Liver ascites volume calculation method based on artificial intelligence semantic segmentation

Abstract

The invention relates to the technical field of medical image processing, and discloses a liver ascites volume calculation method based on artificial intelligence semantic segmentation, and the method comprises the following steps: obtaining medical image data of the abdomen of a patient; the medical image data is preprocessed; inputting the preprocessed data into a deep learning model for intelligent segmentation to obtain a segmentation result of the ascites region; calculating the volume of liver ascites according to the segmentation result; and analyzing the liver ascites distribution conditions at different time points and visually outputting analysis results. Through a deep learning semantic segmentation strategy, the model can learn and identify a complex mode in a medical image, ascites and other tissues or liquids with similar densities are more accurately distinguished, wrong classification and missing classification conditions are reduced, the accuracy of ascites area identification is improved, the deep learning model has very strong generalization ability, and the method is suitable for popularization and application. The system can adapt to image data from different patients, different devices and different operation conditions, and can be effectively applied in a wide clinical environment.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and specifically to a method for calculating the volume of ascites based on artificial intelligence semantic segmentation. Background Art

[0002] The generation of ascites mainly stems from the imbalance of the body's fluid balance and involves multiple complex pathological processes. Liver diseases such as liver cirrhosis trigger portal hypertension and hypoalbuminemia through liver tissue fibrosis, which not only reduces the plasma colloid osmotic pressure but also promotes the extravasation of intravascular fluid into the abdominal cavity. At the same time, the release of inflammatory mediators increases capillary permeability and is accompanied by the activation of the renin-angiotensin-aldosterone system (RAAS), further causing excessive retention of sodium and water. In addition, malignant tumors in the abdominal cavity change local vascular permeability by directly invading the peritoneum or secreting pro-inflammatory cytokines, further exacerbating fluid exudation and accumulation. The interaction of these factors makes ascites clinically characterized by uneven fluid distribution and complex composition.

[0003] Traditional assessment methods, such as physical examination and ultrasound (B-ultrasound), although simple to operate, rely on the experience of operators and individual differences of patients, lack objective quantitative indicators, and have limited accuracy and repeatability. With the progress of medical imaging technology, CT (Computed Tomography) and MRI (Magnetic Resonance Imaging) are increasingly widely used in the diagnosis of abdominal diseases, providing a more accurate imaging basis for the objective assessment of ascites volume. However, when segmenting and quantifying the ascites area in CT / MRI images, the following challenges are still faced: manual segmentation is labor-intensive and subjective, and simply relying on fixed thresholds or single features to segment the ascites area is easily affected by noise, imaging parameters, and lesion severity, resulting in difficult-to-guarantee accuracy. Summary of the Invention

[0004] The present invention provides a method for calculating the volume of ascites based on artificial intelligence semantic segmentation, aiming to achieve automatic and high-precision segmentation of the ascites area in abdominal CT / MRI. By introducing an artificial intelligence-based semantic segmentation algorithm and combining the powerful image detail and context feature extraction ability of deep learning, through three-dimensional reconstruction or voxel statistics, the present invention can accurately estimate the ascites volume, provide more objective data support for clinical diagnosis and treatment, and at the same time, through improvements in network design, preprocessing, and postprocessing, enhance the detection ability for irregular and small-volume ascites, reduce the dependence on manual intervention, and improve clinical applicability.

[0005] To achieve the above object, the present invention is realized through the following technical solutions: A method for calculating the volume of ascites based on artificial intelligence semantic segmentation, comprising the following steps:

[0006] Obtain medical image data of the patient's abdomen;

[0007] Preprocess the medical image data, including image registration, HU value calibration, and ROI extraction;

[0008] Input the preprocessed data into a deep learning model for intelligent segmentation to obtain the segmentation result of the ascites region;

[0009] Calculate the volume of hepatic ascites according to the segmentation result;

[0010] Analyze the distribution of hepatic ascites at different time points and visually output the analysis results.

[0011] By adopting the above technical solution, this method comprehensively applies multiple steps such as image data processing, intelligent segmentation, and data analysis, can accurately identify and quantify the situation of hepatic ascites, dynamically monitor its changes, provide comprehensive and intuitive information for doctors, and help improve the accuracy and efficiency of the diagnosis and treatment of hepatic ascites.

[0012] Preferably, the obtaining of the medical image data of the patient's abdomen is specifically to collect the abdominal image data of the patient through a 64-slice spiral CT scanner or an MRI scanner, and output the abdominal sequence images in DICOM format.

[0013] By adopting the above technical solution, high-resolution and high-quality abdominal image data can be obtained. The standardized output in DICOM format facilitates subsequent data processing and transmission, ensures the accuracy and universality of the data, and provides a reliable data basis for subsequent analysis.

[0014] Preferably, the image registration adopts a non-rigid B-spline registration algorithm. The HU value calibration includes linearly mapping the original CT value and applying morphological operations to specific regions, and the N4ITK algorithm is used for bias field correction of MRI images.

[0015] By adopting the above technical solution, the non-rigid B-spline registration algorithm can accurately align multi-phase images, facilitating the comparison of ascites changes at different times; the HU value calibration can enhance the ascites characteristics and suppress fat interference, making the ascites region in the CT image clearer; the N4ITK algorithm performs bias field correction on MRI images, which can improve the image quality, reduce the image non-uniformity, and enhance the accuracy of subsequent segmentation and analysis.

[0016] Preferably, the deep learning model is an improved 3D Attention U-Net model. The encoder of the model includes multiple levels of downsampling and spatial attention modules, the decoder includes multiple levels of upsampling and channel attention gating mechanisms, and multi-scale feature fusion is performed through skip connections.

[0017] By adopting the above technical solution, the improved 3D Attention U-Net model combines a spatial attention module and a channel attention gating mechanism, which can effectively extract and fuse multi-scale features and accurately segment the ascites area. The 4-level downsampling and upsampling structure can capture image information at different scales, improve the model's recognition ability for the ascites area, and adapt to complex medical imaging scenarios.

[0018] Preferably, when training the improved 3D Attention U-Net model, anatomical constraint Dice Loss and edge-sensitive Focal Loss are used as loss functions, and data augmentation is performed.

[0019] By adopting the above technical solution, anatomical constraint Dice Loss can prompt the model to better learn anatomical structure information and improve the accuracy of segmentation; edge-sensitive Focal Loss can enhance the attention to the edges of the ascites area and reduce edge segmentation errors. Data augmentation increases the diversity of training data, improves the generalization ability of the model, and enables the model to maintain good performance when facing different patients and different imaging conditions.

[0020] Preferably, when calculating the ascites volume, first convert the binary mask of the segmentation result into a three-dimensional point cloud, load a preset liver capsule surface model, remove the abnormal points that are more than a certain distance from the liver surface, and then calculate the ascites volume through the voxel integration formula.

[0021] By adopting the above technical solution, converting the binary mask into a three-dimensional point cloud can more intuitively represent the spatial distribution of ascites. Loading the liver capsule surface model and removing abnormal points can exclude the interference of other tissues, making the calculation result more accurately reflect the true volume of hepatic ascites.

[0022] Preferably, perform time series alignment on the ascites distribution at different time points, and use the dynamic time warping algorithm to calculate the volume change rate;

[0023] The visual output includes displaying the area of ascites growth and decline in a three-dimensional heat map and automatically generating a PDF report containing ascites volume values, change trend curves, and clinical advice threshold prompts;

[0024] It also includes using other deep learning models to replace the improved 3D Attention U-Net model. The other deep learning models include SegNet, Fully Convolutional Network (FCN), or Swin Transformer, etc., and adapting the corresponding network layers and loss functions;

[0025] In the data preprocessing stage, classical methods such as adaptive threshold and Histogram distribution analysis are combined with machine learning classifiers to make a secondary judgment on the candidate regions; during post-processing, the watershed algorithm is used to refine and segment the low-density regions in the dimensional space and fuse them.

[0026] By adopting the above technical solution, time series alignment can accurately compare the ascites changes at different time points, and the calculation of the volume change rate helps doctors understand the development trend of the disease. The three-dimensional heat map and the automatically generated PDF report visually display the analysis results to doctors, enabling doctors to quickly and comprehensively understand the situation of liver ascites and providing strong support for diagnosis and treatment decisions.

[0027] A liver ascites volume calculation system based on artificial intelligence semantic segmentation, for the above-mentioned liver ascites volume calculation method based on artificial intelligence semantic segmentation, the system includes:

[0028] A data acquisition module, used to obtain medical image data of the patient's abdomen;

[0029] A preprocessing module, used to preprocess the medical image data;

[0030] An intelligent segmentation module, used to input the preprocessed data into a deep learning model for intelligent segmentation to obtain the segmentation result of the ascites region;

[0031] A volume calculation module, used to calculate the liver ascites volume according to the segmentation result;

[0032] An analysis and output module, used to analyze the liver ascites distribution at different time points and visually output the analysis results.

[0033] By adopting the above technical solution, the system modularizes the entire liver ascites volume calculation method. Each module has a clear division of labor and works together to achieve the full-process automated processing from data acquisition to result output, improving the calculation efficiency and accuracy and providing a convenient and reliable tool for clinical applications.

[0034] A computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned liver ascites volume calculation method based on artificial intelligence semantic segmentation.

[0035] By adopting the above technical solution, the computer device can utilize its powerful computing power and storage function to efficiently execute the liver ascites volume calculation method, providing doctors with fast and accurate diagnostic basis and being convenient for wide application in places such as hospitals.

[0036] A readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, it implements the above-mentioned method for calculating the volume of ascites based on artificial intelligence semantic segmentation.

[0037] By adopting the above technical solution, the readable storage medium can conveniently store and spread the computer program of this calculation method, enabling different computer devices to run this program and realize the calculation of the volume of ascites, thus expanding the application scope of this method.

[0038] The present invention provides a method for calculating the volume of ascites based on artificial intelligence semantic segmentation. It has the following

[0039] Beneficial effects:

[0040] 1. Through the deep learning semantic segmentation strategy, the model of the present invention can learn and recognize complex patterns in CT or MRI images, more accurately distinguish ascites from other tissues or liquids with similar densities, reduce misclassification and missed classification situations, and improve the accuracy of ascites region recognition. The deep learning model has strong generalization ability and can adapt to imaging data from different patients, different devices, and different operating conditions, and can be effectively applied in a wide range of clinical environments.

[0041] 2. The present invention automatically performs preprocessing such as noise removal, gray normalization, and resampling to improve the image quality and provide high-quality input data for the deep learning model. The automated processing flow reduces the need for manual intervention, shortens the processing time, improves the standardization of the processing process, and helps improve clinical work efficiency and ensure result consistency.

[0042] 3. The present invention uses three-dimensional connected component analysis to identify and connect adjacent suspected ascites regions in each layer of slices, reduce missed classification, ensure the continuity and integrity of the ascites region, improve the accuracy of volume estimation, and calculate the ascites volume by combining three-dimensional voxel statistics with the physical size of voxels to provide an accurate volume calculation method.

[0043] 4. The automated image processing and segmentation technology of the present invention reduces the workload of professionals, enables non-experts to obtain fast and reliable analysis results, reduces the dependence on advanced professional skills, and accurately and quickly quantifying the ascites volume helps evaluate the disease state, formulate treatment plans, and monitor treatment effects. Especially in the management of chronic disease patients, it can help doctors more effectively manage ascites-related diseases and improve the efficiency of disease management. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic flow chart of a method for calculating the volume of ascites based on artificial intelligence semantic segmentation of the present invention;

[0045] Figure 2Flow diagram of a method for calculating the volume of ascites in the liver based on artificial intelligence semantic segmentation according to the present invention;

[0046] Figure 3 Schematic diagram of the hardware platform in an embodiment of the present invention;

[0047] Figure 4 System architecture diagram of a method for calculating the volume of ascites in the liver based on artificial intelligence semantic segmentation according to the present invention. Detailed implementation manners

[0048] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0049] Embodiment 1:

[0050] Please refer to the attached Figure 1 - attached Figure 3 , an embodiment of the present invention provides a method for calculating the volume of ascites in the liver based on artificial intelligence semantic segmentation, including the following steps:

[0051] Obtain medical image data of the patient's abdomen;

[0052] Preprocess the medical image data, including image registration, HU value calibration, and ROI extraction;

[0053] Input the preprocessed data into a deep learning model for intelligent segmentation to obtain the segmentation result of the ascites area;

[0054] Calculate the volume of ascites in the liver according to the segmentation result;

[0055] Analyze the distribution of ascites in the liver at different time points and visually output the analysis results.

[0056] This method comprehensively uses multiple steps such as image data processing, intelligent segmentation, and data analysis, can accurately identify and quantify the ascites in the liver, and dynamically monitor its changes, providing comprehensive and intuitive information for doctors, which helps to improve the accuracy and efficiency of the diagnosis and treatment of ascites in the liver.

[0057] This method specifically includes the following steps:

[0058] I. Data collection and input

[0059] 1. Equipment Selection and Parameter Setting: Select a 64-slice spiral CT scanner or an MRI scanner with high-resolution imaging capabilities to collect abdominal image data of patients. Before scanning, set the scanning parameters reasonably according to the specific conditions and clinical needs of the patients. For example, for CT scanning, adjust parameters such as tube voltage, tube current, and slice thickness to obtain clear abdominal CT images with rich details; for MRI scanning, select appropriate scanning sequences (such as T1WI, T2WI, etc.), repetition time (TR), echo time (TE), etc. parameters to ensure that the imaging information of ascites and surrounding tissues can be accurately captured.

[0060] 2. Data Transmission and Format Conversion: After the scanning is completed, the equipment will output abdominal sequence images in DICOM (Digital Imaging and Communications in Medicine) format. Through the hospital's internal PACS (Picture Archiving and Communication System) system or a dedicated data transmission interface, the DICOM format images are transmitted to the data input layer. The data input layer has the function of automatically identifying and parsing DICOM format data, can quickly extract the key information of the images (such as patient basic information, scanning time, image dimensions, etc.), and transmit it to the subsequent processing module. At the same time, it supports the input of multi-phase images, providing a data basis for the comprehensive analysis of the changes in ascites.

[0061] II. Data Preprocessing

[0062] 1. Image Registration

[0063] Algorithm Implementation: Use the non-rigid B-spline registration algorithm to register multi-phase images. In actual operation, use a dedicated image registration software library (such as ITK - Insight Segmentation and Registration Toolkit) to implement this algorithm. First, select one of the multi-phase images as the reference image, and the remaining images as the floating images. Then, according to the optimization function, in each iteration process, continuously adjust the spatial transformation parameters of the floating image to gradually increase the similarity between the floating image and the reference image. The similarity is evaluated by calculating the difference between image pixels (such as gray value difference). When the difference is less than the preset threshold or reaches the maximum number of iterations, stop the iteration and complete the registration.

[0064] Optimization Function:

[0065] Among them, λ = 0.8, the number of iterations ≤ 50 times, λ is a weight parameter used to balance the importance of the two items; T * represents the optimal spatial transformation parameter; argmin TIt means to find the one T that makes the following expression reach the minimum value among all possible transformations; x is the pixel position in the image, Ω represents the spatial domain of the image, and I f (χ) represents the gray value of the floating image at pixel x; m (T(χ)) means the reference image I m After the spatial transformation T, the gray value at the pixel point x. This value is the same as I f (χ) comparison, the smaller the difference, the better the two images match near this pixel; calculate the sum of squares of the gray value differences between the floating image and the transformed reference image at all pixels. By minimizing this sum of squares, the gray value distributions of the two images can be made as similar as possible, which is a key indicator for measuring the effect of image registration. It is the Laplace operator of transformation T at pixel point x, which is used to measure the smoothness of the transformation. The larger the value of the Laplace operator, the more drastic the change of the transformation at that point, that is, the transformation is not smooth enough).

[0066] Accuracy verification: After the registration is completed, the registration accuracy is verified by calculating the spatial error of the registered image. For example, multiple feature points in the image (such as specific anatomical landmarks at the edge of the liver) are selected to compare the spatial position changes of these feature points before and after the registration. If the registration accuracy reaches the design requirement of 0.3mm, it means that the registration is successful and the registered image can be used for subsequent processing; if the accuracy does not meet the standard, the algorithm parameters are readjusted or other reference images are selected for registration.

[0067] 2. HU value calibration: Ascites feature enhancement: For CT images, the original CT value (-50 to 150 HU) is linearly mapped to the grayscale interval [0, 255]. Use image processing software (such as OpenCV) to write code to implement this conversion process. By traversing each pixel of the CT image, according to the linear mapping formula Calculate the new gray value of each pixel to enhance the contrast of the ascites area in the image, making it easier to analyze and identify later. Among them, new_gray represents the new gray value obtained after linear mapping, and its value range is [0,255]. old_HU represents the original CT value, that is, the CT value corresponding to each pixel in the medical CT image. In the scenario of liver ascites detection, its range is usually between -50 and 150HU.

[0068] Fat interference suppression: A morphological opening operation (kernel size 3×3) is applied to the area with CT value <-10HU. In the code implementation, the morphological operation function in OpenCV is used to define a 3×3 structure element (such as a rectangular structure element) to perform an opening operation on the area with CT value less than -10HU. The opening operation first performs an erosion operation to remove small noise points and isolated areas, and then performs an expansion operation to restore the eroded target area, thereby effectively suppressing the interference of fat tissue on the recognition of ascites area.

[0069] 3. Bias field correction (for MRI images): Use the N4ITK algorithm to perform bias field correction on MRI images. In practice, with the help of the N4ITK implementation tool provided by the ITK software, the number of iterations is set to 200. When the MRI image is input into the N4ITK algorithm, the algorithm will automatically detect and correct the low-frequency intensity inhomogeneity in the image. By adjusting the grayscale value of the image, the overall brightness and contrast of the image are made more uniform, the image quality is improved, and more reliable data is provided for subsequent segmentation and analysis.

[0070] 3. Intelligent Segmentation

[0071] 1. Model construction and training

[0072] Model construction: Build an improved 3DAttention U-Net model. Use a deep learning framework (such as PyTorch) to build the network structure and define the encoder, decoder, and jump connection parts. The encoder contains 4 levels of downsampling, each of which consists of 2 3×3×3 convolutional layers and a spatial attention module. In the code implementation, the feature extraction of the input image and the application of the attention mechanism are realized by defining the convolutional layer class and the attention module class. The decoder has 4 levels of upsampling and adopts a channel attention gating mechanism (Channel Gate structure). The relevant modules are also defined through code to implement the upsampling and channel attention functions. The jump connection part introduces a multi-scale feature fusion unit (MSFF) and uses the multi-scale feature fusion formula (Weight w i Multi-scale feature fusion is performed through 1×1×1 convolution (dynamic generation) to ensure that the model can make full use of information at different scales and improve segmentation accuracy.

[0073] Among them, F fused Represents the fused features; Represents the features of the encoder at level i. Indicates the mode of operation of feature fusion, which usually refers to element-by-element addition. is a feature of the decoder stage 4-i.

[0074] Training process: Prepare a large amount of labeled abdominal CT / MRI image data as the training set. During the training process, anatomical constraint Dice Loss and edge-sensitive Focal Loss (γ = 2) are used as the loss functions. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the model parameters are adjusted by an optimizer (such as the Adam optimizer) to gradually reduce the loss function. At the same time, data augmentation operations are performed, including online generation of elastic deformations (σ = 10, α = 15) and random mirror flips. By data augmentation, the diversity of the training data is increased, the generalization ability of the model is improved, and overfitting is prevented. The training process is carried out on a medical imaging workstation equipped with an NVIDIA A100 GPU to accelerate the training speed and improve the training efficiency.

[0075] 2. Model inference: Input the preprocessed abdominal imaging data into the trained improved 3D Attention U-Net model for inference. The model will automatically extract image features, and through the collaborative work of the encoder, decoder, and skip connections, output the segmentation result of the ascites area, that is, a binary mask image, where the white area represents the ascites area and the black area represents other tissues.

[0076] IV. Volume calculation

[0077] 1. Three-dimensional point cloud generation: Use an image processing algorithm to convert the binary mask output by the intelligent segmentation module into a three-dimensional point cloud. In the Python environment, this conversion process is implemented with the help of relevant point cloud processing libraries (such as Open3D). By traversing each pixel of the binary mask, the spatial coordinates of the pixels belonging to the ascites area (pixels with a value of 1) are extracted and converted into the three-dimensional point cloud format, and the point spacing is set to 0.5mm to ensure that the point cloud can accurately reflect the spatial distribution of the ascites area.

[0078] 2. Anatomical constraint processing

[0079] Model loading: Load the preset liver capsule surface model (STL format). Use an STL file reading library (such as numpy-stl) to load the liver capsule surface model into memory, and obtain the vertex and triangular facet information of the model for subsequent anatomical constraint processing.

[0080] Outlier removal: According to the liver capsule surface model, calculate the distance from each point in the three-dimensional point cloud to the liver surface. By writing a distance calculation algorithm, outliers with a distance greater than 20mm from the liver surface are removed. These outliers are usually caused by segmentation errors or other tissue interferences. Eliminating these points can effectively avoid the influence of interferences such as intestinal fluid accumulation on the ascites volume calculation.

[0081] 3. Voxel integral calculation: According to the voxel integral formula (where d irepresents the distance from the i-th voxel to the peritoneum, k = 0.2, d 0 = 5 mm. V represents the calculated volume of ascites. N represents the total number of voxels involved in the calculation. v i is the volume of the i-th voxel. k is a constant, taking the value of 0.2 in this formula. d 0 is a reference distance.) Calculate the volume of ascites. By traversing the three-dimensional point cloud after anatomical constraint processing, calculate the voxel volume corresponding to each point, and accumulate according to the formula to finally obtain the volume value of ascites.

[0082] V. Dynamic Monitoring and Result Output

[0083] 1. Time Series Alignment: Use the Dynamic Time Warping (DTW) algorithm to perform time series alignment on the ascites distributions at different time points. Implement this algorithm using the DTW algorithm library in Python (such as dtw-python). Take the ascites segmentation results (binary masks or three-dimensional point cloud data) collected at different time points as input. The DTW algorithm will find the optimal alignment path between the two time series to make them accurately correspond in time for subsequent accurate calculation of the volume change rate.

[0084] 2. Calculation of Volume Change Rate: According to the formula Calculate the volume change rate. Among them, ΔV represents the volume change rate of ascites, which reflects the speed of change of the ascites volume within a certain time interval and is an important indicator for evaluating the development of the disease or the treatment effect. V t is the ascites volume value measured at the current time point and is the current state data for calculating the volume change rate. V t-Δt represents the ascites volume measured at time point t - Δt. Δt is the time interval between the two measurement time points. V t-Δt as the starting state data for comparison, used to calculate the change in volume (V t -V t-Δt ).

[0085] Δt refers to the time interval between two measurements of the ascites volume. It determines the time span based on which the volume change rate is calculated.

[0086] This part is a time conversion factor. Since the unit of Δt may be different time units such as days and hours, to unify the time scale of the calculation results, convert the time interval to hours. For example, if Δt is in days, then multiply by (here the number of hours = 1) to convert the time interval to hours and make the unit of the volume change rate unified as "volume change per hour".

[0087] In the code implementation, obtain the ascites volume V at different time pointst and V t-Δt , and the time interval Δt, are calculated according to the formula to obtain the change rate of ascites volume over time, providing an important basis for clinicians to evaluate the development of the disease.

[0088] 3. Visualization output

[0089] Three-dimensional heat map generation: Use a visualization library (such as Mayavi) to display the area of ascites growth and decline in the form of a three-dimensional heat map. Map the ascites volume change data at different time points onto a three-dimensional model, and intuitively show the changes of ascites in space and time through color coding (red represents the area of newly added ascites, and blue represents the area of reduced ascites), helping doctors more intuitively understand the trend of disease changes.

[0090] PDF report generation: Use a report generation library in Python (such as reportlab) to automatically generate a PDF report. The report content includes ascites volume values, volume change trend curves (generated by plotting line graphs of volume over time), clinical advice threshold prompts (setting thresholds for ascites volume and change rate based on clinical experience and relevant research, and prompting doctors to pay attention to abnormal situations in the report), etc., providing a comprehensive and intuitive reference for clinical diagnosis and treatment decisions.

[0091] In one embodiment, models such as SegNet, Fully Convolutional Network (FCN), or Swin Transformer can be used to replace U-Net for semantic segmentation of abdominal CT images, and the corresponding network layers and loss functions are adapted. A multi-model fusion strategy can also be adopted, such as using U-Net and FCN to segment the same CT image successively, and then merging the output masks or taking the intersection in a weighted form to improve the robustness of the detection results.

[0092] In one embodiment, in the preprocessing stage, classical methods such as adaptive thresholding and Histogram distribution analysis can be combined with machine learning classifiers (such as random forest, SVM) to make a secondary judgment on the candidate regions. During post-processing, the watershed algorithm can be used to refine and fuse the low-density regions in three-dimensional space to reduce the missed segmentation of discontinuous ascites regions.

[0093] A lightweight network (such as MobileNet, Lightweight U-Net) can be deployed on an embedded device or an edge computing node in the hospital's local area network to achieve near-real-time segmentation and volume calculation, which is suitable for small primary medical institutions or mobile diagnosis scenarios. The entire AI segmentation process can also be fully migrated to the cloud platform, and the client is only responsible for uploading CT images and viewing the results, facilitating multi-institutional collaboration or remote diagnosis.

[0094] In one embodiment, for multi-modal images such as MRI and PET-CT in the same time period, the suspected ascites region in CT can be cross-validated by combining other sequence features, or a segmentation network with multi-channel input can be designed to automatically learn cross-modal features. In addition, a regression-based approach can be attempted, where the network directly outputs an estimated value of the "total ascites volume" to meet scenarios with higher speed requirements.

[0095] The present invention has the following advantages:

[0096] A semantic segmentation model based on deep learning is used to automatically identify and segment the ascites region in abdominal CT / MRI images. The convolutional neural network is used to accurately identify the ascites region, maintaining good performance in complex situations. The deep learning semantic segmentation strategy enables the model to learn and recognize complex patterns in CT or MRI images, more accurately distinguish ascites from other tissues or fluids with similar densities, reduce misclassification and missed classification, and improve the accuracy of ascites region recognition. The deep learning model has strong generalization ability and can adapt to imaging data from different patients, different devices, and different operating conditions, and can be effectively applied in a wide range of clinical environments.

[0097] The automated processing flow reduces the need for manual intervention, shortens the processing time, improves the standardization of the processing process, and helps improve clinical work efficiency and ensure result consistency. Preprocessing such as noise removal, gray-scale normalization, and resampling is automatically performed to improve the image quality and provide high-quality input data for the deep learning model.

[0098] Three-dimensional connected component analysis is used to identify and connect adjacent suspected ascites regions in each layer of slices, reducing missed classification, ensuring the continuity and integrity of the ascites region, and improving the volume estimation accuracy. The ascites volume is calculated by three-dimensional voxel statistics combined with the physical size of the voxels, providing an accurate volume calculation method.

[0099] Automated image processing and segmentation techniques reduce the workload of professionals, enabling non-experts to obtain fast and reliable analysis results and reducing the dependence on advanced professional skills.

[0100] Accurately and quickly quantifying the ascites volume helps evaluate the disease state, develop treatment plans, and monitor treatment effects. Especially in the management of chronic disease patients, it can help doctors more effectively manage ascites-related diseases and improve disease management efficiency. Provide intuitive ascites analysis results for clinicians, including ascites region visualization and volume data, to assist in rapid diagnosis and treatment decision-making.

[0101] In one embodiment of the present invention, taking the CT sequence of the venous phase of liver cirrhosis patients (slice thickness 1 mm, matrix 512×512) as an example, the preprocessing takes 32 seconds (including multi-phase registration and HU value adjustment), and the model inference time is 18 seconds per case (using the TensorRT acceleration engine). The output result is the ascites volume of 1523±15 ml (the gold standard manual measurement value: 1580 ml), and the 7-day volume change rate is -9.7% (the error from the clinical puncture drainage volume record is <3%).

[0102] The technical solution of the present invention is realized by the cooperation of a hardware platform and software modules.

[0103] The hardware platform consists of the following:

[0104] Medical imaging acquisition device: A 64-slice spiral CT scanner or an MRI scanner is used to obtain abdominal sequence images in DICOM format.

[0105] Data processing server: A medical imaging workstation equipped with an NVIDIA A100 GPU, running a deep learning inference engine, providing powerful computing capabilities for complex image analysis and model operations.

[0106] Terminal display device: A medical-grade color monitor with a resolution of ≥2048×2560 is used to ensure clear display of the processing results and image information, facilitating doctors' observation and analysis.

[0107] The interaction process of the software modules is as follows:

[0108] Data input layer: Receives the original CT / MRI data, supports multi-phase image input, and provides a data basis for subsequent processing.

[0109] Preprocessing module: Performs image registration, HU value calibration, and ROI extraction in sequence. Among them, non-rigid B-spline registration algorithm is used for image registration, and the optimization function is (where λ = 0.8, the number of iterations ≤ 50 times, and the registration accuracy reaches 0.3 mm); HU value calibration includes linearly mapping the original CT value (-50~150 HU) to the [0, 255] gray scale interval to enhance the ascites characteristics, applying morphological opening operation (kernel size 3×3) to the area where the CT value < -10 HU to suppress fat interference; using the N4ITK algorithm to correct the bias field of the MRI image, and setting the number of iterations to 200 times to eliminate low-frequency intensity inhomogeneity.

[0110] Intelligent Segmentation Module: Deploy an improved 3D Attention U-Net model. The encoder of this model has 4 levels of downsampling, each level containing 2 3×3×3 convolutions and a spatial attention module; the decoder has 4 levels of upsampling, adopting a channel attention gating mechanism; by introducing a multi-scale feature fusion unit (MSFF), multi-scale features are fused using the formula (weights w i dynamically generated by 1×1×1 convolution) to fuse multi-scale features. During training, anatomical constraint Dice Loss and edge-sensitive Focal Loss (γ = 2) are used as loss functions, and data augmentation of online generation of elastic deformation (σ = 10, α = 15) and random mirror flipping is performed.

[0111] Volume Calculation Core Module: First, convert the binary mask of the segmentation result into a 3D point cloud (point spacing 0.5mm), load a preset liver capsule surface model (STL format), remove abnormal points more than 20mm away from the liver surface to eliminate the interference of intestinal fluid accumulation, and then use the formula (where d i represents the distance from the i-th voxel to the peritoneum, k = 0.2, d 0 = 5mm) for voxel integration calculation.

[0112] Dynamic Monitoring Interface: Adopt the dynamic time warping (DTW) algorithm to match the ascites distribution at different time points, calculate the volume change rate through the formula display the growth and decline area of ascites with a 3D heat map (red for new, blue for reduction), and automatically generate a PDF report (including volume value, change trend curve, clinical recommendation threshold prompt).

[0113] Example 2:

[0114] Reference Figure 4 , a liver ascites volume calculation system based on artificial intelligence semantic segmentation, the system includes:

[0115] Data Acquisition Module, used to obtain medical image data of the patient's abdomen;

[0116] Preprocessing Module, used to preprocess the medical image data;

[0117] Intelligent Segmentation Module, used to input the preprocessed data into a deep learning model for intelligent segmentation to obtain the segmentation result of the ascites area;

[0118] Volume Calculation Module, used to calculate the liver ascites volume according to the segmentation result;

[0119] An analysis and output module is used to analyze the distribution of ascites in the liver at different time points and visually output the analysis results. This system modularizes the entire calculation method of liver ascites volume. Each module has a clear division of labor and works collaboratively to achieve the full-process automated processing from data collection to result output, improving the calculation efficiency and accuracy, and providing a convenient and reliable tool for clinical applications.

[0120] Embodiment 3:

[0121] In the third embodiment of the present invention, based on the same inventive concept, a computer-readable storage medium is proposed. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for calculating the volume of liver ascites based on artificial intelligence semantic segmentation in the above embodiment are implemented.

[0122] Embodiment 4

[0123] In the fourth embodiment of the present invention, based on the same inventive concept, a computer device is proposed, including: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory and execute the method for calculating the volume of liver ascites based on artificial intelligence semantic segmentation in the above embodiment.

[0124] It should be understood that each part of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits with suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0125] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for calculating liver ascites volume based on artificial intelligence semantic segmentation, characterized in that: The following steps are involved: Acquire the patient's abdominal imaging data through medical imaging acquisition equipment; Preprocessing the abdominal image data, including image registration, HU value calibration and ROI extraction; The preprocessed data is input into the deep learning model for intelligent segmentation to obtain the segmentation results of the ascites area; Calculating the ascites volume according to the segmentation result; The ascites distribution at different time points was processed and the analysis results were visualized and output.

2. The method for calculating liver ascites volume based on artificial intelligence semantic segmentation according to claim 1, characterized in that: The medical image acquisition device is a 64-row spiral CT machine or an MRI scanner, and the acquired image data are abdominal sequence images in DICOM format.

3. The method for calculating liver ascites volume based on artificial intelligence semantic segmentation according to claim 1, characterized in that: The image registration adopts a non-rigid B-spline registration algorithm, the HU value calibration includes linear mapping of the original CT value and applying morphological operations to a specific area, and the N4ITK algorithm is used to perform bias field correction on the MRI image.

4. The method for calculating liver ascites volume based on artificial intelligence semantic segmentation according to claim 1, characterized in that: The deep learning model is an improved 3D Attention U-Net model. The encoder of the model includes multi-level downsampling and spatial attention modules, the decoder includes multi-level upsampling and channel attention gating mechanism, and multi-scale feature fusion is performed through jump connections.

5. The method for calculating liver ascites volume based on artificial intelligence semantic segmentation according to claim 4, characterized in that: When training the improved 3D Attention U-Net model, anatomically constrained DiceLoss and edge-sensitive FocalLoss are used as loss functions, and data enhancement is performed.

6. The method for calculating liver ascites volume based on artificial intelligence semantic segmentation according to claim 1, characterized in that: When calculating the ascites volume, the binary mask of the segmentation result is first converted into a three-dimensional point cloud, the preset liver capsule surface model is loaded, abnormal points that are more than a certain distance from the liver surface are removed, and then the ascites volume is calculated using the voxel integration formula.

7. The method for calculating liver ascites volume based on artificial intelligence semantic segmentation according to claim 5, characterized in that: The time series of ascites distribution at different time points were aligned, and the volume change rate was calculated using the dynamic time warping algorithm; The visual output includes displaying the ascites growth and decline area in a three-dimensional heat map and automatically generating a PDF report containing ascites volume values, change trend curves, and clinically recommended threshold prompts; In the data preprocessing stage, adaptive threshold, classical method based on histogram distribution analysis and machine learning classifier are used to make secondary judgments on candidate regions; During post-processing, the watershed algorithm is used to refine the segmentation and fusion of low-density areas in the dimensional space.

8. A system for calculating liver ascites volume based on artificial intelligence semantic segmentation, characterized in that: The method for calculating the volume of hepatic ascites based on artificial intelligence semantic segmentation according to any one of claims 1 to 7, wherein the system comprises: A data acquisition module, used to obtain medical imaging data of the patient's abdomen; A preprocessing module, used for preprocessing the medical image data; An intelligent segmentation module is used to input the preprocessed data into a deep learning model for intelligent segmentation to obtain the segmentation result of the ascites area; A volume calculation module, used for calculating the volume of liver ascites according to the segmentation result; The analysis and output module is used to analyze the distribution of liver ascites at different time points and visualize the analysis results.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for calculating the volume of ascites based on artificial intelligence semantic segmentation as described in any one of claims 1 to 7 is implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for calculating the volume of ascites based on artificial intelligence semantic segmentation as described in any one of claims 1 to 7 is implemented.

Images