Portal vein medical image segmentation system and method based on automatic parameter configuration

Abstract

The invention discloses a portal vein medical image segmentation system and method based on automatic parameter configuration, and relates to the field of medical image processing, artificial intelligence and computer-aided diagnosis. Multi-center medical image key statistical features are obtained through a data set feature extraction module, and an automatic parameter configuration module automatically generates optimal network training parameters based on the features without manual intervention; and then model training is completed through a network training module, key parameters are recorded by an empirical parameter output module for cross-center migration, and a three-dimensional segmentation result and a visual interface are output by a prediction visualization module. According to the method, a full-intelligent process from data input to segmentation output is realized, the method is adaptive to multi-center and multi-mode medical images, the segmentation precision is high, the robustness is high, the method can be widely applied to liver operation planning, portal hypertension evaluation, liver cancer diagnosis and treatment monitoring, and the clinical diagnosis efficiency and consistency are greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, artificial intelligence, and computer-aided diagnosis. Specifically, it relates to a portal vein medical image segmentation system and method based on automated parameter configuration. This system can achieve high-precision automatic segmentation and three-dimensional reconstruction of the portal vein in multi-center, multimodal medical images such as enhanced computed tomography (CT) or magnetic resonance imaging (MRI), and is widely used in liver surgery planning, portal hypertension assessment, and liver cancer diagnosis and treatment monitoring. Background Technology

[0002] The portal vein is the main blood vessel supplying the liver, collecting blood from the spleen, pancreas, small intestine, and stomach, and is crucial for maintaining liver metabolism and detoxification functions. Its morphology, course, and hemodynamic changes are closely related to the development of diseases such as cirrhosis, portal hypertension, and hepatocellular carcinoma (HCC), and are important bases for clinical treatment planning. Therefore, accurate imaging assessment of the portal vein has significant clinical value in disease diagnosis, efficacy evaluation, and prognosis prediction. Especially in primary liver cancer, portal vein invasion (such as portal vein tumor thrombus, PVTT) significantly increases the risk of tumor recurrence and metastasis, and limits the choice of treatment methods such as surgical resection, interventional therapy, and radiotherapy. Accurate acquisition of the three-dimensional anatomical structure of the portal vein and its branches is crucial for surgical planning, liver transplant donor evaluation, and liver function reserve assessment.

[0003] However, current imaging assessment of the portal vein mainly relies on enhanced CT or MRI scans, as well as manual segmentation and measurement by radiologists, which has the following prominent problems:

[0004] 1. High time and labor costs: Manually outlining the three-dimensional structure of the portal vein requires annotating multiple images one by one, which takes several hours or even longer, placing a heavy burden on clinicians;

[0005] 2. High subjectivity: The segmentation results are affected by the physician's experience, fatigue level and operating habits, resulting in poor repeatability and consistency, making it difficult to meet the needs of clinical quantitative research;

[0006] 3. Diverse data sources: The scanning parameters, resolution, grayscale distribution, and noise levels of different hospitals and imaging equipment vary, making it difficult to generalize fixed-parameter algorithms to real-world multi-center clinical data;

[0007] 4. Complex vascular features: The portal vein has a slender, tree-like branching structure, with small diameter and low contrast in the distal branches. It is also similar in gray to the surrounding liver parenchyma and adjacent vessels, making segmentation difficult.

[0008] 5. Limitations of traditional algorithms: Traditional methods based on thresholding, region growing, or simple deep networks usually require manual parameter tuning, are not sensitive to small blood vessel branches, and are prone to segmentation discontinuities, artifacts, or missed detections.

[0009] In recent years, deep learning (especially convolutional neural networks, CNNs) has made significant breakthroughs in the field of medical image segmentation. nnU-Net, as a self-configuring network framework, can automatically adjust data preprocessing, network structure, and training parameters, and has performed excellently in several international medical segmentation challenges, demonstrating its versatility in multimodal imaging tasks. However, directly applying nnU-Net to portal vein segmentation still faces several challenges:

[0010] 1. Insufficient task specificity: The portal vein forms a long, thin, tree-like vascular network with extremely complex branching patterns;

[0011] 2. The distal branch pipe has a very small diameter and blurred boundaries, which places extremely high demands on feature extraction and continuity modeling;

[0012] 3. Low image contrast: The signal intensity of the portal vein is often similar to that of the surrounding liver parenchyma and adjacent vessels (especially in small or distal branches), which increases the risk of confusion, missed detection and discontinuous segmentation.

[0013] 4. Limited high-quality training data: Portal vein 3D annotation requires professional radiologists to meticulously track vascular structures slice by slice. Public datasets with reliable labels are scarce, which limits large-scale training and reduces the ability to generalize across centers.

[0014] With the increasing demand for precision medicine and automated analysis, there is an urgent need to develop a portal vein segmentation system that can adapt to the characteristics of multi-center data, automatically generate training parameters, and provide interpretable results. This system can significantly reduce the workload of radiologists, improve diagnostic efficiency, and provide valuable objective evidence for early detection of liver diseases, surgical planning, and prognostic assessment, laying the technical foundation for future intelligent imaging diagnosis and personalized treatment. Summary of the Invention

[0015] This invention aims to overcome the problems of low efficiency of manual annotation, complex parameter tuning, and poor model generalization ability in existing portal vein segmentation technology. It proposes an nnU-Net portal vein segmentation system and method based on automated parameter configuration, which realizes automatic extraction of dataset features from enhanced CT or MRI images from multiple centers and multiple devices, generates the optimal training scheme, completes high-precision automatic segmentation and three-dimensional reconstruction of the portal vein, and constructs a fully intelligent process.

[0016] Specific technical solution:

[0017] The portal vein medical image segmentation system based on automated parameter configuration includes a dataset feature extraction module, an automated parameter configuration module, a network training module, an empirical parameter output module, and a prediction visualization module. These modules communicate via an internal data pipeline to achieve information exchange and collaborative operation.

[0018] Dataset feature extraction module: used to automatically read multi-center 3D medical image data, extract key statistical features such as median shape, imaging modality, intensity distribution, spatial resolution distribution, and sample size, and generate feature reports;

[0019] The automated parameter configuration module is used to automatically generate network patch size, batch size, network topology, image resampling strategy and intensity normalization method based on dataset features and using rule-driven algorithms. At the same time, it completes network configuration by combining fixed parameters, including optimizer, loss function, learning rate and data augmentation strategy.

[0020] Network training module: Used to train the nnU-Net model with automatically generated parameters, employing five-fold cross-validation and a hybrid loss function to improve the segmentation accuracy of small branches and low-contrast regions;

[0021] Empirical parameter output module: used to record key empirical parameters during the training process and form a parameter file. The key empirical parameters include Dice coefficient, optimal learning rate, convergence curve and optimal patching strategy.

[0022] Prediction visualization module: It is used to receive the trained model and new image data, automatically output the portal vein 3D segmentation mask and vessel volume, and generate an interactive 3D visualization interface.

[0023] Furthermore, the multicenter three-dimensional medical imaging data is enhanced computed tomography or magnetic resonance imaging data.

[0024] The data augmentation strategies include rotation, scaling, mirroring, Gaussian noise, Gaussian blur, gamma correction, and low-resolution simulation.

[0025] The portal vein medical image segmentation method based on automated parameter configuration, using the aforementioned portal vein medical image segmentation system based on automated parameter configuration, includes the following steps:

[0026] S1: Read multi-center 3D medical image data through the dataset feature extraction module, extract key statistical features such as median shape, imaging modality, intensity distribution, spatial resolution distribution, and sample size, and generate a feature report;

[0027] S2: Based on the feature report, the automated parameter configuration module automatically generates network patch size, batch size, network topology, image resampling strategy and intensity normalization method using a rule-driven algorithm, and completes network configuration by combining fixed parameters.

[0028] S3: Train the nnU-Net model using the parameters generated in step S2 through the network training module, and use five-fold cross-validation and a hybrid loss function for training;

[0029] S4: Record key empirical parameters during the training process through the empirical parameter output module to form a parameter file;

[0030] S5: Receives the trained model and new image data through the prediction visualization module, automatically outputs the portal vein 3D segmentation mask and vessel volume, and generates an interactive 3D visualization interface.

[0031] Furthermore, the multicenter three-dimensional medical imaging data mentioned in step S1 is enhanced CT or MRI imaging data.

[0032] The fixed parameters mentioned in step S2 include optimizer, loss function, learning rate, and data augmentation strategy. The data augmentation strategy includes rotation, scaling, mirroring, Gaussian noise, Gaussian blur, gamma correction, and low-resolution simulation.

[0033] The key empirical parameters mentioned in step S4 include the Dice coefficient, the optimal learning rate, the convergence curve, and the optimal patching strategy. The parameter file is used for cross-center transfer and model fine-tuning.

[0034] The interactive 3D visualization interface described in step S5 is used in clinical scenarios for preoperative planning or treatment assessment.

[0035] Compared with the prior art, the present invention has the following significant advantages:

[0036] 1. Fully automated process: From feature extraction to network training and prediction, no manual parameter tuning is required throughout the entire process, greatly reducing deployment complexity and alleviating the workload of clinicians;

[0037] 2. High precision and robustness: Automated parameter configuration combined with diverse data augmentation strategies can effectively identify small vascular branches and low-contrast areas, reducing segmentation discontinuities, artifacts, and missed detections.

[0038] 3. Cross-center adaptability: Empirical parameter outputs provide direct guidance for quickly adapting to data from different hospitals and imaging equipment, improving the model's generalization ability;

[0039] 4. Clinical interpretability: Outputs three-dimensional visualization results, providing a reliable basis for liver surgery planning, portal hypertension staging, and liver cancer risk assessment, and assisting in the formulation of clinical treatment plans.

[0040] Through the above technical solutions, this invention achieves precise segmentation and three-dimensional reconstruction of the portal vein in multi-center, multi-modal medical images, significantly improving the efficiency and consistency of clinical diagnosis and providing strong support for precision medicine. Detailed Implementation

[0041] The specific technical solutions of the present invention will be described with reference to the embodiments.

[0042] The portal vein medical image segmentation system based on automated parameter configuration includes a dataset feature extraction module, an automated parameter configuration module, a network training module, an empirical parameter output module, and a prediction visualization module. These modules communicate via an internal data pipeline to achieve information exchange and collaborative operation.

[0043] Dataset Feature Extraction Module: Automatically reads multi-center 3D medical image data, extracts key statistical features such as median shape, imaging modality, intensity distribution, spatial resolution distribution, and sample size, and generates a feature report to guide subsequent parameter configuration. The multi-center 3D medical image data consists of enhanced CT or MRI images.

[0044] The automated parameter configuration module automatically generates network patch size, batch size, network topology, image resampling strategy, and intensity normalization method based on dataset features and a rule-driven algorithm. It also completes network configuration using fixed parameters, requiring no manual intervention. These fixed parameters include the optimizer, loss function, learning rate, and data augmentation strategies, specifically rotation, scaling, mirroring, Gaussian noise, Gaussian blur, gamma correction, and low-resolution simulation.

[0045] Network training module: The nnU-Net model is trained using automatically generated parameters, and five-fold cross-validation and a hybrid loss function are employed to improve the segmentation accuracy of small branches and low-contrast regions.

[0046] Empirical parameter output module: Records key empirical parameters during the training process, including Dice coefficients, optimal learning rate, convergence curve, and best patching strategy, forming a parameter file for cross-center transfer and model fine-tuning.

[0047] Predictive visualization module: Receives the trained model and new image data, automatically outputs the portal vein 3D segmentation mask and vessel volume, and generates an interactive 3D visualization interface, which can be directly used in clinical scenarios such as preoperative planning or treatment assessment.

[0048] Using the above system, a portal vein medical image segmentation method based on automated parameter configuration includes the following steps:

[0049] S1: Read multi-center 3D medical image data through the dataset feature extraction module, extract key statistical features such as median shape, imaging modality, intensity distribution, spatial resolution distribution, and sample size, and generate a feature report;

[0050] The multicenter three-dimensional medical imaging data are enhanced CT or MRI imaging data.

[0051] S2: Based on the feature report, the automated parameter configuration module automatically generates network patch size, batch size, network topology, image resampling strategy and intensity normalization method using a rule-driven algorithm, and completes network configuration by combining fixed parameters.

[0052] The fixed parameters include optimizer, loss function, learning rate, and data augmentation strategies, including rotation, scaling, mirroring, Gaussian noise, Gaussian blur, gamma correction, and low-resolution simulation.

[0053] S3: Train the nnU-Net model using the parameters generated in step S2 through the network training module, and use five-fold cross-validation and a hybrid loss function for training;

[0054] S4: Record key empirical parameters during the training process through the empirical parameter output module to form a parameter file;

[0055] The key empirical parameters include the Dice coefficient, optimal learning rate, convergence curve, and best patching strategy. The parameter file is used for cross-center transfer and model fine-tuning.

[0056] S5: Receives the trained model and new image data through the prediction visualization module, automatically outputs the portal vein 3D segmentation mask and vessel volume, and generates an interactive 3D visualization interface.

Claims

1. A portal vein medical image segmentation system based on automated parameter configuration, characterized in that, It includes a dataset feature extraction module, an automated parameter configuration module, a network training module, an empirical parameter output module, and a prediction visualization module. These modules communicate via an internal data pipeline to achieve information exchange and collaborative operation. Dataset feature extraction module: used to automatically read multi-center 3D medical image data, extract key statistical features such as median shape, imaging modality, intensity distribution, spatial resolution distribution, and sample size, and generate feature reports; The automated parameter configuration module is used to automatically generate network patch size, batch size, network topology, image resampling strategy and intensity normalization method based on dataset features and using rule-driven algorithms. At the same time, it completes network configuration by combining fixed parameters, including optimizer, loss function, learning rate and data augmentation strategy. Network training module: Used to train the nnU-Net model with automatically generated parameters, employing five-fold cross-validation and a hybrid loss function to improve the segmentation accuracy of small branches and low-contrast regions; Empirical parameter output module: used to record key empirical parameters during the training process and form a parameter file. The key empirical parameters include Dice coefficient, optimal learning rate, convergence curve and optimal patching strategy. Prediction visualization module: It is used to receive the trained model and new image data, automatically output the portal vein 3D segmentation mask and vessel volume, and generate an interactive 3D visualization interface.

2. The portal vein medical image segmentation system based on automated parameter configuration according to claim 1, characterized in that, The multicenter three-dimensional medical imaging data are enhanced computed tomography or magnetic resonance imaging data.

3. The portal vein medical image segmentation system based on automated parameter configuration according to claim 1, characterized in that, The data augmentation strategies include rotation, scaling, mirroring, Gaussian noise, Gaussian blur, gamma correction, and low-resolution simulation.

4. A portal vein medical image segmentation method based on automated parameter configuration, characterized in that, The portal vein medical image segmentation system based on automated parameter configuration as described in any one of claims 1 to 3, wherein the method comprises the following steps: S1: Read multi-center 3D medical image data through the dataset feature extraction module, extract key statistical features such as median shape, imaging modality, intensity distribution, spatial resolution distribution, and sample size, and generate a feature report; S2: Based on the feature report, the automated parameter configuration module automatically generates network patch size, batch size, network topology, image resampling strategy and intensity normalization method using a rule-driven algorithm, and completes network configuration by combining fixed parameters. S3: Train the nnU-Net model using the parameters generated in step S2 through the network training module, and use five-fold cross-validation and a hybrid loss function for training; S4: Record key empirical parameters during the training process through the empirical parameter output module to form a parameter file; S5: Receives the trained model and new image data through the prediction visualization module, automatically outputs the portal vein 3D segmentation mask and vessel volume, and generates an interactive 3D visualization interface.

5. The portal vein medical image segmentation method based on automated parameter configuration according to claim 4, characterized in that, The multicenter three-dimensional medical imaging data mentioned in step S1 is enhanced CT or MRI imaging data.

6. The portal vein medical image segmentation method based on automated parameter configuration according to claim 4, characterized in that, The fixed parameters mentioned in step S2 include optimizer, loss function, learning rate, and data augmentation strategy. The data augmentation strategy includes rotation, scaling, mirroring, Gaussian noise, Gaussian blur, gamma correction, and low-resolution simulation.

7. The portal vein medical image segmentation method based on automated parameter configuration according to claim 4, characterized in that, The key empirical parameters mentioned in step S4 include the Dice coefficient, the optimal learning rate, the convergence curve, and the optimal patching strategy. The parameter file is used for cross-center transfer and model fine-tuning.