Automatic Segmentation Method, System, Device and Storage Medium for Nasopharyngeal Carcinoma PET Tumors

Abstract

A method, system, device and storage medium for automatic segmentation of nasopharyngeal carcinoma PET tumors. The automatic segmentation method of nasopharyngeal carcinoma PET tumors obtains the final prediction probability of each pixel point and the corresponding segmentation image through 6 steps based on the object's PET image. The integrated model proposed by the automatic segmentation method of nasopharyngeal carcinoma PET tumors in the present invention shows better generalization ability compared with single models and non-machine learning methods. Compared with the original single 3D U-Net segmentation method, the DSC, Jaccard and ASSD evaluation indexes of the segmentation image in the present invention are respectively improved by 7.8%, 10.8% and 24.5%, and the segmentation effect is better.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and particularly to an automatic segmentation method, system, device and storage medium for nasopharyngeal carcinoma PET tumors. Background Art

[0002] At present, the delineation of nasopharyngeal carcinoma PET tumors is mainly completed manually by doctors. This manual delineation method has the following problems: 1. Manual delineation is difficult and extremely time-consuming. Doctors need to check the tomographic images of patients layer by layer and delineate the contours in each layer of images; 2. Manual delineation highly depends on doctors' experience, and there are significant differences in the delineation results of doctors with different experiences. Since PET is a functional imaging that can depict the metabolic intensity of tissues, the lesion areas usually show high brightness in PET images. In current nasopharyngeal carcinoma PET / CT images, it is difficult to define the lesions. The spatial resolution of PET images is low, and it is affected by the partial volume effect and artifacts generated by scattered photons. Some non-adjacent tissues with high SUV values may appear adjacent and are difficult to distinguish. Limited resolution usually leads to blurred lesion boundaries, making it difficult to define the lesion boundaries during the segmentation process. In addition, nasopharyngeal carcinoma lesions rarely show observable boundaries in CT (except those invading the sinuses). The most commonly used methods include: 2.5SUV threshold, linear adaptive SUV threshold function method, and local maximum uptake value. However, these methods are limited by simple thresholds and usually result in inaccurate tumor delineation. In recent years, more advanced methods have been proposed one after another, such as the Markov random field method that simultaneously utilizes the gradients and intensities of PET and CT. The methods based on region growing, edge information, and texture analysis are relatively simple and suitable for delineating small lesions, but they still have insufficient segmentation capabilities for adjacent and overlapping regions. The active contour method realizes segmentation by minimizing the energy function to evolve the contour. Although these methods have certain improvements in noise compatibility, they are still not robust enough to the volume changes and uncertain factors between different tumors.

[0003] The PET-CT joint segmentation method constructs a model using the gray-scale distribution and gradient of pixels. Its basic assumption is that pixels belonging to the same tissue have consistent gray-scale values in PET and CT images. However, this type of method highly depends on manually selected seed points. The main difference between different models lies in the way of defining the penalty function. Although the model can be optimized by methods such as the maximum flow minimum cut algorithm or the random walk algorithm, overall, the joint segmentation method still depends on prior information.

[0004] In recent years, the application of machine learning methods in the fields of image reconstruction and image analysis has become increasingly widespread. Machine learning can be divided into traditional learning and deep learning. Traditional machine learning methods train classifiers by extracting manually defined features from images and complete image segmentation at the pixel level. Deep learning methods directly obtain segmentation maps in an end-to-end manner. At the same time, some semantic segmentation methods, including the Fully Convolutional Network (FCN), SegNet, and DenseNet, have been widely used in medical images. Although deep learning methods have problems such as lack of interpretability and the black box effect.

[0005] However, in the segmentation of medical images by deep learning models, the performance of the model mainly depends on the selection of hyperparameters. Traditional hyperparameters include the depth of the convolutional layer, the number of features, the size of the convolutional kernel, the learning rate, etc. Generally speaking, the more learnable parameters there are, the higher the overall performance of the model, but at the same time, the model complexity will increase, which is likely to lead to overfitting.

[0006] Therefore, in view of the deficiencies of the existing technology, it is very necessary to provide a method, system, device, and storage medium for automatic segmentation of nasopharyngeal carcinoma PET tumors to solve the deficiencies of the existing technology. Summary of the Invention

[0007] The purpose of the present invention is to avoid the deficiencies of the existing technology and provide a method for automatic segmentation of nasopharyngeal carcinoma PET tumors. The method for automatic segmentation of nasopharyngeal carcinoma PET tumors has better generalization ability and higher segmentation accuracy.

[0008] The above object of the present invention is achieved by the following technical measures:

[0009] Provide a method for automatic segmentation of nasopharyngeal carcinoma PET tumors, which is obtained by the following steps:

[0010] S1. Obtain the PET image of the object;

[0011] S2. Reconstruct the PET image obtained in S1 to obtain a three-dimensional PET reconstructed image;

[0012] S3. Convert the PET reconstructed image obtained in S2 into an SUV image;

[0013] S4. Adjust the resolution of the SUV image obtained in S3 to obtain a PET-SUV image;

[0014] S5. Input the PET-SUV image obtained in S4 into an integrated model composed of a V-NET model and a U-NET model, and then predict each pixel point in the PET-SUV image. Multiple first prediction probabilities and multiple second prediction probabilities are obtained corresponding to each pixel point, and the integrated model is composed of multiple trained V-NET models and multiple trained U-NET models;

[0015] S6. Weight the multiple first prediction probabilities and multiple second prediction probabilities of each pixel point obtained in S5 to obtain the final prediction probability of each pixel point and the corresponding segmentation image.

[0016] Preferably, during the training process of the above V-NET model and U-NET model, the loss value is calculated through a combined loss function composed of a weighted cross-entropy loss function and a Dice loss function.

[0017] Preferably, S5 is to input the PET-SUV image obtained in S4 into each trained V-NET model to obtain the first prediction probability of each pixel point; input the PET-SUV image obtained in S4 into each trained U-NET model to obtain the second prediction probability of each pixel point. The number of the first prediction probabilities of each pixel point is the same as the number of trained V-NET models, and the number of the second prediction probabilities of each pixel point is the same as the number of trained U-NET models.

[0018] Preferably, the above combined loss function is represented by formula (1):

[0019]

[0020] Among them, is the weighted cross-entropy loss function term, is the Dice loss function term, L is the loss value, α is a constraint term, and 0 < α < 1, β is the weight of the foreground, and 0 < β < 1, N is the number of voxels, i is a voxel, p i is the prediction probability when voxel i is the foreground, t i is the annotation of voxel i in the labeled PET-SUV image, and when i is the foreground, t i is 1, and when i is the background, t i is 0.

[0021] Each of the V-NET models of the present invention is trained by the following steps:

[0022] A1. Randomly select PET-SUV images in the database; the database is set with multiple pairs of image pairs, and each group of image pairs consists of a PET-SUV image and a labeled PET-SUV image corresponding to the PET-SUV image;

[0023] A2. Input the PET-SUV image selected in A1 into the V-NET model to obtain the prediction probability of each pixel point;

[0024] A3. Calculate the prediction probability of each pixel point obtained in A2 and the annotation loss value of the annotated PET-SUV image through the combined loss function;

[0025] A4. Determine whether the number of training times has reached the set number of times. If not, go to A5; if so, go to A6, where the set number of times is greater than 200;

[0026] A5. Update the parameters of the V-NET model and return to A1;

[0027] A6. End the training and use the V-NET model with the current parameters as the trained V-NET model.

[0028] Each of the U-NET models of the present invention is trained by the following steps:

[0029] B1. Randomly select a PET-SUV image from the database;

[0030] B2. Input the PET-SUV image selected in B1 into the U-NET model to obtain the prediction probability of each pixel point;

[0031] B3. Calculate the prediction probability of each pixel point obtained in B2 and the annotation loss value of the annotated PET-SUV image through the combined loss function;

[0032] B4. Determine whether the number of training times has reached the set number of times. If not, go to B5; if so, go to A6, where the set number of times is greater than 200;

[0033] B5. Update the parameters of the U-NET model and return to B1;

[0034] B6. End the training and use the U-NET model with the current parameters as the trained U-NET model.

[0035] Preferably, the construction method of the above database is as follows:

[0036] C1. Obtain multiple PET images;

[0037] C2. Reconstruct the PET images obtained in C1 to obtain three-dimensional PET reconstruction images;

[0038] C3. Convert the PET reconstruction images obtained in C2 into SUV images;

[0039] C4. Obtain a PET-SUV image by adjusting the resolution of the SUV image obtained in C3.

[0040] C5. Annotate each pixel point in the PET-SUV image obtained in C4 to obtain the annotated PET-SUV image.

[0041] In S6, take the average of the multiple first prediction probabilities and multiple second prediction probabilities of each pixel point obtained in S5 to obtain the final prediction probability of each pixel point and the corresponding segmentation image.

[0042] Preferably, in S2 or C2 above, specifically, reconstruct the PET image obtained in S1 by the OSEM algorithm to obtain a three-dimensional PET reconstruction image in DICOM format.

[0043] In S3 or C3, calculate the SUV value of each pixel point in the PET reconstruction image by formula (2), and the SUV values of all pixel points constitute the SUV image.

[0044]

[0045] Where the unit of the SUV value is g / mL, tissue activity is the tracer activity of the diseased tissue per unit volume, the unit is Bq / mL, injected dose is the injection dose of the imaging agent and the unit is Bq, and body weight is the body weight and the unit is g.

[0046] In S4 or C4, adjust the resolution of the SUV image to 512×512 to obtain a PET-SUV image in NiFTI format.

[0047] Preferably, the activation function of the above U-NET model is the ReLU function, ELU function, SELU function or LeakyReLU function, and the learning rate lr ranges from le -4 ~le -2 The L2 regularization parameter λ ranges from le -25 ~le -9, the discard probability d ranges from 0 to 0.3, the number of convolutional layers ranges from 1 to 4, the number of features ranges from 16 to 40, the convolutional kernel size is 3×3×3 or 5×5×5, the initialization method is Xavier uniform, Xavier normal, Kaiming uniform or Kaiming normal, the upsampling method is max-pooling, average pooling or convolution, and the downsampling method is transpose convolution, nearest up-sampling or trilinear up-sampling.

[0048] Preferably, the activation function of the above V-NET model is the ReLU function, ELU function, SELU function or LeakyReLU function, and the learning rate lr ranges from le -4 ~le -2 , the L2 regularization parameter λ ranges from le -25 ~le -9 , the discard probability d ranges from 0 to 0.3, the number of convolutional layers ranges from 1 to 4, the number of features ranges from 12 to 32, and the convolutional kernel size is 3×3×3 or 5×5×5.

[0049] Preferably, the above integrated model is composed of 6 trained V-NET models and 6 trained U-NET models.

[0050] The second object of the present invention is to provide a nasopharyngeal carcinoma PET tumor automatic segmentation system to avoid the deficiencies of the prior art. The nasopharyngeal carcinoma PET tumor automatic segmentation system has higher segmentation accuracy for nasopharyngeal carcinoma PET tumors.

[0051] The above object of the present invention is achieved by the following technical measures:

[0052] Provide a nasopharyngeal carcinoma PET tumor automatic segmentation system, which is provided with:

[0053] Dataset module - obtain PET-SUV images and corresponding annotated PET-SUV images to construct a training set;

[0054] Model training module - train the V-NET model and U-NET model through the training set to obtain an integrated model;

[0055] Image processing module - reconstruct the PET image to be segmented to obtain a three-dimensional PET reconstructed image; then convert the PET reconstructed image into an SUV image, and finally adjust the resolution of the SUV image to obtain the PET-SUV image to be segmented;

[0056] Image segmentation module - Input the PET-SUV image to be segmented into the integrated model and output the segmentation result.

[0057] The third object of the present invention is to provide a nasopharyngeal carcinoma PET tumor automatic segmentation device to avoid the deficiencies of the prior art. The nasopharyngeal carcinoma PET tumor automatic segmentation device has higher segmentation accuracy for nasopharyngeal carcinoma PET tumors.

[0058] The above object of the present invention is achieved by the following technical measures:

[0059] Provide a nasopharyngeal carcinoma PET tumor automatic segmentation device provided with a memory and a processor. Computer-executable instructions are stored on the memory, and when the processor runs the computer-executable instructions on the memory, the above-mentioned nasopharyngeal carcinoma PET tumor automatic segmentation method is implemented.

[0060] The fourth object of the present invention is to provide a computer-readable storage medium to avoid the deficiencies of the prior art. The computer-readable storage medium has higher segmentation accuracy for nasopharyngeal carcinoma PET tumors.

[0061] The above object of the present invention is achieved by the following technical measures:

[0062] Provide a computer-readable storage medium storing a computer program, which when executed by a processor, implements the above-mentioned nasopharyngeal carcinoma PET tumor automatic segmentation method.

[0063] A method, system, device and storage medium for automatic segmentation of nasopharyngeal carcinoma PET tumors according to the present invention, wherein the method for automatic segmentation of nasopharyngeal carcinoma PET tumors is obtained by the following steps: S1. Obtain the PET image of an object; S2. Reconstruct the PET image obtained in S1 to obtain a three-dimensional PET reconstructed image; S3. Convert the PET reconstructed image obtained in S2 into an SUV image; S4. Adjust the resolution of the SUV image obtained in S3 to obtain a PET-SUV image; S5. Input the PET-SUV image obtained in S4 into an integrated model composed of a V-NET model and a U-NET model, and then predict each pixel point in the PET-SUV image. Multiple first prediction probabilities and multiple second prediction probabilities are obtained corresponding to each pixel point, and the integrated model is composed of multiple trained V-NET models and multiple trained U-NET models; S6. Weight the multiple first prediction probabilities and multiple second prediction probabilities of each pixel point obtained in S5 to obtain the final prediction probability of each pixel point and the corresponding segmentation image. During the training process, both the V-NET model and the U-NET model calculate the loss value through a combined loss function composed of a weighted cross-entropy loss function and a Dice loss function. Compared with a single model and a non-machine learning method, the integrated model proposed by the present invention shows better generalization ability. Compared with using the original single 3D U-Net segmentation method, the DSC, Jaccard and ASSD evaluation indexes of the segmentation image of the present invention are respectively improved by 7.8%, 10.8% and 24.5%, and its segmentation effect is better. Description of the Drawings

[0064] The present invention is further described with reference to the accompanying drawings, but the content in the drawings does not constitute any limitation to the present invention.

[0065] Figure 1 It is a curve graph of the DSC evaluation index obtained by the present invention and each model.

[0066] Figure 2 It is a curve graph of the ASSD evaluation index obtained by the present invention and each model.

[0067] Figure 3 It is a segmentation image obtained by using a single U-Net and a Dice loss function for segmentation.

[0068] Figure 4 It is a segmentation image obtained by using a single V-Net and a Dice loss function for segmentation.

[0069] Figure 5 It is a segmentation image obtained by using traditional PET-CT combined segmentation.

[0070] Figure 6 It is a segmentation image obtained by using a single U-Net and a BCE loss function for segmentation.

[0071] Figure 7 The segmented image is segmented using a single V-Net and a BCE loss function.

[0072] Figure 8 The segmented image is segmented using the integrated model and combined loss function of the present invention. Detailed implementation manners

[0073] The technical solution of the present invention will be further described in conjunction with the following embodiments.

[0074] Embodiment 1

[0075] An automatic segmentation method for nasopharyngeal carcinoma PET tumors, as Figure 1 shown, is obtained by the following steps:

[0076] S1. Obtain the PET image of the object;

[0077] S2. Reconstruct the PET image obtained in S1 to obtain a three-dimensional PET reconstructed image;

[0078] S3. Convert the PET reconstructed image obtained in S2 into an SUV image;

[0079] S4. Adjust the resolution of the SUV image obtained in S3 to obtain a PET-SUV image;

[0080] S5. Input the PET-SUV image obtained in S4 into an integrated model composed of a V-NET model and a U-NET model, and then predict each pixel point in the PET-SUV image. Multiple first prediction probabilities and multiple second prediction probabilities are obtained corresponding to each pixel point, and the integrated model is composed of multiple trained V-NET models and multiple trained U-NET models; specifically, when the PET-SUV image obtained in S4 is input into each trained V-NET model, the first prediction probability of each pixel point is obtained; when the PET-SUV image obtained in S4 is input into each trained U-NET model, the second prediction probability of each pixel point is obtained. The number of first prediction probabilities of each pixel point is the same as the number of trained V-NET models, and the number of second prediction probabilities of each pixel point is the same as the number of trained U-NET models;

[0081] S6. Weight the multiple first prediction probabilities and multiple second prediction probabilities of each pixel point obtained in S5 to obtain the final prediction probability of each pixel point and the corresponding segmented image.

[0082] During the training process, both the V-NET model and the U-NET model calculate the loss value through a combined loss function composed of a weighted cross-entropy loss function and a Dice loss function. The specific combined loss function is represented by Equation (1):

[0083]

[0084] Among them, is the weighted cross-entropy loss function term, is the Dice loss function term, L is the loss value, α is the constraint term, and 0 < α < 1, β is the weight of the foreground, and 0 < β < 1, N is the number of voxels, i is the voxel, p i is the predicted probability when voxel i is the foreground, t i is the annotation of voxel i in the annotated PET-SUV image, and when i is the foreground, t i is 1, and when i is the background, t i is 0.

[0085] The combined loss function of the present invention combines the Dice loss function and the weighted cross-entropy loss function, and introduces two hyperparameters α and β in the Dice loss function term and the weighted cross-entropy loss function term. The purpose is to balance the weights between different tumor volumes and between false positives and false negatives.

[0086] Each V-NET model of the present invention is trained by the following steps:

[0087] A1. Randomly select PET-SUV images from the database; the database is set with multiple pairs of image pairs, and each group of image pairs consists of a PET-SUV image and an annotated PET-SUV image corresponding to the PET-SUV image;

[0088] A2. Input the PET-SUV image selected in A1 into the V-NET model to obtain the predicted probability of each pixel point;

[0089] A3. Calculate the predicted probability of each pixel point obtained in A2 and the annotation loss value of the annotated PET-SUV image through the combined loss function;

[0090] A4. Determine whether the number of training times has reached the set number. If not, go to A5; if so, go to A6, where the set number is greater than 200, and the set number is specifically 500 times;

[0091] A5. Update the parameters of the V-NET model and return to A1;

[0092] A6. End the training, and use the V-NET model with the current parameters as the trained V-NET model.

[0093] Each U-NET model of the present invention is trained by the following steps:

[0094] B1. Randomly select PET-SUV images from the database;

[0095] B2. Input the PET-SUV images selected in B1 into the U-NET model to obtain the prediction probability of each pixel point;

[0096] B3. Calculate the prediction probability of each pixel point obtained in B2 and the annotation loss value of the annotated PET-SUV image through the combined loss function;

[0097] B4. Determine whether the number of training times has reached the set number of times. If not, enter B5; if so, enter A6, where the set number of times is greater than 200, and the specific set number of times is 500 times;

[0098] B5. Update the parameters of the U-NET model and return to B1;

[0099] B6. End the training, and use the U-NET model with the current parameters as the trained U-NET model.

[0100] It should be noted that the V-NET model or U-NET model in A5 and B5 of the present invention can be tuned by the tree-structured Parzen estimation method (TPE), or the parameters of the V-NET model or U-NET model can be randomly updated. Moreover, the tree-structured Parzen estimation method is represented by the following formula:

[0101]

[0102] where χ is the domain to be searched, x is the set of hyperparameters, and f(x) is the evaluation function.

[0103] The construction method of the database of the present invention is as follows:

[0104] C1. Obtain multiple PET images;

[0105] C2. Reconstruct the PET images obtained in C1 to obtain three-dimensional PET reconstructed images;

[0106] C3. Convert the PET reconstructed images obtained in C2 into SUV images;

[0107] C4. Resolve the resolution of the SUV images obtained in C3 to obtain PET-SUV images;

[0108] C5. Annotate each pixel point in the PET-SUV images obtained in C4 to obtain annotated PET-SUV images.

[0109] It should be noted that in C5 of the present invention, each pixel point in the PET-SUV image is manually labeled. The labeling of the PET-SUV image of the present invention refers to outlining the PET tumor in the nasopharyngeal carcinoma in the PET-SUV image, that is, the outlined area is the tumor, and its voxels are all foreground, and other non-outlined areas are the background.

[0110] In S6, the average value of the multiple first prediction probabilities and multiple second prediction probabilities of each pixel point obtained in S5 is taken to obtain the final prediction probability of each pixel point and the corresponding segmentation image.

[0111] In S2 or C2, specifically, the PET image obtained in S1 is reconstructed by the OSEM algorithm to obtain a three-dimensional PET reconstructed image in DICOM format.

[0112] It should be noted that the reconstruction parameters of the OSEM algorithm of the present invention are set to 3 iterations and 21 subsets. The voxel size of the PET image is 4.07×4.07×5mm 3 , and its resolution is 200×200. The CT scan (80 mA, 120 KVp) is used for attenuation correction. The CT voxel size is 0.98×0.98×3mm 3 , and the resolution is 512×512. Therefore, in the present invention, through S4 or C4, the resolution of the PET reconstructed image is adjusted to the same resolution as the CT image.

[0113] In S3 or C3, the SUV value of each pixel point in the PET reconstructed image is calculated by formula (2) to obtain the SUV values of all pixel points to form an SUV image;

[0114]

[0115] where the unit of the SUV value is g / mL, tissue activity is the tracer activity of the diseased tissue per unit volume, the unit is Bq / mL, injected dose is the imaging agent injection dose and the unit is Bq, and body weight is the body weight and the unit is g.

[0116] It should be noted that the tracer activity of the diseased tissue per unit volume of the present invention is data obtained from the attenuation-corrected image by the software system, and the tracer activity of the diseased tissue per unit volume can be directly measured.

[0117] In S4 or C4, the resolution of the SUV image is adjusted to 512×512 to obtain a PET-SUV image in NiFTI format.

[0118] The activation function of the U-NET model is the ReLU function, ELU function, SELU function, or LeakyReLU function, and the learning rate lr ranges from le -4 ~le -2 The learning rate lr ranges from le -4 ~le -2 The L2 regularization parameter λ ranges from le -25 ~le -9 The dropout probability d is 0 to 0.3, the number of convolutional layers is 1 to 4, the number of features is 16 to 40, the convolutional kernel size is 3×3×3 or 5×5×5, the initialization method is Xavier uniform, Xavier normal, Kaiming uniform, or Kaiming normal, the upsampling method is max-pooling, average pooling, or convolution, and the downsampling method is transposeconvolution, nearest up-sampling, or trilinear up-sampling. The specific parameters of the U-NET model are shown in Table 1. The U-NET model of the present invention is selected within these parameter ranges.

[0119] Table 1. Parameters of the U-NET model

[0120]

[0121] The activation function of the V-NET model is the ReLU function, ELU function, SELU function, or LeakyReLU function, and the learning rate lr ranges from le -4 ~le -2 The L2 regularization parameter λ ranges from le -25 ~le -9 The dropout probability d is 0 to 0.3, the number of convolutional layers is 1 to 4, the number of features is 12 to 32, and the convolutional kernel size is 3×3×3 or 5×5×5. The specific parameters of the V-NET model are shown in Table 2. The V-NET model of the present invention is selected within these parameter ranges.

[0122] Table 2. Parameters of the V-NET model

[0123]

[0124] A method for automatic segmentation of nasopharyngeal carcinoma PET tumors. The proposed integrated model shows better generalization ability compared with single models and non-machine learning methods. Its segmentation effect is better than that of using the original single 3D U-Net or V-Net segmentation method.

[0125] Example 2

[0126] A method for automatic segmentation of nasopharyngeal carcinoma PET tumors, with other features being the same as those in Example 1. The integrated model in this example consists of 6 trained V-NET models and 6 trained U-NET models. The database in this example was retrospectively studied using 201 nasopharyngeal carcinoma subjects (average age: 47.3 ± 11.8 years, range 11–78 years; among them, 163 were male and 38 were female). All subjects were initially diagnosed as non-keratinizing undifferentiated nasopharyngeal carcinoma by histopathology. According to the AJCC standard (8th edition), they were divided into 22 cases of stage I–II and 179 cases of stage III–IV using the TNM staging system, for a total of 201 cases. Among them, 160 cases were used as the training set to train the model, and 41 cases were used as the test set to test. The subject data are shown in Table 3.

[0127] Table 3. Subject data

[0128]

[0129] This retrospective study was approved by the hospital ethics committee, and informed consent of the subjects was waived. All patients fasted for at least 6 hours before injection of the tracer. Imaging was performed 62 minutes (58 ± 5 minutes, range: 52–67 minutes) after intravenous injection of 306 - 468 MBq (8.27 - 12.65 mCi) of 18F-FDG (about 150 μCi / kg body weight). According to the procedure guidelines of the Society of Nuclear Medicine & Molecular Imaging (SNMMI), whole-body PET / CT scans were performed on a Siemens Biograph-128mCT scanner. In this example, segmentation was performed according to the method for automatic segmentation of nasopharyngeal carcinoma PET tumors in Example 1, and then the segmentation results were evaluated through the following steps. In this example, when the final predicted probability of a pixel point is less than 0.5, then this pixel point is judged as the background, that is, non-tumor.

[0130] In this example, a comparison was made between the present invention and a single V-NET model, a single U-NET model, and a traditional PET-CT combined segmentation model. Among them, the single V-NET model and U-NET model were trained through separate loss functions, such as the Dice loss function and the BCE loss function. The BCE loss function is the binary cross-entropy loss, which is represented by the following formula:

[0131]

[0132] The Dice loss function is represented by the following formula:

[0133]

[0134] The Dice loss function is differentiable, and its gradient calculation is expressed as follows:

[0135]

[0136] In the present invention, the manually delineated tumor region is used as the ground truth (gold standard) to evaluate the generalization performance of the model. The evaluation metrics can be divided into two categories: voxel-based metrics and surface-based metrics.

[0137] The voxel-based metrics for measuring the overlap between the segmented image and the ground truth, namely the Jaccard Coefficient (JC), DSC, and ASSD, are calculated using the following formulas respectively:

[0138]

[0139]

[0140]

[0141] Where |TP|, |FP|, and |FN| are the numbers of True Positive, False Positive, and False Negative respectively, and their values are between (0, 1). The higher the value, the more similar the segmentation value is to the ground truth. N X and N Y are the total number of pixels on the surfaces of the segmented image and the gold standard image respectively. The larger the values of JC and DSC, the more similar the two images are, while for ASSD, the closer it is to 1, the more similar.

[0142] The parameters of the 6 trained V-NET models, the 6 trained U-NET models, the parameters of the trained U-NET model using the Dice loss function, the parameters of the trained U-NET model using the BCE loss function, the parameters of the trained V-NET model using the Dice loss function, and the parameters of the trained V-NET model using the BCE loss function in this embodiment are shown in Table 3.

[0143]

[0144] The evaluation metrics of DSC, Jaccard, and ASSD for the 6 trained V-NET models, the 6 trained U-NET models, the integrated model of the present invention, and the control trained U-NET model using the Dice loss function, the trained U-NET model using the BCE loss function, the trained V-NET model using the Dice loss function, and the trained V-NET model using the BCE loss function in the present invention are shown in Table 4.

[0145] Table 4. Evaluation indicators of average DSC, Jaccard, and ASSD for each model

[0146] Segmentation method DSC Jaccard ASSD (mm) U-Net + Dice loss function 0.753±0.113 0.616±0.139 1.560±0.837 U-Net + BCE loss function 0.738±0.112 0.603±0.150 1.893±1.079 V-Net + Dice loss function 0.743±0.079 0.597±0.098 1.593±0.717 V-Net + BCE loss function 0.729±0.130 0.590±0.157 1.928±1.229 Joint segmentation 0.663±0.179 0.519±0.180 2.771±1.912 U-Net 1 + combined loss function 0.781±0.108 0.651±0.134 1.414±0.810 U-Net 2 + combined loss function 0.796±0.085 0.669±0.107 1.310±0.526 U-Net 3 + combined loss function 0.795±0.090 0.667±0.111 1.253±0.532 U-Net 4 + combined loss function 0.801±0.075 0.673±0.100 1.258±0.429 U-Net 5 + combined loss function 0.793±0.097 0.667±0.120 1.253±0.482 U-Net 6 + combined loss function 0.800±0.093 0.676±0.117 1.229±0.482 V-Net 1 + combined loss function 0.792±0.075 0.661±0.099 1.375±0.673 V-Net 2 + combined loss function 0.792±0.081 0.662±0.107 1.318±0.573 V-Net 3 + combined loss function 0.793±0.087 0.665±0.111 1.286±0.539 V-Net 4 + combined loss function 0.796±0.086 0.669±0.110 1.272±0.561 V-Net 5 + combined loss function 0.791±0.089 0.662±0.114 1.327±0.625 V-Net 6 + combined loss function 0.797±0.084 0.670±0.109 1.274±0.492 The integrated model of the present invention 0.812±0.084 0.683±0.109 1.195±0.508

[0147] As can be seen from Table 4, the DSC of the integrated model of the present invention is 0.812 ± 0.084, the Jaccard is 0.683 ± 0.109, and the ASSD is 1.195 ± 0.508. Compared with the original single 3D U-Net segmentation method, the performance of the present invention under the three evaluation indicators of DSC, Jaccard, and ASSD is far better than that of U-Net, V-Net, and traditional combined segmentation methods (p < 0.05). The accuracy of the segmentation images of the present invention is improved by 7.8%, 10.8%, and 24.5% respectively under the DSC, Jaccard, and ASSD evaluation indicators

[0148] while Figure 1 and Figure 2 are the evaluation indicators of DSC and ASSD corresponding to 41 test cases of each model, and the 41 tests are sorted from small to large according to the tumor volume size. The results show that for the vast majority of test cases, the segmentation accuracy of the present invention is better than that of other several segmentation methods. Generally speaking, the present invention not only shows a significant improvement in segmentation accuracy, but also shows consistency in the segmentation accuracy of tumors with different volume sizes

[0149] Figures 3 to 8 are the horizontal (left), sagittal (middle), and coronal (right) views of test case ID 20 under different segmentation methods, where the red shows the manually drawn contour and the green shows the automatically segmented contour. Through the segmentation images segmented by each model and the integrated model of the present invention, it can be seen from the figure that the traditional combined segmentation method cannot distinguish between tumors and normal tissues. The segmentation method proposed by the present invention distinguishes normal tissues from real tumors by comprehensively considering the outputs of all single models and does not require any post-processing. Therefore, the present invention obtains a more accurate segmentation result than other segmentation methods. An automatic segmentation method for nasopharyngeal carcinoma PET tumors of the present invention has better segmentation effect

[0150] Example 3

[0151] An automatic segmentation system for nasopharyngeal carcinoma PET tumors is provided with:

[0152] Dataset module - obtain PET-SUV images and corresponding annotated PET-SUV images to construct a training set

[0153] Model training module - train the V-NET model and U-NET model through the training set to obtain an integrated model

[0154] Image processing module - reconstruct the PET image to be segmented to obtain a three-dimensional PET reconstructed image; then convert the PET reconstructed image into an SUV image, and finally adjust the resolution of the SUV image to obtain the PET-SUV image to be segmented;

[0155] Image segmentation module - input the PET-SUV image to be segmented into the integrated model and output the segmentation result.

[0156] A nasopharyngeal carcinoma PET tumor automatic segmentation system according to this embodiment can execute a nasopharyngeal carcinoma PET tumor automatic segmentation method provided in Embodiment 1 or 2 of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0157] Embodiment 4

[0158] A nasopharyngeal carcinoma PET tumor automatic segmentation device is provided with a memory and a processor. A computer executable instruction is stored on the memory, and when the processor runs the computer executable instruction on the memory, a nasopharyngeal carcinoma PET tumor automatic segmentation method is implemented.

[0159] A nasopharyngeal carcinoma PET tumor automatic segmentation device according to this embodiment can execute a nasopharyngeal carcinoma PET tumor automatic segmentation method provided in Embodiment 1 or 2 of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0160] Embodiment 5

[0161] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the nasopharyngeal carcinoma PET tumor automatic segmentation method in the above Embodiment 1 or 2 is implemented.

[0162] A computer-readable storage medium according to this embodiment can execute a nasopharyngeal carcinoma PET tumor automatic segmentation method provided in the method embodiment of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the protection scope of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. An automatic segmentation method for PET tumors of nasopharyngeal carcinoma, characterized in that, It is obtained by the following steps: S1. Obtain the PET image of the object; S2. Reconstruct the PET image obtained in S1 to obtain a three-dimensional PET reconstructed image; S3. Convert the PET reconstructed image obtained in S2 into an SUV image; S4. Adjust the resolution of the SUV image obtained in S3 to obtain a PET-SUV image; S5. Input the PET-SUV image obtained in S4 into the integrated model, and then predict each pixel point in the PET-SUV image. Multiple first prediction probabilities and multiple second prediction probabilities are obtained corresponding to each pixel point, and the integrated model is composed of multiple trained V-NET models and multiple trained U-NET models; S6. Weight the multiple first prediction probabilities and multiple second prediction probabilities of each pixel point obtained in S5 to obtain the final prediction probability of each pixel point and the corresponding segmentation image; In the training process of the V-NET model and the U-NET model, the loss value is calculated through a combined loss function composed of a weighted cross-entropy loss function and a Dice loss function; S5 is to input the PET-SUV image obtained in S4 into each trained V-NET model to obtain the first prediction probability of each pixel point; input the PET-SUV image obtained in S4 into each trained U-NET model to obtain the second prediction probability of each pixel point; The number of the first prediction probabilities of each pixel point is the same as the number of trained V-NET models; The number of the second prediction probabilities of each pixel point is the same as the number of trained U-NET models; In S6, take the average of the multiple first prediction probabilities and multiple second prediction probabilities of each pixel point obtained in S5 to obtain the final prediction probability of each pixel point and the corresponding segmentation image; In S2, specifically, reconstruct the PET image obtained in S1 by the OSEM algorithm to obtain a three-dimensional PET reconstructed image in DICOM format; In S3, calculate the SUV value of each pixel point in the PET reconstructed image through formula (2), and the SUV values of all pixel points constitute the SUV image; ……Formula (2); where the unit of the SUV value is g / mL, tissue activity is the tracer activity of the diseased tissue per unit volume, the unit is Bq / mL, injected dose is the imaging agent injection dose and the unit is Bq, and body weight is the body weight and the unit is g; In S4, adjust the resolution of the SUV image to 512×512 to obtain a PET-SUV image in NiFTI format; Each of the V-NET models is trained by the following steps: A1. Randomly select a PET-SUV image from the database; the database is set with multiple pairs of image pairs, and each group of image pairs consists of a PET-SUV image and a labeled PET-SUV image corresponding to the PET-SUV image; A2. Input the PET-SUV image selected in A1 into the V-NET model to obtain the prediction probability of each pixel point; A3. Calculate the prediction probability of each pixel point obtained in A2 through the combined loss function and the annotation loss value of the annotated PET-SUV image; A4. Determine whether the number of training times has reached the set number of times. If not, go to A5; if so, go to A6, where the set number of times is greater than 200; A5. Update the parameters of the V-NET model and return to A1; A6. End the training, and use the V-NET model with the current parameters as the trained V-NET model; Each of the U-NET models is trained by the following steps: B1. Randomly select a PET-SUV image from the database; B2. Input the PET-SUV image selected in B1 into the U-NET model to obtain the prediction probability of each pixel point; B3. Calculate the prediction probability of each pixel point obtained in B2 through the combined loss function and the annotation loss value of the annotated PET-SUV image; B4. Determine whether the number of training times has reached the set number of times. If not, go to B5; if so, go to A6, where the set number of times is greater than 200; B5. Update the parameters of the U-NET model and return to B1; B6. End the training, and use the U-NET model with the current parameters as the trained U-NET model.

2. The automatic segmentation method of nasopharyngeal carcinoma PET tumor according to claim 1, characterized in that: The combined loss function is represented by Equation (1): ……Formula (1); Among them, is the weighted cross-entropy loss function term, is the Dice loss function term, L is the loss value, α is the constraint term, and 0 < α < 1, β is the weight of the foreground, and 0 < β < 1, N is the number of voxels, is a voxel, is the voxel is the predicted probability when it is the foreground, is the annotation of the voxel in the annotated PET-SUV image, and when is the foreground is 1, and when is the background is 0.

3. The automatic segmentation method of nasopharyngeal carcinoma PET tumor according to claim 2, wherein: The construction method of the database is as follows: C1. Obtain multiple PET images; C2. Reconstruct the PET images obtained in C1 to obtain three-dimensional PET reconstructed images; C3. Convert the PET reconstructed images obtained in C2 into SUV images; C4. Resolve the resolution of the SUV images obtained in C3 to obtain PET-SUV images; C5. Annotate each pixel point in the PET-SUV images obtained in C4 to obtain the annotated PET-SUV images.

4. The automatic segmentation method of nasopharyngeal carcinoma PET tumor according to claim 3, wherein: The activation function of the U-NET model is the ReLU function, ELU function, SELU function, or LeakyReLU function, and the learning rate lr ranges from le -4 to le -2 , the L2 regularization parameter λ ranges from le -25 to le -9 , the dropout probability d is 0 to 0.3, the number of convolutional layers is 1 to 4, the number of features is 16 to 40, the convolutional kernel size is 3×3×3 or 5×5×5, the initialization method is Xavier uniform, Xavier normal, Kaiming uniform, or Kaiming normal, the upsampling method is max-pooling, average pooling, or convolution, and the downsampling method is transpose convolution, nearest up-sampling, or trilinear up-sampling; The activation function of the V-NET model is the ReLU function, ELU function, SELU function, or LeakyReLU function, and the learning rate lr ranges from le -4 to le -2 , the L2 regularization parameter λ ranges from le -25 to le -9 , the dropout probability d is 0 to 0.3, the number of convolutional layers is 1 to 4, the number of features is 12 to 32, and the convolutional kernel size is 3×3×3 or 5×5×5.

5. An automatic segmentation device for nasopharyngeal carcinoma PET tumors, characterized in that: There is a memory and a processor. Computer-executable instructions are stored on the memory. When the processor runs the computer-executable instructions on the memory, the automatic segmentation method for nasopharyngeal carcinoma PET tumors described in any one of claims 1-4 is implemented.

6. A computer-readable storage medium, characterized in that: A computer program is stored. When the computer program is executed by the processor, the automatic segmentation method for nasopharyngeal carcinoma PET tumors described in any one of the above claims 1-4 is implemented.

Images