error values ​​into a continuous distribution. Error levels were assigned colors: blue (BE < 1 mm), yellow (1 mm ≤ BE < 3 mm), and red (BE ≥ 3 mm).

[0170] Dynamic calibration: Overlay a semi-transparent heat map on the DICOM image and annotate statistical values ​​(such as "Maximum boundary error: 2.4 mm").

[0171] Finally, an Error-Calibrated Diagnostic Report (ECDR) is generated: Format compliance: Conforms to the DICOM SR standard, encapsulating SDT text, DICOM-RT structure set, and error heat map into a single DICOM file, for example: SOP Class UID=1.2.840.10008.5.1.4.1.1.88.33.

[0172] Interactive function integration: Click on the "Lesion 1" text → automatically jump to the corresponding CT slice and highlight the contour.

[0173] Hover over the error heat map → display the historical error distribution box plot of this region.

[0174] Multi-modal output: generate a PDF version of the report (with clickable navigation links) and an HL7 message (trigger LIS system alerts) simultaneously.

[0175] The report is sent to the PACS archive via the DICOM Storage SCP service, completing the full-process closed loop.

[0176] The system intelligently matches the detection results with clinical guidelines (such as LI-RADS standards) and automatically generates a structured description containing key diagnostic elements. The specially designed navigation links allow direct jumping to the lesion corresponding slice, and 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 assurance for clinical decision-making.

[0177] As can be seen, multi-dimensional liver CT image data is received and screened, a standardized DICOM data set is generated based on preset layer thickness threshold, resolution threshold and artifact filtering rules; a two-stage lesion positioning analysis is performed on the standardized DICOM data set, and a lesion positioning atlas containing lesion coordinates, size and average CT value is output; based on the lesion positioning atlas, an interactive image optimization engine is started, real-time response to physician's window width and window level adjustment instructions and measurement tool operation, dynamic enhanced lesion visual field is generated and lesion parameters are updated synchronously; the visual field data and the clinical rule base are integrated to generate a structured report, and the image findings field is automatically filled by checking the lesion list, and the DICOM standard diagnostic advice is output by the graphic report module, so as to realize the standardization and intelligentization of auxiliary diagnosis, reduce human error, and improve the accuracy and efficiency of liver lesion detection.

[0178] Another embodiment of the present application provides a CT image assisted detection system for liver focal lesions, as shown in Figure 5 , the system can comprise: The receiving module 501 is used for receiving and screening multi-dimensional liver CT image data, and 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; Analysis module 502 is used to perform a two-stage lesion localization analysis on the standardized DICOM dataset. It first generates a global liver segmentation mask through a cascaded deep learning model, then extracts local ROI regions based on the mask coordinates for sub-pixel lesion detection, and outputs a lesion localization map containing lesion coordinates, size, and average CT value; The generation module 503 is used to start the interactive image optimization engine based on the lesion localization map, respond to the physician's window width and window position adjustment instructions and measurement tool operations in real time, generate a dynamically enhanced lesion visual field, and synchronously update the lesion parameters; The output module 504 is used to integrate the visual domain data with the clinical rule base to generate a structured report. By checking the lesion list, the image findings field is automatically filled in, and the graphic report module is linked to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation is embedded with a lesion location navigation link and measurement error calibration.

[0179] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0180] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps: S201, receiving and screening multi-dimensional liver CT image data, generating a standardized DICOM dataset based on preset slice thickness thresholds, resolution thresholds, and artifact filtering rules, wherein the dataset is dynamically associated with patient age, scanning device parameters, and lesion size thresholds; S202, performing a two-stage lesion localization analysis on the standardized DICOM dataset. A cascaded deep learning model is used to first generate a global liver segmentation mask. Then, based on the mask coordinates, local ROI regions are extracted for sub-pixel lesion detection. The lesion localization map containing lesion coordinates, size, and average CT value is output. S203, starting an interactive image optimization engine based on the lesion localization map, responding in real time to the physician's window width and window position adjustment instructions and measurement tool operations, generating a dynamically enhanced lesion visual field and synchronously updating lesion parameters; S204 integrates the visual domain data and the clinical rule base to generate a structured report. By checking the lesion list, the image findings field is automatically filled in, and the graphic report module is linked to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation is embedded with a lesion location navigation link and measurement error calibration.

[0181] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0182] Specifically, the electronic device may 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.

[0183] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201, receiving and screening multi-dimensional liver CT image data, generating a standardized DICOM dataset based on preset slice thickness thresholds, resolution thresholds, and artifact filtering rules, wherein the dataset is dynamically associated with patient age, scanning device parameters, and lesion size thresholds; S202, performing a two-stage lesion localization analysis on the standardized DICOM dataset. A cascaded deep learning model is used to first generate a global liver segmentation mask. Then, based on the mask coordinates, local ROI regions are extracted for sub-pixel lesion detection. The lesion localization map containing lesion coordinates, size, and average CT value is output. S203, starting an interactive image optimization engine based on the lesion localization map, responding in real time to the physician's window width and window position adjustment instructions and measurement tool operations, generating a dynamically enhanced lesion visual field and synchronously updating lesion parameters; S204 integrates the visual domain data and the clinical rule base to generate a structured report. By checking the lesion list, the image findings field is automatically filled in, and the graphic report module is linked to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation is embedded with a lesion location navigation link and measurement error calibration.

[0184] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A CT image-assisted detection method for focal liver lesions, characterized in that: The method comprises: Receive and filter multi-dimensional liver CT image data and generate a standardized DICOM dataset based on preset slice thickness thresholds, resolution thresholds, and artifact filtering rules, where the dataset is dynamically associated with patient age, scanning device parameters, and lesion size thresholds; A two-stage lesion localization analysis is performed on standardized DICOM datasets. A cascaded deep learning model is used to first generate a global liver segmentation mask. Then, based on the mask coordinates, local ROI regions are extracted for sub-pixel lesion detection. The resulting lesion localization map includes lesion coordinates, size, and average CT value. An interactive image optimization engine is activated based on the lesion localization map, responding in real time to the physician's window width and window position adjustment instructions and measurement tool operations, generating a dynamically enhanced lesion visual field and synchronously updating lesion parameters. Integrate visual domain data with the clinical rule base to generate structured reports. Automatically fill in the image findings field by checking the lesion list, and link the graphic report module to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation embeds a lesion location navigation link and measurement error calibration.

2. The method according to claim 1, characterized in that The receiving and screening of multi-dimensional liver CT image data, generating a standardized DICOM dataset based on a preset slice thickness threshold, a resolution threshold, and an artifact filtering rule, wherein the dataset is dynamically associated with patient age, scanning device parameters, and a lesion size threshold, includes: Receive the original DICOM stream from the PACS system, synchronously access the patient age, device model and scan protocol metadata, and generate a raw data cube with time and space labels; The raw data cube is input into an adaptive frequency domain filter, and the convolution kernel size is dynamically selected based on the device parameters to remove metal artifacts and motion blur, and output a purified image sequence; Inter-slice consistency analysis is performed on the purified image sequence, adjacent slices are aligned using a non-rigid registration algorithm, and sub-millimeter volume data are generated by interpolation based on a preset slice thickness threshold; The patient's age and lesion size threshold are encoded as feature vectors and input into the graph neural network together with the submillimeter volume data to generate a dynamically weighted standardized DICOM dataset.

3. The method according to claim 2, characterized in that The method performs a two-stage lesion localization analysis on the standardized DICOM dataset. A global liver segmentation mask is first generated through a cascaded deep learning model. Then, local ROI regions are extracted based on the mask coordinates for sub-pixel lesion detection. The output is a lesion localization map containing lesion coordinates, size, and average CT value, including: The standardized DICOM dataset was input into the cascade U-Net model. In the first stage, 3D dilated convolution was used to extract liver boundary features and output a liver probability heat map. Perform topological connectivity analysis on the liver probabilistic heat map, correct the edges using the gradient vector flow field, and generate a subvoxel-level liver mask and the corresponding spatial coordinate tree; Based on the spatial coordinate tree, the local region of interest (ROI) is delineated, and the deformable convolutional network is used to adaptively adjust the ROI size and output a set of candidate lesion blocks. Phase consistency edge detection is applied to the lesion candidate block set, and the lesion center offset is calculated in combination with the Hessian matrix to generate sub-pixel positioning coordinates; Sub-pixel positioning coordinates and CT value histogram statistics are fused and radial basis function interpolation is used to generate a lesion localization map.

4. The method according to claim 3, characterized in that The interactive image optimization engine is started based on the lesion localization map, responds to the physician's window width and window position adjustment instructions and measurement tool operations in real time, generates a dynamically enhanced lesion visual field, and synchronously updates lesion parameters, including: Receive the physician's window width and position adjustment signals and measurement tool trajectory data in real time, analyze the operation type through gesture intention recognition algorithm, and generate window position and width parameter vectors and measurement coordinate sequences; Based on the spatial coordinate tree of the lesion localization map, the corresponding ROI volume data is preloaded into the GPU memory, and the ray casting algorithm is dynamically adjusted in combination with the window position and width parameter vectors to output a real-time volume rendering image. Anisotropic diffusion filtering is superimposed on the real-time volume rendering image to adaptively enhance the lesion edge contrast according to the window width parameter, and the annotation trajectory of the measurement coordinate sequence is simultaneously embedded to generate an enhanced annotation image; Compare the measured coordinate sequence with the sub-pixel positioning coordinates of the lesion localization map, calculate the lesion size deviation and CT value fluctuation, and update the lesion dynamic parameter table in real time; The enhanced annotated image is fused with the lesion dynamic parameter table, and a dynamic enhanced lesion visual field with parameter overlay is generated through multi-plane reconstruction technology.

5. The method according to claim 4, characterized in that The integration of visual domain data and clinical rule base generates a structured report, automatically populates the image findings field by checking the lesion list, and links the graphic report module to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation embeds a lesion location navigation link and measurement error calibration, including: Analyze the quantitative data in the lesion dynamic parameter table, perform feature matching with the clinical rule base, and generate structured lesion data units; Map the list of lesions selected by the physician to structured lesion data units, drive natural language generation through the medical knowledge graph, and output structured diagnostic text; Generate a DICOM spatial coordinate system based on the spatial coordinate tree of the lesion localization atlas, convert the sub-pixel localization coordinates into a DICOM-RT structure set, and create a lesion spatial navigation engine; Combining structured diagnostic text with the lesion spatial navigation engine, adding the error heat map of the measurement coordinate sequence, and generating a diagnostic recommendation with error calibration.

6. A CT image-assisted detection system for focal liver lesions, characterized in that: The system comprises: a receiving module, configured to receive and filter multi-dimensional liver CT image data and generate a standardized DICOM dataset based on preset slice thickness thresholds, resolution thresholds, and artifact filtering rules, wherein the dataset is dynamically associated with patient age, scanning device parameters, and lesion size thresholds; The analysis module performs a two-stage lesion localization analysis on standardized DICOM datasets. It first generates a global liver segmentation mask through a cascaded deep learning model, then extracts local ROI regions based on the mask coordinates for sub-pixel lesion detection, and outputs a lesion localization map containing lesion coordinates, size, and average CT value. A generation module is used to start the interactive image optimization engine based on the lesion localization map, respond to the physician's window width and window position adjustment instructions and measurement tool operations in real time, generate a dynamically enhanced lesion visual field, and synchronously update the lesion parameters; The output module is used to integrate visual domain data with the clinical rule base to generate structured reports. By checking the lesion list, the image fields are automatically filled in, and the graphic report module is linked to output a diagnosis recommendation that complies with the DICOM standard. The diagnosis recommendation is embedded with a lesion location navigation link and measurement error calibration.

7. The system according to claim 6, characterized in that The receiving module is specifically configured to: Receive the original DICOM stream from the PACS system, synchronously access the patient age, device model and scan protocol metadata, and generate a raw data cube with time and space labels; The raw data cube is input into an adaptive frequency domain filter, and the convolution kernel size is dynamically selected based on the device parameters to remove metal artifacts and motion blur, and output a purified image sequence; Inter-slice consistency analysis is performed on the purified image sequence, adjacent slices are aligned using a non-rigid registration algorithm, and sub-millimeter volume data are generated by interpolation based on a preset slice thickness threshold; The patient's age and lesion size threshold are encoded as feature vectors and input into the graph neural network together with the submillimeter volume data to generate a dynamically weighted standardized DICOM dataset.

8. The system according to claim 7, characterized in that The analysis module is specifically used to: The standardized DICOM dataset was input into the cascade U-Net model. In the first stage, 3D dilated convolution was used to extract liver boundary features and output a liver probability heat map. Perform topological connectivity analysis on the liver probabilistic heat map, correct the edges using the gradient vector flow field, and generate a subvoxel-level liver mask and the corresponding spatial coordinate tree; Based on the spatial coordinate tree, the local region of interest (ROI) is delineated, and the deformable convolutional network is used to adaptively adjust the ROI size and output a set of candidate lesion blocks. Phase consistency edge detection is applied to the lesion candidate block set, and the lesion center offset is calculated in combination with the Hessian matrix to generate sub-pixel positioning coordinates; Sub-pixel positioning coordinates and CT value histogram statistics are fused and radial basis function interpolation is used to generate a lesion localization map.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 5 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 5.

Images