Automatic segmentation method, system and device for nasopharyngeal carcinoma PET tumor and storage medium

Abstract

The invention discloses a nasopharyngeal carcinoma PET tumor automatic segmentation method, system and device and a storage medium, and the nasopharyngeal carcinoma PET tumor automatic segmentation method obtains the final prediction probability of each pixel point and a corresponding segmentation image through six steps based on a PET image of an object. According to the nasopharynx cancer PET tumor automatic segmentation method, compared with a single model and a non-machine learning method, the provided integrated model shows better generalization ability. Compared with an original single 3D U-Net segmentation method, the precision is improved by 7.8%, 10.8% and 24.5% under the evaluation indexes of DSC, Jaccard and ASSD of the segmented image, and the segmentation effect is better.

Description

Technical Field

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

[0002] At present, NPC PET tumor delineation is mainly done 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 patient's tomographic images layer by layer and draw the outline in each layer of the image; 2. Manual delineation is highly dependent on the doctor's experience, and the delineation results of doctors with different experience vary greatly. Since PET is a functional imaging, it can depict the metabolic intensity of tissues, and the lesion area usually shows high brightness on the PET image. At present, lesions in NPC PET / CT images are difficult to define. PET images have low spatial resolution and are affected by artifacts generated by partial volume effects and scattered photons. Some non-adjacent tissues with high SUV may appear adjacent and difficult to distinguish. Limited resolution usually leads to blurred lesion boundaries, which makes it difficult to define lesion boundaries during segmentation. In addition, NPC lesions rarely have observable boundaries in CT (except for 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 lead to inaccurate tumor delineation. In recent years, more advanced methods have been proposed, such as the Markov random field method that uses the gradient and intensity of PET and CT simultaneously. Methods based on region growing, edge information, and texture analysis are relatively simple and suitable for outlining small lesions, but their ability to segment adjacent and overlapping areas is still insufficient. The active contour method achieves segmentation by minimizing the energy function to evolve the contour. Although these methods have improved in noise compatibility, they are still not robust enough to volume changes and uncertainty factors between different tumors.

[0003] The PET-CT joint segmentation method uses the grayscale distribution and gradient of pixels to build a model. Its basic assumption is that pixels belonging to the same tissue have consistent grayscale values ​​in PET and CT images. However, this type of method relies heavily on manually selected seed points. The difference between different models lies mainly in the different ways of defining penalty functions. 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 relies on prior information.

[0004] In recent years, the application of machine learning methods in the field 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 to 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 Fully Convolutional Network (FCN), SegNet and DenseNet, have been widely cited in medical images. Although deep learning methods have the problems of lack of interpretability and black box effect.

[0005] However, in the medical image segmentation of deep learning models, the model performance mainly depends on the selection of hyperparameters, and traditional hyperparameters include the depth of the convolution layer, the number of features, the size of the convolution 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 also increase, which can easily lead to overfitting.

[0006] Therefore, in view of the shortcomings of the existing technology, it is necessary to provide a method, system, device and storage medium for automatic segmentation of NPC PET tumor to solve the shortcomings of the existing technology. Summary of the invention

[0007] The purpose of the present invention is to avoid the shortcomings of the prior art and provide a method for automatic segmentation of nasopharyngeal carcinoma PET tumors. The method for automatic segmentation of nasopharyngeal carcinoma PET tumors has good generalization ability and higher segmentation accuracy.

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

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

[0010] S1. Acquire a PET image of the subject;

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

[0012] S3, converting the PET reconstructed image obtained in S2 into a SUV image;

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

[0014] S5, inputting the PET-SUV image obtained in S4 into an integrated model composed of a V-NET model and a U-NET model, and then predicting each pixel in the PET-SUV image, obtaining a plurality of first prediction probabilities and a plurality of second prediction probabilities corresponding to each pixel, and the integrated model is composed of a plurality of trained V-NET models and a plurality of trained U-NET models;

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

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

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

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

[0019]

[0020] in, 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 label of voxel i in the PET-SUV image, and when i is the foreground, t i is 1, when i is background, t i is 0.

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

[0022] A1. Randomly select a PET-SUV image in a database; the database is provided with a plurality of image pairs, and each image pair 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 predicted probability of each pixel;

[0024] A3, calculating 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, judging whether the number of training times reaches the set number, if not, entering A5, if yes, entering A6, wherein 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 described in the present invention is trained by the following steps:

[0029] B1, randomly selecting a PET-SUV image in the database;

[0030] B2, input the PET-SUV image selected in B1 into the U-NET model to obtain the predicted probability of each pixel;

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

[0032] B4, determine whether the number of training times reaches the set number, if not, go to B5, if yes, go to A6, where the set number 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 method for constructing the above database is as follows:

[0036] C1. Acquire multiple PET images;

[0037] C2, reconstructing the PET image obtained in C1 to obtain a three-dimensional PET reconstructed image;

[0038] C3, converting the PET reconstructed image obtained in C2 into a SUV image;

[0039] C4, obtaining a PET-SUV image by increasing the resolution of the SUV image obtained in C3;

[0040] C5. Label each pixel in the PET-SUV image obtained in C4 to obtain the labeled PET-SUV image.

[0041] In S6, the multiple first prediction probabilities and the multiple second prediction probabilities of each pixel point obtained in S5 are averaged to obtain a final prediction probability of each pixel point and a corresponding segmented image.

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

[0043] In the S3 or C3, the SUV value of each pixel in the PET reconstructed image is calculated by formula (2), and the SUV values ​​of all pixels constitute the SUV image;

[0044]

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

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

[0047] Preferably, the activation function of the U-NET model is a ReLU function, an ELU function, a SELU function or a 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 is 0 to 0.3, the number of convolutional layers is 1 to 4, the number of features is 16 to 40, the convolution 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.

[0048] Preferably, the activation function of the V-NET model is a ReLU function, an ELU function, a SELU function or a 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 is 0 to 0.3, the number of convolutional layers is 1 to 4, the number of features is 12 to 32, and the convolution kernel size is 3×3×3 or 5×5×5.

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

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

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

[0052] Provided is a nasopharyngeal carcinoma PET tumor automatic segmentation system, comprising:

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

[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 a SUV image, and finally adjust the resolution of the SUV image to obtain a PET-SUV image to be segmented;

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

[0057] The third object of the present invention is to avoid the shortcomings of the prior art and provide an automatic segmentation device for NPC PET tumors. The automatic segmentation device for NPC PET tumors has a higher segmentation accuracy for NPC PET tumors.

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

[0059] Provided is a device for automatic segmentation of nasopharyngeal carcinoma PET tumors, which is provided with a memory and a processor, wherein the memory stores computer executable instructions, and the processor implements the above-mentioned automatic segmentation method of nasopharyngeal carcinoma PET tumors when running the computer executable instructions on the memory.

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

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

[0062] A computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the above-mentioned method for automatic segmentation of nasopharyngeal carcinoma PET tumor is implemented.

[0063] The present invention discloses a method, system, device and storage medium for automatic segmentation of PET tumors of nasopharyngeal carcinoma, wherein the method for automatic segmentation of PET tumors of nasopharyngeal carcinoma is obtained by the following steps: S1, obtaining a PET image of an object; S2, reconstructing the PET image obtained by S1 to obtain a three-dimensional PET reconstructed image; S3, converting the PET reconstructed image obtained by S2 into an SUV image; S4, adjusting the resolution of the SUV image obtained by S3 to obtain a PET-SUV image; S5, inputting the PET-SUV image obtained by S4 into an integrated model composed of a V-NET model and a U-NET model, and then predicting each pixel point in the PET-SUV image, and each pixel point correspondingly obtains a plurality of first prediction probabilities and a plurality of second prediction probabilities, and the integrated model is composed of a plurality of trained V-NET models and a plurality of trained U-NET models; S6, weighting the plurality of first prediction probabilities and the plurality of second prediction probabilities of each pixel point obtained by S5 to obtain the final prediction probability of each pixel point, and the corresponding segmented 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. The integrated model proposed 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 segmentation accuracy of the present invention is improved by 7.8%, 10.8% and 24.5% under the DSC, Jaccard and ASSD evaluation indicators respectively, and the segmentation effect is better. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0066] Figure 2 The following is a graph showing the ASSD evaluation index obtained by the present invention and various models.

[0067] Figure 3 Segmented image segmented using a single U-Net and Dice loss function.

[0068] Figure 4 Segmented image segmented using a single V-Net and Dice loss function.

[0069] Figure 5 Segmented images using conventional PET-CT joint segmentation.

[0070] Figure 6 Segmented image segmented using a single U-Net and BCE loss function.

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

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

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

[0074] Example 1

[0075] A method for automatic segmentation of nasopharyngeal carcinoma PET tumors, such as Figure 1 As shown, it is obtained by the following steps:

[0076] S1. Acquire a PET image of the subject;

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

[0078] S3, converting the PET reconstructed image obtained in S2 into a SUV image;

[0079] S4, adjusting 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 in the PET-SUV image, and obtain multiple first prediction probabilities and multiple second prediction probabilities for each pixel, and the integrated model is composed of multiple trained V-NET models and multiple trained U-NET models; specifically, input the PET-SUV image obtained in S4 into a trained V-NET model each time to obtain the first prediction probability of each pixel; input the PET-SUV image obtained in S4 into a trained U-NET model each time to obtain the second prediction probability of each pixel, the number of first prediction probabilities for each pixel is the same as the number of trained V-NET models, and the number of second prediction probabilities for each pixel is the same as the number of trained U-NET models;

[0081] S6. Weight the multiple first prediction probabilities and the multiple second prediction probabilities of each pixel point obtained in S5 to obtain a final prediction probability of each pixel point and a 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 consisting of a weighted cross entropy loss function and a Dice loss function. The specific combined loss function is expressed by formula (1):

[0083]

[0084] in, 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 label of voxel i in the PET-SUV image, and when i is the foreground, t i is 1, when i is 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 into the Dice loss function term and the weighted cross entropy loss function term, so as to balance the weights between different tumor volumes and false positives and false negatives.

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

[0087] A1. Randomly select a PET-SUV image in the database; the database is provided with multiple pairs of images, and each pair of images consists of a PET-SUV image and a labeled 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;

[0089] 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;

[0090] A4, judging whether the number of training times reaches the set number of times, if not, entering A5, if yes, entering A6, wherein the set number of times is greater than 200, and the set number of times 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 in the database;

[0095] B2, input the PET-SUV image selected in B1 into the U-NET model to obtain the predicted probability of each pixel;

[0096] B3, calculate the prediction probability of each pixel 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 reaches the set number of times, if not, go to B5, if yes, go to A6, where the set number of times is greater than 200, and the set number of times is specifically 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 of A5 and B5 of the present invention can be adjusted by the tree structure Parzen estimation method (TPE), or by randomly updating the parameters of the V-NET model or U-NET model. And the tree structure 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 database construction method of the present invention is as follows:

[0104] C1. Acquire multiple PET images;

[0105] C2, reconstructing the PET image obtained in C1 to obtain a three-dimensional PET reconstructed image;

[0106] C3, converting the PET reconstructed image obtained in C2 into a SUV image;

[0107] C4, obtain a PET-SUV image by increasing the resolution of the SUV image obtained in C3;

[0108] C5. Label each pixel in the PET-SUV image obtained in C4 to obtain a labeled PET-SUV image.

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

[0110] In S6, the multiple first prediction probabilities and the multiple second prediction probabilities of each pixel point obtained in S5 are averaged to obtain a final prediction probability of each pixel point and a corresponding segmented image.

[0111] In S2 or C2, the PET image obtained in S1 is reconstructed by using the OSEM algorithm to obtain a three-dimensional PET reconstructed image in the 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 , with a resolution of 200×200. CT scanning (80mA, 120KVp) is used for attenuation correction. The CT voxel size is 0.98×0.98×3mm 3 , the resolution is 512×512, so the present invention adjusts the resolution of the PET reconstructed image to the same resolution as the CT image through S4 or C4.

[0113] In S3 or C3, the SUV value of each pixel in the PET reconstructed image is calculated by formula (2), and the SUV values ​​of all pixels are obtained to form a SUV image;

[0114]

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

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

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

[0118] The activation function of the U-NET model is ReLU function, ELU function, SELU function or LeakyReLU function, and the learning rate lr range is le -4 ~le -2 , the learning rate lr range is le -4 ~le -2 , the L2 regularization parameter λ ranges from le -25 ~le -9 , the discard 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 convolution 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 U-NET model

[0120]

[0121] The activation function of the V-NET model is ReLU function, ELU function, SELU function or LeakyReLU function, and the learning rate lr range is le -4 ~le -2 , the L2 regularization parameter λ ranges from le -25 ~le -9 , the discard probability d is 0-0.3, the number of convolutional layers is 1-4, the number of features is 12-32, and the convolution 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 V-NET model

[0123]

[0124] An automatic segmentation method for nasopharyngeal carcinoma PET tumors. The proposed ensemble model shows better generalization ability than single models and non-machine learning methods. The segmentation effect is better than 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, other features are the same as those in Example 1, and the integrated model of this embodiment consists of 6 trained V-NET models and 6 trained U-NET models. The database of this embodiment used 201 nasopharyngeal carcinoma subjects (average age: 47.3±11.8 years, range 11–78 years; 163 males and 38 females) for a retrospective study. All subjects were initially diagnosed with non-keratinizing undifferentiated nasopharyngeal carcinoma by histopathology. According to the AJCC standard (8th edition), the TNM staging system was used to divide the patients into 22 cases of stage I–II and 179 cases of stage III–IV, for a total of 201 cases, of which 160 cases were used as training sets for training models and 41 cases were used as test sets for testing. The subject data are detailed in Table 3.

[0127] Table 3. Subject information

[0128]

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

[0130] This embodiment is compared with the present invention by using a separate V-NET model, a U-NET model and a traditional PET-CT joint segmentation model. The separate V-NET model and the U-NET model are trained using separate loss functions, such as the Dice loss function and the BCE loss function, wherein the BCE loss function is a binary cross entropy loss, which is represented by the following formula:

[0131]

[0132] The Dice loss function is expressed as follows:

[0133]

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

[0135]

[0136] The present invention uses the manually delineated tumor area as the true value (gold standard) to evaluate the generalization performance of the model. The evaluation indicators can be divided into two categories: voxel-based indicators and surface-based indicators.

[0137] The voxel-based indicators Jaccard Coefficient (JC), DSC and ASSD, which measure the overlap between the segmented image and the true value, are calculated using the following formulas:

[0138]

[0139]

[0140]

[0141] Among them, |TP|, |FP| and |FN| are the number 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 true value. X and N Y They are the total number of pixels on the surface of the segmented image and the gold standard image, respectively. The larger the JC and DSC values ​​are, the more similar the two images are, while the closer ASSD is to 1, the more similar they are.

[0142] In this embodiment, the 6 trained V-NET models, the 6 trained U-NET model parameters, the parameters corresponding to the Dice loss function used by the trained U-NET model, the parameters corresponding to the BCE loss function used by the trained U-NET model, the parameters corresponding to the Dice loss function used by the trained V-NET model, and the parameters corresponding to the BCE loss function used by the trained V-NET model are as shown in Expression 3.

[0143]

[0144] The evaluation indicators of DSC, Jaccard and ASSD of the six trained V-NET models of the present invention, the six trained U-NET model parameters, 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 are shown in Table 4.

[0145] Table 4. Average DSC, Jaccard and ASSD evaluation indicators of 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 shown in 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 much better than that of U-Net, V-Net and the traditional joint segmentation method (p < 0.05). The DSC, Jaccard and ASSD evaluation indicators of the segmented image of the present invention improve the accuracy by 7.8%, 10.8% and 24.5% respectively.

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

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

[0150] Example 3

[0151] A nasopharyngeal carcinoma PET tumor automatic segmentation system is provided with:

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

[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 a SUV image, and finally adjust the resolution of the SUV image to obtain a PET-SUV image to be segmented;

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

[0156] The automatic segmentation system for PET tumor of NPC of the present embodiment can execute the automatic segmentation method for PET tumor of NPC provided by the embodiment 1 or 2 of the present invention, and can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0157] Example 4

[0158] A nasopharyngeal carcinoma PET tumor automatic segmentation device is provided with a memory and a processor, the memory stores computer executable instructions, and the processor implements the nasopharyngeal carcinoma PET tumor automatic segmentation method according to any one of claims 1 to 8 when running the computer executable instructions on the memory.

[0159] The automatic segmentation device for PET tumor of NPC in this embodiment can execute the automatic segmentation method for PET tumor of NPC provided in Embodiment 1 or 2 of the present invention, and can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0160] Example 5

[0161] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for automatic segmentation of nasopharyngeal carcinoma PET tumors of the above-mentioned embodiment 1 or 2.

[0162] A computer-readable storage medium of this embodiment can execute a method for automatic segmentation of NPC PET tumors provided by 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 solution of the present invention rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A method for automatic segmentation of nasopharyngeal carcinoma PET tumors, characterized in that: Obtained by the following steps: S1. Acquire a PET image of the subject; S2, reconstructing the PET image obtained in S1 to obtain a three-dimensional PET reconstructed image; S3, converting the PET reconstructed image obtained in S2 into a SUV image; S4, adjusting the resolution of the SUV image obtained in S3 to obtain a PET-SUV image; S5, inputting the PET-SUV image obtained in S4 into the integrated model, and then predicting each pixel in the PET-SUV image, obtaining a plurality of first prediction probabilities and a plurality of second prediction probabilities corresponding to each pixel, and the integrated model is composed of a plurality of trained V-NET models and a plurality of trained U-NET models; S6, weighting the multiple first prediction probabilities and the multiple second prediction probabilities of each pixel point obtained in S5 to obtain a final prediction probability of each pixel point and a corresponding segmented image; During the training process, the V-NET model and the U-NET model both calculate the loss value by a combined loss function consisting of a weighted cross entropy loss function and a Dice loss function.

2. The method for automatic segmentation of NPC PET tumor according to claim 1, characterized in that: S5 is to input the PET-SUV image obtained in S4 into a trained V-NET model to obtain a first prediction probability of each pixel; and input the PET-SUV image obtained in S4 into a trained U-NET model to obtain a second prediction probability of each pixel; The number of the first prediction probabilities for each pixel is the same as the number of the trained V-NET models; The number of the second prediction probabilities for each pixel is the same as the number of the trained U-NET models.

3. The method for automatic segmentation of NPC PET tumor according to claim 2, characterized in that: The combined loss function is expressed by formula (1): in, 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 label of voxel i in the PET-SUV image, and when i is the foreground, t i is 1, when i is background, t i is 0.

4. The method for automatic segmentation of NPC PET tumor according to claim 3, characterized in that: Each of the V-NET models is trained by the following steps: A1. Randomly select a PET-SUV image in a database; the database is provided with a plurality of image pairs, and each image pair 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 predicted probability of each pixel; A3, calculating 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; A4, judging whether the number of training times reaches the set number, if not, entering A5, if yes, entering A6, wherein 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 selecting a PET-SUV image in the database; B2, input the PET-SUV image selected in B1 into the U-NET model to obtain the predicted probability of each pixel; B3, calculating the prediction probability of each pixel point obtained in B2 and the annotation loss value of the annotated PET-SUV image by the combined loss function; B4, determine whether the number of training times reaches the set number, if not, go to B5, if yes, go to A6, where the set number 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.

5. The method for automatic segmentation of NPC PET tumor according to claim 4, characterized in that: The database is constructed as follows: C1. Acquire multiple PET images; C2, reconstructing the PET image obtained in C1 to obtain a three-dimensional PET reconstructed image; C3, converting the PET reconstructed image obtained in C2 into a SUV image; C4, obtaining a PET-SUV image by increasing the resolution of the SUV image obtained in C3; C5. Label each pixel in the PET-SUV image obtained in C4 to obtain the labeled PET-SUV image.

6. The method for automatic segmentation of NPC PET tumor according to claim 5, characterized in that: In S6, the multiple first prediction probabilities and the multiple second prediction probabilities of each pixel point obtained in S5 are averaged to obtain a final prediction probability of each pixel point and a corresponding segmented image; In the S2 or C2, the PET image obtained in the S1 is reconstructed by using an OSEM algorithm to obtain a three-dimensional PET reconstructed image in a DICOM format; In the S3 or C3, the SUV value of each pixel in the PET reconstructed image is calculated by formula (2), and the SUV values ​​of all pixels constitute the SUV image; The unit of SUV value is g / mL, tissue activity is the activity of tracer per unit volume of diseased tissue, the unit is Bq / mL, injected dose is the injected dose of imaging agent, the unit is Bq, and body weight is the body weight, the unit is g; In the S4 or the C4, the resolution of the SUV image is adjusted to 512×512 to obtain a PET-SUV image in a NiFTI format.

7. The method for automatic segmentation of NPC PET tumor according to claim 6, characterized in that: The activation function of the U-NET model is 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 drop 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 convolution 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 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 is 0 to 0.3, the number of convolutional layers is 1 to 4, the number of features is 12 to 32, and the convolution kernel size is 3×3×3 or 5×5×5.

8. A nasopharyngeal carcinoma PET tumor automatic segmentation system, characterized in that: The settings are: Dataset module - obtain PET-SUV images and corresponding annotated PET-SUV images to construct training sets; Model training module: train the V-NET model and U-NET model through the training set to obtain an integrated model; 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 a SUV image, and finally adjust the resolution of the SUV image to obtain a PET-SUV image to be segmented; Image segmentation module: inputs the PET-SUV image to be segmented into the integrated model and outputs the segmentation result.

9. An automatic segmentation device for nasopharyngeal carcinoma PET tumor, characterized by: A memory and a processor are provided, the memory stores computer executable instructions, and the processor implements the method for automatic segmentation of nasopharyngeal carcinoma PET tumors according to any one of claims 1 to 7 when running the computer executable instructions on the memory.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the method for automatic segmentation of nasopharyngeal carcinoma PET tumor according to any one of claims 1 to 7 is implemented.

Images