An image segmentation method for evaluating a neutron therapy dose in hepatocellular carcinoma

Abstract

The application discloses an image segmentation method for evaluating a neutron treatment dose in hepatocellular carcinoma, and relates to the technical field of neutron treatment. The application trains a 3D U-Net GAN network model through a medical image dataset, alternately trains a generator model and a discriminator model in the 3D U-Net GAN network model by using real images and false images randomly generated by the generator model, so that the generator model after training can generate a segmentation result which is more accurate and detailed for the medical image, the discriminator model after training can more accurately evaluate the authenticity of the output of the generator model, and the network after training can accurately reflect the anatomical structure of the patient itself when segmenting the medical image, 10 The distribution of B in the body and the radionuclide kinetics are favorable for fully displaying information in the medical image accurately and conveniently.

Description

Technical Field

[0001] This invention relates to the field of neutron therapy technology, and in particular to an image segmentation method, apparatus, computer device, and storage medium for evaluating neutron therapy doses in hepatocellular carcinoma. Background Technology

[0002] Boron neutron capture therapy (BNCT) utilizes the boron neutrons that occur in cancer cells. 10 B(n, α) 7 The energy released from the Li nuclear reaction kills cancer cells; this process generates alpha particles with extremely high linear energy transfer (LET) and 7 Li recoil nuclei, with average energies of 190 keV / μm and 160 keV / μm respectively, can destroy diseased cells in the human body while having an extremely short range.

[0003] Hepatocellular carcinoma (HCC) is an aggressive primary liver cancer, accounting for more than 80% of primary liver cancers worldwide. In 2018, it ranked 6th in global malignant tumor incidence and 4th in cancer-related deaths. In China, HCC ranked 4th in malignant tumor incidence and 3rd in cancer-related deaths in 2015. Common treatments for HCC include surgery, liver transplantation, local treatments (radiofrequency ablation, microwave therapy, and ethanol injection), transarterial chemoembolization, radiotherapy, and supportive care. Compared to traditional therapies, neutron therapy, with its relatively short range, can more precisely destroy cancer cells, thereby reducing damage to surrounding normal tissues.

[0004] However, since it is impossible to directly measure the metrological parameters inside the human body, current methods typically rely on manually dividing tissues and organs in medical images to support subsequent analysis. However, this approach struggles to reflect the individual anatomical structures of different patients. 10 The distribution of B in the body and the dynamics of radionuclides make it difficult to fully display the information in medical images. Summary of the Invention

[0005] This invention provides an image segmentation method for evaluating neutron therapy dose in hepatocellular carcinoma, which addresses the shortcomings of existing methods that fail to reflect the individual anatomical structures of different patients. 10 The distribution of B in the body and the dynamics of radionuclides make it difficult to fully display information in medical images.

[0006] This invention provides an image segmentation method for evaluating neutron therapy dose in hepatocellular carcinoma, comprising the following steps:

[0007] Obtain a medical image dataset of hepatocellular carcinoma regions;

[0008] Construct a 3D U-Net generator model and a 3D CNN discriminator model, and couple the 3D U-Net generator model and the 3D CNN discriminator model into a 3D U-Net GAN network model;

[0009] The 3D U-Net GAN network model was trained using a medical image dataset. Real images from the medical image dataset and randomly generated fake images from the generator model were used as inputs to the 3D U-Net GAN network model. Preliminary segmentation results were obtained through forward propagation in the 3D U-Net GAN network model. The losses of the generator model and the discriminator model in the 3D U-Net GAN network model were obtained according to the loss function. Based on the preliminary segmentation results and the losses of the generator model and the discriminator model, the parameters of the 3D U-Net GAN network model were updated through backpropagation. At the same time, the discriminator model was fixed to train the generator model, and the generator model was fixed to train the discriminator model, resulting in the trained 3D U-Net GAN network model.

[0010] Medical images are input into a trained 3D U-Net GAN network model. The generator model within the trained 3D U-Net GAN network model is used to extract features from the liver and tumor regions of the medical images, obtaining structural feature maps containing three-dimensional structure and morphological structure, and radioactive element feature maps containing the distribution of radioactive elements and radionuclides, respectively. The discriminator model is used to filter and remove false features from the structural feature maps and radioactive element feature maps of the liver and tumor regions. The remaining true features in the structural feature maps and radioactive element feature maps of the liver and tumor regions are summed and then visualized to obtain segmented images of the liver and tumor regions, respectively. The segmented images of the liver and tumor regions are then concatenated to obtain the segmented image.

[0011] Preferably, the acquisition of the medical image dataset includes:

[0012] Acquire medical images of hepatocellular carcinoma regions;

[0013] Medical images were processed using Gaussian filtering, median filtering, bilateral filtering, histogram equalization, adaptive histogram equalization, contrast stretching, and geometric transformation to obtain a medical image dataset of the hepatocellular carcinoma region.

[0014] To enhance the data in a medical image dataset, random rotation, horizontal flipping, vertical flipping, random cropping, random brightness adjustment, or random contrast adjustment are performed.

[0015] Preferably, the construction of the 3D U-Net generator model and the 3D CNN discriminator model includes:

[0016] The 3D U-Net generator model adds multiple convolutional and pooling layers between the encoder and decoder within the model to adapt to 3D data.

[0017] The 3D U-Net generator model generates realistic samples from random noise or other inputs, and learns the distribution and features of the sample data to generate new samples similar to real samples.

[0018] The 3D CNN discriminator model uses different network structures to distinguish between real samples and generated new samples similar to real samples.

[0019] Preferably, the forward propagation algorithm includes:

[0020] The forward propagation of the generator part is as follows: the input data is passed through a series of neural network layers to gradually increase the dimension of the input data and generate an image. The generator usually consists of multiple layers, including fully connected or convolutional layers, activation functions, and upsampling layers. In some variants, the generator may contain skip connections. Finally, the generator outputs the generated image through an activation function.

[0021] The forward propagation of the discriminator part is as follows: it simultaneously receives real images and fake images generated by the generator as input, extracts features and classifies them through a series of neural network layers, and gradually reduces the spatial dimension of the image through pooling layers or straddle convolutions to extract more abstract features. The final output is the discriminator's prediction of whether the input image is real or fake.

[0022] Preferably, the fixed discriminator model trains the generator model, and then the fixed generator model trains the discriminator model, so that the discriminator model and the generator model compete with each other to improve the recognition rate.

[0023] This invention also provides an image segmentation apparatus for evaluating neutron therapy doses in hepatocellular carcinoma, comprising:

[0024] The image module is used to acquire medical image datasets of hepatocellular carcinoma regions;

[0025] The network model module is used to construct a 3D U-Net generator model and a 3D CNN discriminator model, and to couple the 3D U-Net generator model and the 3D CNN discriminator model into a 3D U-Net GAN network model.

[0026] The 3D U-Net GAN network model was trained using a medical image dataset. Real images from the medical image dataset and randomly generated fake images from the generator model were used as inputs to the 3D U-Net GAN network model. Preliminary segmentation results were obtained through forward propagation in the 3D U-Net GAN network model. The losses of the generator model and the discriminator model in the 3D U-Net GAN network model were obtained according to the loss function. Based on the preliminary segmentation results and the losses of the generator model and the discriminator model, the parameters of the 3D U-Net GAN network model were updated through backpropagation. At the same time, the discriminator model was fixed to train the generator model, and the generator model was fixed to train the discriminator model, resulting in the trained 3D U-Net GAN network model.

[0027] The segmentation module is used to input medical images into a trained 3D U-Net GAN network model. The generator model within the trained 3D U-Net GAN network model extracts features from the liver and tumor portions of the medical images, obtaining structural feature maps containing three-dimensional structure and morphology, and radioactive element feature maps containing the distribution and radionuclides of radioactive elements, respectively. A discriminator model filters and removes false features from the structural and radioactive element feature maps of the liver and tumor portions. The remaining true features in the structural and radioactive element feature maps of the liver and tumor portions are summed and then visualized to obtain segmented images of the liver and tumor portions. The segmented images of the liver and tumor portions are then concatenated to obtain the segmented image.

[0028] This invention also provides an electronic device, including a memory and a processor;

[0029] The memory is used to store computer programs;

[0030] When the processor executes the computer program stored in the memory, it implements the steps of an image segmentation method for evaluating neutron therapy doses for hepatocellular carcinoma as described above.

[0031] This invention also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of an image segmentation method for evaluating neutron therapy doses for hepatocellular carcinoma as described above.

[0032] This invention provides an image segmentation method for evaluating neutron therapy dose in hepatocellular carcinoma, which has the following advantages compared to existing technologies:

[0033] This invention trains a 3D U-Net GAN network model using a medical image dataset. It alternately trains the generator and discriminator models within the 3D U-Net GAN network using real images and randomly generated fake images from the generator model. This allows the trained generator model to produce more accurate and detailed segmentation results for medical images, while the trained discriminator model can more precisely evaluate the realism of the generator model's output. The trained network can accurately reflect the patient's anatomical structure when segmenting medical images. 10 The distribution of B in the body and the kinetics of radionuclides enable it to accurately and fully display information in medical images, facilitating its application. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of a medical image preprocessing and segmentation process based on 3D U-Net GAN, provided as an embodiment of the present invention, for evaluating the neutron therapy dose in hepatocellular carcinoma.

[0035] Figure 2 This is a flowchart illustrating a Monte Carlo dosing software interface method for evaluating neutron therapy dose in hepatocellular carcinoma, provided in an embodiment of the present invention.

[0036] Figure 3 This is a schematic diagram of the structure of a 3D U-Net GAN for image segmentation method used to evaluate neutron therapy dose in hepatocellular carcinoma, provided in an embodiment of the present invention.

[0037] Figure 4 This is a schematic diagram of the patient's liver CT image segmentation result provided by an image segmentation method for evaluating neutron therapy dose in hepatocellular carcinoma according to an embodiment of the present invention; wherein (a) is a CT image of the patient's liver region; wherein (b) is the extracted CT image of the patient's liver; and wherein (c) is the segmentation result of the patient's liver.

[0038] Figure 5 A schematic diagram of a patient liver voxel model provided in this embodiment of the invention for an image segmentation method to evaluate neutron therapy dose in hepatocellular carcinoma;

[0039] Figure 6 This is a schematic diagram of the patient's liver energy deposition results, provided by an embodiment of the present invention, for evaluating the dose of neutron therapy for hepatocellular carcinoma. Detailed Implementation

[0040] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0041] See Figures 1-2 This invention provides an image segmentation method for evaluating neutron therapy dose in hepatocellular carcinoma, including a deep learning-based medical image segmentation method and a Monte Carlo software interface method.

[0042] Specifically, the deep learning-based medical image segmentation method includes the following steps:

[0043] S1: Read images from a medical image dataset of hepatocellular carcinoma patients and convert them into a specific format for easier subsequent processing.

[0044] S2: Preprocess the image after it has been converted to a specific format, including noise reduction and contrast enhancement, to ensure the accuracy and stability of the subsequent segmentation process.

[0045] S3: Perform data augmentation on the preprocessed image, including but not limited to rotation, flipping, scaling, etc., to increase data diversity and improve the model's generalization ability.

[0046] S4: Construct the generator model.

[0047] S5: Construct the discriminator model.

[0048] S6: Define the loss function.

[0049] S7: Train the 3D U-Net GAN model using the prepared training dataset, and iteratively optimize the model parameters to better adapt it to segmentation tasks of complex structures such as liver and tumors.

[0050] S8: Use the trained 3D U-Net GAN model to segment the patient's liver and tumor.

[0051] S9: Outputs the segmented image to provide doctors with a visual reference to assist them in making more accurate diagnoses and treatment plans.

[0052] Specifically, the Monte Carlo software interface method includes the following steps:

[0053] S1: Establish an individualized voxel model of the patient based on the segmented images.

[0054] S2: Based on medical imaging technology and chemical analysis methods, determine the parameters required for Monte Carlo software calculations, such as the density and elemental composition of tissues and organs.

[0055] S3: Converts the user-input model's geometric, physical, and chemical information into the input file format of the Monte Carlo software.

[0056] S4: Call Monte Carlo software to calculate the dose.

[0057] S5: Read the data provided by the radiation monitoring system to determine the total number of neutron nuclear reactions.

[0058] S6: Displays the absorbed dose calculation results and plots related images.

[0059] A specific embodiment of the deep learning-based medical image segmentation method of the present invention comprises the following steps:

[0060] Step S1: Read images from the medical image dataset of hepatocellular carcinoma patients and convert them into a specific format for subsequent processing.

[0061] In step S1, the PyDicom library in Python is used to read medical images in Dicom format, or the pynrrd library is used to read medical images in nnrd format, the h5py library is used to read medical images in h5 format, and the PIL library is used to read medical images in other formats. The NumPy library in Python is used to convert the medical images that have been read into the specific format it provides for subsequent steps.

[0062] Step S2: Preprocess the image converted to a specific format, including noise reduction and contrast enhancement, to ensure the accuracy and stability of the subsequent segmentation process.

[0063] In step S2, image preprocessing is performed using the OpenCV, Scikit-image, and PIL libraries. This includes applying Gaussian filtering, median filtering, or bilateral filtering techniques to effectively reduce noise in the image and improve the accuracy of subsequent processing steps; applying histogram equalization, adaptive histogram equalization, and contrast stretching techniques to enhance image contrast and make different regions in the image easier to distinguish, thereby contributing to the accuracy of the segmentation process; and applying geometric transformations to adjust image size and crop the image to optimize the image data, making it more suitable for subsequent processing and analysis, and improving the performance and accuracy of the model.

[0064] Step S3: Perform data augmentation on the preprocessed image, including but not limited to rotation, flipping, scaling, etc., to increase data diversity and improve the model's generalization ability.

[0065] In step S3, the Albumentations library in Python is used to perform data augmentation operations on the preprocessed image, including random rotation, horizontal flipping, vertical flipping, random cropping and resizing, and random adjustment of brightness and contrast. This increases the diversity and robustness of the data samples, improves the generalization ability of the model, and reduces the overfitting phenomenon that may occur during training.

[0066] Step S4: Construct the generator model.

[0067] In step S4, a 3D U-Net network model is built based on the PyTorch library of the Python language. It adds some convolutional and pooling layers between the encoder and decoder to adapt to 3D data. The generator needs to generate realistic samples from random noise or other inputs and generate new samples similar to real samples by learning the distribution and features of the data.

[0068] Step S5: Construct the discriminator model.

[0069] In step S5, the PyTorch library of Python constructs a 3D CNN model, using different network structures to distinguish between real and generated samples. The discriminator needs to focus more on the classification energy and authenticity judgment of the input samples, rather than just performing feature extraction and image reconstruction. The input of the discriminator is the segmentation result generated by the generator or the real label, and the output is a scalar value (0 represents a generated fake sample, and 1 represents a real sample).

[0070] Step S6: Define the loss function.

[0071] In step S6, loss functions provided by the PyTorch library are called, such as mean squared error loss function, cross-entropy loss function, binary cross-entropy loss function, weighted binary cross-entropy loss function, etc. In addition, custom loss functions are also supported for training neural networks.

[0072] Step S7: Train the 3D U-Net GAN model using the prepared training dataset. The structure of the 3D U-Net GAN model is as follows: Figure 3 As shown, by iteratively optimizing the model parameters, it can be better adapted to segmentation tasks of complex structures such as the liver and tumors.

[0073] In step S7, the real image and the fake image generated by the generator are used as inputs. The generated segmentation result is obtained through forward propagation. The loss of the generator and the discriminator is calculated according to the loss function. Then, directional propagation and parameter updates are performed. An alternating training strategy is adopted. First, the discriminator is fixed to train the generator, and then the generator is fixed to train the discriminator, so that the two models compete with each other and gradually improve performance. During this process, the use of CUDA acceleration is selected based on the actual hardware device. If it is true, CUDA acceleration is used. If it is false, CUDA acceleration is not used, and the CPU is used directly for computation.

[0074] Step S8: Use the trained 3D U-Net GAN model to segment the patient's liver and tumor portions.

[0075] In step S8, the preprocessed 3D medical image is passed as input to the 3D U-Net GAN model. Through the forward propagation process, the corresponding segmentation results are generated based on the features and distribution information of the input image. These results can be regarded as predictions of structural boundaries or regions of interest in the image.

[0076] Step S9: Output the segmented image to provide doctors with a visual reference to assist them in making more accurate diagnoses and treatment plans.

[0077] In step S9, the OpenCV and PIL libraries in Python are used to output segmented images in different formats, such as JPEG, PNG, BMP, and TIFF.

[0078] In the encoder part, the spatial dimension of the image is gradually reduced through convolutional layers, activation functions, and pooling layers, while feature information is extracted. For each neuron in the layer (starting from the first hidden layer), its input value is calculated, which usually involves the following operations: calculating a weighted sum, adding a bias, and applying an activation function.

[0079] In the decoder section: the spatial dimension of the image is gradually restored through deconvolutional layers and skip connections, while the segmentation result is generated by combining the feature information extracted by the encoder. For each neuron in the layer (starting from the first hidden layer), its input value is calculated, which usually involves the following operations: calculating the weighted sum, adding the bias, and applying the activation function.

[0080] The forward propagation of the generator part involves progressively increasing the dimensionality of the input data through a series of neural network layers (usually convolutional layers, activation functions, and upsampling layers) to generate an image. The generator typically consists of multiple layers, including but not limited to: fully connected or convolutional layers, activation functions, and upsampling layers. In some variations, the generator may contain skip connections. Finally, the generator outputs the generated image through an activation function.

[0081] The forward propagation of the discriminator part receives both real images and fake images generated by the generator as input. It extracts features and classifies them through a series of neural network layers (usually convolutional layers and activation functions). The spatial dimension of the image is gradually reduced through pooling layers or straddle convolutions to extract more abstract features. The final output is the discriminator's prediction of whether the input image is real or fake.

[0082] During forward propagation, various regularization techniques, such as dropout and weight decay, may be applied to prevent overfitting and improve the model's generalization ability.

[0083] A specific embodiment of the Monte Carlo software interface method of the present invention comprises the following steps:

[0084] Step S1: Establish an individualized voxel model of the patient based on the segmented images.

[0085] In step S1, a voxel model consisting of one or more hexahedrons is established, such as... Figure 5 As shown, it is used to represent the anatomical structure of a liver cancer patient for subsequent Monte Carlo calculations.

[0086] Step S2: Based on medical imaging technology and chemical analysis methods, determine the parameters required for Monte Carlo software calculations, such as the density and elemental composition of tissues, organs, and lesions.

[0087] In step S2, medical imaging techniques such as X-rays, CT scans, and MRI are used. Figure 4 The images shown are segmentation results of a patient's liver CT image. (a) is a CT image of the patient's liver region; (b) is a CT image of the patient's liver extracted from the CT image of the patient's liver region; and (c) is the segmentation result of the patient's liver. These techniques can provide information about density and elemental composition. Chemical analysis methods, such as atomic absorption spectrometry, atomic emission spectrometry, mass spectrometry, metabolomics, and proteomics, usually require the collection of tissue samples and laboratory analysis to determine the elemental composition and content in tissues or organs.

[0088] Step S3: Convert the user-input model geometric, physical, and chemical information into the input file format of the Monte Carlo software.

[0089] In step S3, the geometric information of the model, such as the length, width, and height of voxels reflecting geometric structure, and spatial information such as the distance between the radioactive source and the patient's target area; the physical information of the model, such as particle type, source strength of the radioactive source, beam intensity, emission angle, energy spectrum, and density of tissues and organs reflecting physical parameters; and the chemical information of the model, including the elemental composition of tissues and organs reflecting chemical parameters.

[0090] Step S4: Call the Monte Carlo software to calculate the dose.

[0091] In step S4, Monte Carlo software is called to calculate the dose, such as MCNP, Geant4, GATE, etc.

[0092] Step S5: Read the data provided by the radiation monitoring system to determine the total number of neutron nuclear reactions.

[0093] In step S5, data from relevant radiation monitoring instruments such as neutron measuring instruments, activation detectors, thermoluminescent dosimeters, and gamma-ray detectors are input to determine the total number of neutron nuclear reactions.

[0094] Step S6: Display the dosimetric calculation results and draw related images.

[0095] In step S6, the relevant dosimetry calculation results are plotted using the Matplotlib library in Python, such as... Figure 6 As shown.

[0096] In this embodiment, in determining the parameters required for Monte Carlo software calculations, such as the density and elemental composition of tissues, organs, and lesions, in addition to medical imaging techniques and chemical analysis methods, data such as International Commission on Radiological Protection Publication No. 110 and databases can also be used.

[0097] In this embodiment, Monte Carlo software is a general-purpose software package based on the Monte Carlo method for calculating various particle transport problems; examples of Monte Carlo software include MCNP, Geant4, and GATE. Using Monte Carlo software to perform particle transport calculations during treatment, the average dose value of the lesion area in a hepatocellular carcinoma patient during a single nuclear conversion is obtained. The calculation formula is as follows:

[0098]

[0099] Where: D represents the average dose value; the average dose value represents the magnitude of the average ionization energy dE absorbed per unit mass of material dm under the action of ionizing radiation; this method supports calling Monte Carlo software to calculate the average ionization energy of the lesion area during particle transport, and calculates the average dose value based on the mass of the lesion area obtained from the voxel model.

[0100] In this embodiment, according to the definition of the Medical Internal Radiation Dose Committee, the absorbed dose of the target area can be expressed as:

[0101]

[0102] Where: D T The absorbance in the target region is represented by ; S represents the average absorbance in the target region during one nuclear transformation (calculated using Monte Carlo software); This represents the time integral activity.

[0103] In this invention, the 3D U-Net GAN model incorporates the idea of ​​Generative Adversarial Networks (GANs), enabling the model to better preserve the realism of images during segmentation. The generator model generates more accurate and detailed segmentation results, while the discriminator model evaluates the realism of the generator model's output. This adversarial training encourages the generator model to continuously optimize, thereby improving the accuracy of segmentation. The segmented images can provide accurate anatomical information for subsequent dosimetric calculations.

[0104] This invention combines Monte Carlo simulation technology and medical image segmentation results to quickly and accurately establish individualized dose calculation models for patients. The model information, physical information, and chemical information input by the user are converted into input files for Monte Carlo software, facilitating dose calculation. The absorbed dose calculation results are displayed and related images are drawn, providing doctors and patients with intuitive dose assessment results.

[0105] This invention also provides an image segmentation apparatus for evaluating neutron therapy doses in hepatocellular carcinoma, comprising:

[0106] The image module is used to acquire medical image datasets of hepatocellular carcinoma regions.

[0107] The network model module is used to construct a 3D U-Net generator model and a 3D CNN discriminator model, and to couple the 3D U-Net generator model and the 3D CNN discriminator model into a 3D U-Net GAN network model.

[0108] The 3D U-Net GAN network model was trained using a medical image dataset. Real images from the dataset and randomly generated fake images from the generator model were used as inputs to the 3D U-Net GAN network model. Preliminary segmentation results were obtained through forward propagation. The losses of the generator and discriminator models within the 3D U-Net GAN network model were calculated using the loss function. Based on the preliminary segmentation results and the losses of the generator and discriminator models, the parameters of the 3D U-Net GAN network model were updated through backpropagation. Simultaneously, the discriminator model was fixed while training the generator model, and the generator model was fixed while training the discriminator model, resulting in the trained 3D U-Net GAN network model.

[0109] The segmentation module is used to input medical images into a trained 3D U-Net GAN network model. The generator model within the trained 3D U-Net GAN network model extracts features from the liver and tumor portions of the medical images, obtaining structural feature maps containing three-dimensional structure and morphology, and radioactive element feature maps containing the distribution and radionuclides of radioactive elements, respectively. A discriminator model filters and removes false features from the structural and radioactive element feature maps of the liver and tumor portions. The remaining true features in the structural and radioactive element feature maps of the liver and tumor portions are summed and then visualized to obtain segmented images of the liver and tumor portions. The segmented images of the liver and tumor portions are then concatenated to obtain the segmented image.

[0110] This invention also provides an electronic device, including a memory and a processor.

[0111] Memory is used to store computer programs.

[0112] When the processor executes a computer program stored in memory, it implements the steps described above for an image segmentation method for evaluating neutron therapy doses in hepatocellular carcinoma.

[0113] This invention also provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the steps of the image segmentation method described above for evaluating neutron therapy doses for hepatocellular carcinoma.

[0114] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An image segmentation method for evaluating neutron therapy dose in hepatocellular carcinoma, characterized in that, Includes the following steps: Obtain a medical image dataset of hepatocellular carcinoma regions; Construct a 3D U-Net generator model and a 3D CNN discriminator model, and couple the 3D U-Net generator model and the 3D CNN discriminator model into a 3D U-Net GAN network model; The 3D U-Net GAN network model was trained using a medical image dataset. Real images from the medical image dataset and randomly generated fake images from the generator model were used as inputs to the 3D U-Net GAN network model. Preliminary segmentation results were obtained through forward propagation in the 3D U-Net GAN network model. The losses of the generator model and the discriminator model in the 3D U-Net GAN network model were obtained according to the loss function. Based on the preliminary segmentation results and the losses of the generator model and the discriminator model, the parameters of the 3D U-Net GAN network model were updated through backpropagation. At the same time, the discriminator model was fixed to train the generator model, and the generator model was fixed to train the discriminator model, resulting in the trained 3D U-Net GAN network model. The medical image is input into the trained 3D U-Net GAN network model. The generator model in the trained 3D U-Net GAN network model is used to extract features from the liver and tumor parts in the medical image. The structural feature maps of the liver and tumor parts, which include the three-dimensional structure and morphological structure, and the radioactive element feature maps, which include the distribution of radioactive elements and radionuclides, are obtained respectively. The discriminator model is used to filter and remove false features in the structural feature maps and radiometric feature maps of the liver and tumor parts. The remaining true features in the structural feature maps and radiometric feature maps of the liver and tumor parts are summed and then visualized to obtain segmented images of the liver and tumor parts respectively. The segmented images of the liver and tumor parts are then concatenated to obtain the segmented image.

2. The image segmentation method for evaluating neutron therapy dosage in hepatocellular carcinoma according to claim 1, characterized in that, The acquisition of the medical image dataset includes: Acquire medical images of hepatocellular carcinoma regions; Medical images were processed using Gaussian filtering, median filtering, bilateral filtering, histogram equalization, adaptive histogram equalization, contrast stretching, and geometric transformation to obtain a medical image dataset of the hepatocellular carcinoma region. To enhance the data in a medical image dataset, random rotation, horizontal flipping, vertical flipping, random cropping, random brightness adjustment, or random contrast adjustment are performed.

3. The image segmentation method for evaluating neutron therapy dosage in hepatocellular carcinoma according to claim 1, characterized in that, The construction of the 3D U-Net generator model and the 3D CNN discriminator model includes: The 3D U-Net generator model adds multiple convolutional and pooling layers between the encoder and decoder within the model to adapt to 3D data. The 3D U-Net generator model generates realistic samples from random noise or other inputs, and learns the distribution and features of the sample data to generate new samples similar to real samples. The 3D CNN discriminator model uses different network structures to distinguish between real samples and generated new samples similar to real samples.

4. The image segmentation method for evaluating neutron therapy dosage in hepatocellular carcinoma according to claim 1, characterized in that, The forward propagation algorithm includes: The forward propagation of the generator part is as follows: the input data is passed through a series of neural network layers to gradually increase the dimension of the input data and generate an image. The generator usually consists of multiple layers, including fully connected or convolutional layers, activation functions, and upsampling layers. In some variants, the generator may contain skip connections. Finally, the generator outputs the generated image through an activation function. The forward propagation of the discriminator part is as follows: it simultaneously receives real images and fake images generated by the generator as input, extracts features and classifies them through a series of neural network layers, and gradually reduces the spatial dimension of the image through pooling layers or straddle convolutions to extract more abstract features. The final output is the discriminator's prediction of whether the input image is real or fake.

5. The image segmentation method for evaluating neutron therapy dosage in hepatocellular carcinoma according to claim 1, characterized in that, The discriminator model is fixed to train the generator model, and then the generator model is fixed to train the discriminator model, so that the discriminator model and the generator model compete with each other to improve the recognition rate.

6. An image segmentation apparatus for evaluating neutron therapy dose in hepatocellular carcinoma, characterized in that, include: The image module is used to acquire medical image datasets of hepatocellular carcinoma regions; The network model module is used to construct a 3D U-Net generator model and a 3D CNN discriminator model, and to couple the 3D U-Net generator model and the 3D CNN discriminator model into a 3D U-Net GAN network model. The 3D U-Net GAN network model was trained using a medical image dataset. Real images from the medical image dataset and randomly generated fake images from the generator model were used as inputs to the 3D U-Net GAN network model. Preliminary segmentation results were obtained through forward propagation in the 3D U-Net GAN network model. The losses of the generator model and the discriminator model in the 3D U-Net GAN network model were obtained according to the loss function. Based on the preliminary segmentation results and the losses of the generator model and the discriminator model, the parameters of the 3D U-Net GAN network model were updated through backpropagation. At the same time, the discriminator model was fixed to train the generator model, and the generator model was fixed to train the discriminator model, resulting in the trained 3D U-Net GAN network model. The segmentation module is used to input medical images into the trained 3D U-Net GAN network model. The generator model in the trained 3D U-Net GAN network model is used to extract features from the liver and tumor parts in the medical images, and to obtain structural feature maps of the liver and tumor parts that include three-dimensional structure and morphological structure, and radioactive element feature maps that include the distribution of radioactive elements and radionuclides. The discriminator model is used to filter and remove false features in the structural feature maps and radiometric feature maps of the liver and tumor parts. The remaining true features in the structural feature maps and radiometric feature maps of the liver and tumor parts are summed and then visualized to obtain segmented images of the liver and tumor parts respectively. The segmented images of the liver and tumor parts are then concatenated to obtain the segmented image.

7. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of the image segmentation method for evaluating neutron therapy dose for hepatocellular carcinoma as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the steps of an image segmentation method for evaluating neutron therapy doses for hepatocellular carcinoma as described in any one of claims 1 to 5.

Images