A system level doi calibration method for pet imaging

Abstract

The application discloses a system-level DOI calibration method for PET imaging, and steps of the method comprise the following steps: 1) selecting optional detectors in a PET imaging system as reference detectors, and the remaining detectors as to-be-calibrated detectors; collecting side incidence data of the reference detectors and recording corresponding DOI position values, and collecting top incidence data of each to-be-calibrated detector; 2) respectively performing data standardization on the collected data; 3) training a convolutional neural network model by using the standardized side incidence data, so as to obtain a pre-training model; 4) merging the standardized top incidence data of the i-th to-be-calibrated detector and the side incidence data, and training the pre-training model to obtain a DOI calibration model of the i-th to-be-calibrated detector; and 5) using each calibration model to predict the DOI position of the top incidence data of the corresponding to-be-calibrated detector in the PET imaging system. The application greatly saves the time cost of system-level DOI calibration.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of gamma ray imaging, and particularly relates to a system-level DOI calibration method for PET imaging. BACKGROUND

[0002] Positron emission tomography (PET) imaging measures the time, energy and position information of 511 keV gamma photons generated by annihilation of positrons emitted by a radionuclide through coincidence detection by a detector, and cooperates with an image reconstruction algorithm to quantitatively image the in vivo distribution of the radionuclide in a living body in a normal physiological state, in situ and non-invasively.

[0003] In PET imaging, a detector with depth of interaction (DOI) information can reduce parallax error and improve the uniformity of the spatial resolution of a PET system. In terms of DOI detector technology, the current brain PET system mainly adopts a stacked DOI detector scheme. For DOI schemes for different crystal types, the number of crystal layers is limited by the crystal type, and is mostly limited to double stacking, thereby limiting the DOI resolution. In addition, DOI schemes based on crystal layer misplacement or special design of a reflection layer require decoding of multiple crystal layers on the same decoding map, which limits the possibility of further reducing the size of the crystal, and thus results in relatively low spatial resolution. In order to overcome these limitations, further innovation and optimization of DOI detector technology are essential.

[0004] The use of machine learning methods to solve the DOI decoding problem in PET has attracted extensive research. In 2020, Andy LaBella et al. constructed a 1.4x1.4x20mm DOI detector with a DOI resolution of 2.5mm, which is a significant improvement over the DOI resolution of 4.5mm of the previous DOI detector. 3A crystal array detector with specially designed prismatoid light guides on top of the crystal array achieved DOI resolution of 1.84 mm using supervised machine learning algorithms (see A. LaBella, X. Cao, X. Zeng, W. Zhao, and A. H. Goldan, “Sub-2 mm depth of interaction localization in PET detectors with prismatoid light guide arrays and single-ended readout using convolutional neural networks,” Med. Phys., vol. 48, no. 3, pp. 1019-1025, Mar. 2021, doi: 10.1002 / mp.14654). Artem Zatcepin et al. used dense neural networks and convolutional neural networks, as well as a multiple linear regression-based estimation to estimate DOI information for a non-polished PET detector array with single-ended readout, further improving DOI resolution compared to traditional methods (see A. Zatcepin et al., “Improving depth-of-interaction resolution in pixellated PET detectors using neural networks,” Phys. Med. Biol., vol. 65, no. 17, p. 175017, Sep. 2020, doi: 10.1088 / 1361-6560 / ab9efc.). In 2023, Wen He et al. proposed a new four-layer DOI encoding stacked detector consisting of four alternating yttrium lutetium oxyorthosilicate (LYSO) and bismuth germanium oxide (BGO) scintillator arrays and successfully decoded the top and bottom LYSO events and the top and bottom BGO events using a convolutional neural network (W. He et al., “A CNN-based four-layer DOI encoding detector using LYSO and BGO scintillators for small animal PET imaging,” Phys. Med. Biol., vol. 68, no. 9, p. 095021, May 2023, doi: 10.1088 / 1361-6560 / accc07). It can be seen that machine learning-based methods provide an efficient and feasible solution to the DOI problem.

[0005] In 2023, Wen He et al. proposed a PET detector with high TOF, DOI, and spatial resolution (W. He et al., “A multi-resolution TOF-DOI detector for human brain dedicated PET scanner,” Phys. Med. Biol., vol. 69, no. 2, Jan. 2024, doi:10.1088 / 1361-6560 / ad1b6b), which is composed of two layers of yttrium lutetium silicate (LYSO) crystal arrays with different sizes. And by collecting light distribution data of different layer DOI positions through side incidence, a machine learning algorithm is trained to predict and classify single ray events into specific top or bottom layer positions, thereby realizing DOI calibration of the PET detector.

[0006] The above prior art can achieve good performance for PET detectors providing side incidence data by training a machine learning algorithm to predict DOI classification of ray events. For other detectors with the same size and same crystal layering, the performance of the model may be lacking due to slight differences in floodmap caused by temperature, voltage of photomultiplier devices, and coupling of different crystal layers.

[0007] However, for a PET imaging system with a large number of detectors (about 400 brain PET detectors, and thousands of clinical PET detectors), a machine learning algorithm needs to be trained for each detector, and side incidence data of each detector needs to be obtained (at least 600,000 event data are required for each training, and at least 80 minutes are required for data acquisition), and the experimental operation of collecting side incidence data is more complicated (manual updating of experimental devices is required for collecting data at each depth), which has the problems of complicated operation and high time cost. SUMMARY

[0008] To solve the problems of complicated operation and high time cost in DOI calibration of a PET imaging system, the present application provides a DOI calibration method for a PET imaging system, which only needs to be based on side incidence data of a small number of reference detectors to complete DOI calibration of top incidence data of a large number of same kind of to-be-calibrated detectors.

[0009] In the PET system, the detectors are installed on the system gantry (usually arranged in multiple detector rings). The detectors on the gantry are difficult to achieve side collimation incident acquisition data, so a large number of to-be-calibrated detectors need to be removed from the gantry to achieve side collimation incident, which brings great difficulty to the system level calibration, and even is unrealistic. In the use process of the PET system, in order to ensure its imaging performance, the accuracy of the system calibration needs to be checked regularly, and it may need to be calibrated multiple times. After the reference detector side incident data is obtained, the method does not need to collect a large amount of side incident data of the to-be-calibrated detector, only needs to collect a large amount of top incident data of the to-be-calibrated detector, and the radioactive source is placed in the detection field of view of the PET system, and the detector is still on the gantry to realize the collection. This greatly improves the possibility of the deep learning method in the practical application of the system level.

[0010] The application selects a CNN model with high precision trained by side incident data as a pre-training model, and then uses a semi-supervised learning method to input the collected side incident data of the reference detector, the top incident data of the to-be-calibrated detector and the enhanced top incident data of the to-be-calibrated detector into the pre-training model for training.

[0011] The application applies the semi-supervised learning method to the system level DOI calibration of the PET imaging, and only needs a small amount (1-5) of side incident data of the reference detector to complete the DOI calibration of a large amount (more than 400) of top incident data of the to-be-calibrated detector with the same structure, avoids tedious operation, and maintains a high calibration precision (more than 95%) of other detectors. In the case of using only one set of detector DOI calibration experimental platform, the time cost of the system level DOI calibration is greatly saved (so that the time of the system level DOI calibration is reduced from about 3 months to about 2 weeks).

[0012] The technical scheme adopted by the application to solve the technical problems is as follows:

[0013] A system level DOI calibration method for PET imaging, the steps of which comprise:

[0014] 1) For a PET imaging system, the PET imaging system comprises a plurality of detectors with the same structure; a plurality of detectors in the PET imaging system are selected as reference detectors, and the remaining detectors are to-be-calibrated detectors; side incident data of the reference detectors are collected, and corresponding DOI position values are recorded as labels, and top incident data of each to-be-calibrated detector in the PET imaging system are collected;

[0015] 2) The side incident data of the reference detectors and the top incident data of the to-be-calibrated detectors are respectively standardized;

[0016] 3) training a convolutional neural network model by using the side incident data after data standardization processing in a supervised learning manner; taking the trained convolutional neural network model as a pre-trained model;

[0017] 4) merging the top incident data of the i-th to-be-calibrated detector after standardization processing and the side incident data of any reference detector after standardization processing as a training sample of the i-th to-be-calibrated detector;

[0018] 5) training the pre-trained model by using the training sample of the i-th to-be-calibrated detector in a semi-supervised learning manner to obtain a DOI calibration model of the i-th to-be-calibrated detector;

[0019] 6) using the DOI calibration model of the i-th to-be-calibrated detector to predict the DOI position value of the top incident data of the i-th to-be-calibrated detector in the PET imaging system.

[0020] Further, in step 1), the top incident data of the to-be-calibrated detector is subjected to data enhancement to obtain enhanced top incident data; in step 2), the side incident data of the reference detector, the top incident data of the to-be-calibrated detector and the enhanced top incident data are subjected to data standardization respectively; in step 4), the top incident data of the i-th to-be-calibrated detector after standardization processing, the enhanced top incident data of the i-th to-be-calibrated detector after data standardization processing and the side incident data of any reference detector after standardization processing are merged as a training sample of the i-th to-be-calibrated detector.

[0021] Further, the method for training the pre-trained model by using the training sample in a semi-supervised learning manner is as follows: inputting the training sample into the pre-trained model to obtain an output Out1 corresponding to the side incident data, an output Out2 corresponding to the top incident data and an output Out3 corresponding to the enhanced top incident data; then optimizing the pre-trained model according to a loss function Loss = ω1 x Loss1 + ω2 x Loss2 + ω3 x Loss3; wherein the DOI position value corresponding to the side incident data is a label Label1, the output Out2 of the top incident data is subjected to Softmax to obtain a pseudo DOI position value Label2; Loss1 is the cross-entropy loss of Out1 and Label1, Loss2 is the cross-entropy loss of Out2 and Label2, Loss3 is the cross-entropy loss of Out3 and Label2, ω1 is the weight factor of Loss1, ω2 is the weight factor of Loss2 and ω3 is the weight factor of Loss3.

[0022] Further, the data enhancement manner is random rotation and flipping of the top incident data at any angle.

[0023] Further, the side incident data of the reference detector after data standardization processing is used to train the convolutional neural network model in multiple rounds in a supervised learning manner, and the performance of the convolutional neural network model after each round of training is tested, and the best performance convolutional neural network model is selected as the pre-training model.

[0024] Further, the method for standardizing the data is: first, the data is sequentially subjected to deviation standardization and standard deviation standardization to obtain standardized data; wherein the formula for deviation standardization is The standard deviation standardization formula is Where x is the original value of the event feature in the data, x max is the maximum value of all event features in the data, x min is the minimum value of all event features in the data, x * is the feature value after deviation standardization, which maps the feature value to [0, 1]; μ is the average value of all event features in the data after deviation standardization, σ is the standard deviation of all event features in the data after deviation standardization, and x' is the feature value after standard deviation standardization.

[0025] Further, the convolutional neural network model comprises a first convolutional layer, a second convolutional layer, a maximum pooling layer, a flattening layer, a fully connected layer and an output layer which sequentially process the input data; wherein a batch normalization unit and a rectified linear unit are sequentially provided before each convolutional layer.

[0026] Further, the detector is a LYSO crystal detector, a yttrium lutetium silicate crystal detector, a lutetium silicate crystal detector, a bismuth germanate crystal detector, a lutetium fine silicate detector, a garnet crystal detector, a lanthanum bromide crystal detector, a gadolinium silicate crystal detector, a gadolinium lutetium silicate crystal detector or a barium fluoride detector.

[0027] Further, the detector is a multi-layer crystal detector of the same or different size.

[0028] The present application selects a plurality of detectors from a PET imaging system as reference detectors; collects side incident data of the reference detectors and records the DOI position value as pre-training data, and collects top incident data of the to-be-calibrated detectors in the PET imaging system; the detectors in the PET imaging system are detectors of the same structure;

[0029] Further, in order to increase the generalization ability and robustness of the model, the top incident data of the to-be-calibrated detectors is subjected to data enhancement to obtain enhanced top incident data of the to-be-calibrated detectors;

[0030] Further, the collected side incident data of the reference detector, the top incident data of the to-be-calibrated detector and the enhanced top incident data are respectively subjected to data standardization, including deviation standardization and standard deviation (Z-Score) standardization; the data is first subjected to deviation standardization and then subjected to standard deviation standardization;

[0031] The formula of the deviation standardization is as follows:

[0032]

[0033] wherein x is the original value of the event feature in the data, x max is the maximum value of all event features in the data, x min is the minimum value of all event features, x * is the feature value after the deviation standardization processing, and the feature value is mapped to [0, 1]. The data is the side incident data, the top incident data or the enhanced top incident data.

[0034] The formula of the standard deviation standardization is as follows:

[0035]

[0036] wherein μ is the average value of all event features after the deviation standardization (i.e. the average value of each x * obtained above), σ is the standard deviation of all event features after the deviation standardization, and x' is the feature value after the standard deviation standardization processing, and the obtained feature value satisfies the characteristics of the average value being 0 and the standard deviation being 1.

[0037] Further, the side incident data of the reference detector after the data standardization processing is used to train a convolutional neural network (CNN) model in a supervised learning manner;

[0038] Further, the CNN network model is trained multiple rounds using the standardized side incident data, and multiple CNN network models with different parameters are generated. The performance of these trained models is tested, and a model with better performance is selected as a pre-trained model.

[0039] Further, the side incident data of the reference detector after the data processing, the top incident data of the to-be-calibrated detector to be calibrated and the enhanced top incident data are sequentially merged.

[0040] Further, the ordered combined data is input into the pre-trained model, and semi-supervised learning is used, specifically: the model has three outputs, which are the output Out1 of the side incident data passing through the model, the output Out2 of the top incident data, and the output Out3 of the enhanced top incident data, wherein the data formats of Out1, Out2 and Out3 are all two-dimensional arrays. The DOI position value in the side incident data is Label1, and the output Out2 of the top incident data is subjected to Softmax to obtain a pseudo DOI position value Label2.

[0041] The loss function of the semi-supervised learning model is composed of three parts, specifically:

[0042] Loss=ω1×Loss1+ω2×Loss2+ω3×Loss3

[0043] Wherein Loss1 is the cross-entropy loss calculated by Out1 and Label1, Loss2 is the cross-entropy loss calculated by Out2 and Label2, Loss2 is the cross-entropy loss calculated by Out3 and Label2, ω1, ω2 and ω3 are weight factors of Loss1, Loss2 and Loss3 respectively, and Loss is the final loss function.

[0044] Further, the parameters of the pre-trained model are updated through back propagation, and after iteration, the final network model is obtained as a DOI calibration model.

[0045] Further, the DOI calibration model is used to classify and predict the DOI position value of the top incident data of the to-be-calibrated detector.

[0046] The advantages of the present application are as follows:

[0047] In the conventional calibration method, for a large number of detectors of a PET system, side incident data of each detector needs to be collected, and a machine learning algorithm is trained for each detector, which has a high time cost. If only a few detectors are trained to predict other untrained detectors, the flood map of each layer of the detector will have misclassification events, and the classification accuracy is low, as shown in the accompanying Figure 4 .

[0048] The present application uses a semi-supervised machine learning method, which can complete the DOI calibration of a large number of (more than 400) to-be-calibrated detectors of the same kind using only a small amount (1-5) of side incident data of reference detectors, avoids tedious operations, and has high calibration accuracy, and each layer of the flood map has almost no misclassification events, as shown in the accompanying Figure 5 . The accuracy on the experimental data set can reach more than 95%.

[0049] Meanwhile, the model used by the present application has simple results, and a pre-trained model is used, further greatly reducing the training time, greatly saving the time cost of PET system level DOI calibration. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a method flowchart of the present application.

[0051] Figure 2 is a PET probe structure for collecting data.

[0052] Figure 3 is a CNN network model structure.

[0053] Figure 4 is a prediction result of the prior art.

[0054] Figure 5 is a prediction result of the present method.

[0055] Reference signs: 1-first crystal layer, 2-light guide layer, 3-second crystal layer, 4-photodetector layer. DETAILED DESCRIPTION

[0056] The present application will be further described in detail below in conjunction with the drawings, and the examples are only used to explain the present application, and are not used to limit the scope of the present application.

[0057] Referring to the flowchart of the DOI calibration method for the PET imaging system shown in Figure 1 , the flowchart includes the following steps:

[0058] Step S1: Obtain the PET probe structure for collecting data, as shown in Figure 2 , the specific composition is: the first crystal layer 1 is a 16x16 1.53x1.53x6mm 3 LYSO crystal array, the light guide layer 2 is a whole piece of glass with a thickness of 2mm, the second crystal layer 3 is an 8x8 3.0x3.0x15mm 3 LYSO crystal array, and the photodetector layer 4.

[0059] The method for acquiring data is as follows: first, collimate the rays into a collimated plane with a thickness of 1-2 mm along the z direction, and incident into different DOI positions of the detector #0 (reference detector) along the X or Y direction; for each incident DOI position, collect the light distribution data of the detector and record the DOI position value, and take the collected light distribution data and DOI position value as the side incident data and the corresponding label; the light distribution data incident into the detector #7 (to-be-calibrated detector) in the opposite direction of Z is the top incident data. The side incident data and the top incident data are from different PET detectors. Then, the collected top incident data is subjected to data enhancement to obtain enhanced top incident data, wherein the data enhancement is random rotation of the collected top incident data at any angle.

[0060] The data format of the three data sets is a 1x8x8 vector.

[0061] Step S2: normalizing the side incident data, top incident data and enhanced top incident data collected in step S1. Including deviation standardization and standard deviation (Z-Score) standardization.

[0062] First, the three data sets are subjected to deviation standardization, and the formula for deviation standardization is:

[0063]

[0064] where x is the original value of the event feature in the data, x max is the maximum value of all event features, x min is the minimum value of all event features, and x ω is the feature value after deviation standardization, which is mapped to [0, 1].

[0065] The data subjected to deviation standardization is then subjected to standard deviation standardization, and the formula for standard deviation standardization is:

[0066]

[0067] where μ is the average value of all event features after deviation standardization, σ is the standard deviation of all event features after deviation standardization, and x' is the feature value after standard deviation standardization, which has the characteristics of mean value 0 and standard deviation 1.

[0068] Step S3: dividing the side incident data obtained in step S2 into a training set and a test set, wherein the side incident training set accounts for 80% of the total data set, and the test set accounts for 20% of the total data set.

[0069] Step S4: input the event data of the side incident training set and the corresponding DOI position value obtained in step S3 into a CNN for supervised learning training, wherein the data format of the output result of each event data after passing through the convolutional neural network is a 2-dimensional array (L1, L2), and the cross-entropy loss of the DOI position value corresponding to each event is calculated, and the parameters of the model are updated by back propagation.

[0070] The event data of the side incident test set obtained in step S3 is input into the trained convolutional neural network model, the DOI position value predicted for each event data is obtained, the accuracy of the model is calculated by using the predicted DOI position value and the corresponding DOI position value in the test set, and finally a convolutional neural network model with better prediction performance for the test set is selected as a pre-trained model.

[0071] Step S5: sequentially merge the side incident data of detector #0 after normalization, the top incident data of detector #7 after normalization, and the enhanced top incident data of detector #7 after normalization in step S2.

[0072] Step S6: input the data set obtained in step S5 into the pre-trained model selected in step S4 for semi-supervised learning training. Specifically, the model will have three outputs, which are the output Out1 of the side incident data after passing through the model, the output Out2 of the top incident data, and the output Out3 of the enhanced top incident data, wherein the data format of Out1, Out2 and Out3 is a 2-dimensional array. The DOI position value recorded in the side incident data is Label1, and a pseudo DOI position value Label2 is obtained by Softmax on the output Out2 of the top incident data.

[0073] The loss function of the semi-supervised learning model consists of three parts, specifically:

[0074] Loss = ω1 x Loss1 + ω2 x Loss2 + ω3 x Loss3

[0075] Wherein Loss1 is the cross-entropy loss calculated by Out1 and Label1, Loss2 is the cross-entropy loss calculated by Out2 and Label2, Loss2 is the cross-entropy loss calculated by Out3 and Label2, ω1, ω2 and ω3 are weight factors of Loss1, Loss2 and Loss3 respectively, and Loss is the final loss function. The labels of the top incident data and the enhanced top incident data are the same, and training the CNN by using the top incident data before and after enhancement can improve the generalization ability.

[0076] The parameters of the model are updated by back propagation.

[0077] Step S7: using the trained convolutional neural network model in step S6 to make DOI classification prediction on the top incident data collected in step S1.

[0078] The CNN model structure used in the above steps is shown in the accompanying drawings Figure 3 As shown in the accompanying drawings, the input data of the CNN is a 1x8x8 array representing the detector response data of the gamma event, followed by two convolutional layers with a convolution kernel of 3x3, each of which is preceded by batch normalization (BatchNorm) and rectified linear unit (ReLU), and then sequentially followed by a max pooling layer, a flatten layer, a fully connected layer with a dropout of 60%, and an output layer.

[0079] The present application can only need to be based on a small amount (1-5) of side incident data of the detector to complete the DOI calibration of the top incident data of a large number of the same detectors. It avoids collecting the side incident data of each detector of the system through tedious operations, and greatly reduces the time cost (for 400 detectors of the brain PET detector, the time required to collect the side incident data of all detectors is about 32000 minutes, and if the system-level DOI calibration method is applied, it only needs about 80-400 minutes).

[0080] In the present application, a pre-trained model can not be used, and the side incident data of the reference detector, the top incident data of the to-be-calibrated detector, and the enhanced top incident data of the to-be-calibrated detector can be directly input into an untrained model for training, including but not limited to this.

[0081] The crystal detector in the PET imaging system to be calibrated is a double-layer LYSO crystal detector with different sizes, and the number of layers can be replaced, and crystal detectors with 2-4 layers can be used; the crystal material can be replaced, and detectors collecting data can be composed of yttrium lutetium silicate (LYSO), lutetium silicate (LSO), bismuth germanate (BGO), lutetium fine silicate (LFS), garnet crystal (GAGG), lanthanum bromide (LaBr3), gadolinium silicate (GSO), gadolinium lutetium silicate (LGSO), barium fluoride (BaF2), and other scintillation crystals, including but not limited to this.

[0082] Although specific embodiments of the present application are disclosed for illustrative purposes, the purpose is to help understand the content of the present application and to implement it, and those skilled in the art can understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present application and the appended claims. Therefore, the present application should not be limited to the disclosed content of the best embodiment, and the scope of protection claimed by the present application is defined by the scope of the claims.

Claims

1. A system-level DOI calibration method for PET imaging, comprising the steps of: 1) for a PET imaging system, the PET imaging system comprising a plurality of detectors of the same structure; selecting some detectors from the PET imaging system as reference detectors and the remaining detectors as to-be-calibrated detectors; collecting side incidence data of the reference detectors and recording the corresponding DOI position values as labels, and collecting top incidence data of each to-be-calibrated detector in the PET imaging system; 2) performing data standardization on the side incidence data of the reference detectors and the top incidence data of the to-be-calibrated detectors, respectively; 3) training a convolutional neural network model using the data-standardized side incidence data in a supervised learning manner; and using the trained convolutional neural network model as a pre-trained model; 4) for each to-be-calibrated detector, merging the standardized top incidence data of the to-be-calibrated detector and the standardized side incidence data of any reference detector to obtain a training sample of the to-be-calibrated detector; 5) for each to-be-calibrated detector, training the pre-trained model using the training sample of the to-be-calibrated detector in a semi-supervised learning manner to obtain a DOI calibration model of the to-be-calibrated detector; 6) for each to-be-calibrated detector, using the DOI calibration model of the to-be-calibrated detector to predict the DOI position value of the top incidence data of the to-be-calibrated detector in the PET imaging system.

2. The method of claim 1, wherein, In step 1), the top incidence data of the to-be-calibrated detector is data-augmented to obtain augmented top incidence data; in step 2), the side incidence data of the reference detectors, the top incidence data of the to-be-calibrated detectors, and the augmented top incidence data of the to-be-calibrated detectors are standardized, respectively; and in step 4), for each to-be-calibrated detector, the standardized top incidence data of the to-be-calibrated detector, the data-standardized augmented top incidence data of the to-be-calibrated detector, and the standardized side incidence data of any reference detector are merged to obtain a training sample of the to-be-calibrated detector.

3. The method of claim 2, wherein, The method for training the pre-trained model by semi-supervised learning using the training samples is: inputting the training samples into the pre-trained model to obtain an output Out1 corresponding to the side incident data, an output Out2 corresponding to the top incident data and an output Out3 corresponding to the enhanced top incident data; and then optimizing the pre-trained model according to a loss function Loss = ω1 x Loss1 + ω2 x Loss2 + ω3 x Loss3; wherein a DOI position value corresponding to the side incident data is a label Label1, a pseudo DOI position value Label2 is obtained by performing Softmax on the output Out2 of the top incident data; Loss1 is a cross-entropy loss of Out1 and Label1, Loss2 is a cross-entropy loss of Out2 and Label2, Loss3 is a cross-entropy loss of Out3 and Label2, ω1 is a weight factor of Loss1, ω2 is a weight factor of Loss2, and ω3 is a weight factor of Loss3.

4. The method of claim 2, wherein, The data enhancement manner is: randomly rotating and flipping the top incident data at any angle.

5. The method of claim 1, wherein, The side incident data of the reference detector after data standardization processing is used to perform multi-round training on the convolutional neural network model in a supervised learning manner, and the convolutional neural network model after each round of training is tested for performance, and the convolutional neural network model with the best performance is selected as the pre-trained model.

6. The method of claim 5, wherein, The method for normalizing the data is: firstly, sequentially performing dispersion standardization and standard deviation standardization on the data to obtain normalized data; wherein the formula of the dispersion standardization is the formula of the standard deviation standardization is wherein x is an original value of an event feature in the data, x max is a maximum value of all event features in the data, x min is a minimum value of all event features in the data, x * is a feature value after the dispersion standardization processing, and the feature value is mapped to [0, 1]; μ is an average value of all event features in the data after the dispersion standardization, σ is a standard deviation of all event features in the data after the dispersion standardization, and x ′ is a feature value after the standard deviation standardization processing.

7. The method of claim 1, wherein, The convolutional neural network model comprises a first convolutional layer, a second convolutional layer, a maximum pooling layer, a flattening layer, a full connection layer and an output layer which sequentially process the input data; wherein a batch normalization unit and a rectified linear unit are sequentially arranged before each convolutional layer.

8. The method of claim 1, wherein, The detector is a yttrium lutetium silicate crystal detector, a lutetium silicate crystal detector, a bismuth germanate crystal detector, a lanthanum bromide crystal detector, a gadolinium silicate crystal detector, a gadolinium lutetium silicate crystal detector or a barium fluoride detector.

9. The method according to claim 1 or 8, characterized in that, The detector is a multi-layer crystal detector with the same or different sizes. The detector is a multi-layer crystal detector with the same or different sizes.

Images