PET probe tag data mapping method, PET probe depth prediction model training method, and PET probe depth prediction method

Abstract

The application provides a PET detector label data mapping method, a PET detection depth prediction model training method and a PET detection depth prediction method. The method comprises the following steps: calculating DOI distributions of a plurality of depth positions according to an acquired coincidence event data set, wherein the coincidence event data set comprises coincidence event data of a test crystal detector at a plurality of different depth positions, and the DOI distribution comprises depth values and count values of a plurality of event points; for each depth position, calculating a depth distribution mean value of the DOI distribution and a deviation value of the depth value of each event point from the depth distribution mean value; and calculating a depth mapping value of each event point according to each deviation value, a detection width of the test crystal detector and a label value of the current depth position, taking the depth mapping value as a label value of each event point, and obtaining a continuous, accurate and more abundant training data with the label value, so as to improve the prediction ability of the prediction model.

Description

Technical Field

[0001] This application belongs to the field of PET detector technology, specifically, it relates to a PET detector tag data mapping method and a PET detection depth prediction method. Background Technology

[0002] Positron emission tomography (PET) is a molecular imaging technique based on radioactive tracers, used to measure the metabolic distribution of radiolabeled tracers within the body. As a clinical diagnostic tool, PET is widely used in early cancer diagnosis, treatment prognosis assessment, and neuroscience. PET imaging reconstructs a three-dimensional image of the distribution of tissues within the body by detecting paired gamma photons generated after positron-electron annihilation, providing detailed information about the patient's internal structures. In a PET system, a scintillator-based detector is one of the core components, responsible for converting gamma rays into detectable signals.

[0003] With increasing resolution requirements, especially in brain imaging, the resolution of the depth of interaction (DOI) in PET technology is crucial for maintaining high spatial 3D imaging accuracy across the entire field of view. Depth resolution refers to the ability of a PET detector to accurately measure the depth of interaction of gamma photons within the detector. In small animal and brain imaging, the resolution of traditional PET detectors is insufficient to meet millimeter-level imaging requirements. To improve the depth resolution of PET, researchers have proposed various methods, including techniques based on dual-end readout crystal arrays, continuous crystals, and prismatic crystal readout, aiming to improve the accuracy of depth measurement and reduce imaging errors caused by depth uncertainty.

[0004] Traditional PET detectors typically employ depth measurement methods based on energy ratios. In dual-readout detectors, the interaction depth of gamma photons within the detector affects the photon energy reaching the photodetectors at both ends. Different interaction depths result in different gamma photon propagation paths, leading to different total energies reaching the two photodetectors. By calculating the energy ratio received by the two photodetectors, the depth of gamma photon interaction (DOI) can be estimated. The core principle of this method is to infer the interaction location based on changes in energy ratio. Taking a SiPM array as an example, the detector array typically consists of multiple units arranged in an array structure, simultaneously collecting photon information at different depth positions. By analyzing the detector signals at different depth positions and combining them with the energy ratio, the depth information is calculated. However, traditional depth measurement methods rely on the energy difference of scintillation photons reaching the two detectors, and are limited by circuit noise and detector crystal materials. For applications requiring higher depth resolution, high-precision information processing and counting are needed to obtain a more accurate DOI.

[0005] Deep learning has significant advantages, especially in signal analysis and feature extraction. By deeply mining various signal features (such as energy signals, time signals, and waveform signals) acquired by PET detectors, deep learning models can learn the correlation between the acquired signal data and depth information, thereby improving the accuracy of depth measurement.

[0006] Existing technologies primarily employ supervised learning methods such as convolutional neural networks (CNNs) or decision tree models to learn depth information from PET detectors. These methods predict the depth location of each acquisition event by analyzing raw signals collected at different depth positions. In this process, depth information acquisition typically relies on a mobile platform to acquire signal data from the detector at different locations. The acquired signals usually include various features such as gamma photon energy, photon interaction time, and signal waveform amplitude. Multi-feature depth energy distribution information is extracted by processing the raw signals from the detector array. Based on these acquired signals, the model can learn the patterns of depth location, thereby enabling depth prediction for new data.

[0007] In terms of data acquisition, traditional supervised learning and semi-supervised models typically use discrete labels at different depth locations. For example, the detector's depth ranges from 1.5 mm to 19.5 mm, with a 1 mm interval between each depth location. The data acquisition system directly obtains the corresponding depth labels based on the detector's current position and uses these labels along with the original signal data to construct a training set. The signal data is used as feature input, and the depth labels are used as the target output to train the depth prediction model.

[0008] In supervised learning, depth labels serve as the foundation of training data, guiding the training of neural network models through discrete labeled data. Traditional methods typically obtain fixed depth labels based on the detector's current position and then discretize them, resulting in labels with significant intervals. For example, data labels might be a fixed series of values ​​such as 1.5 mm, 2.5 mm, ..., 19.5 mm. While this discrete label approach is simple and direct, it leads to insufficient continuity between depth labels, thus affecting the model's learning performance and prediction accuracy. Although this depth prediction method based on discrete labels has some applications, it also has significant limitations. Due to the discretization of depth labels, the model struggles to fully capture the continuity and subtle differences between depths during training, resulting in larger errors in predictions when dealing with minute depth variations.

[0009] Current techniques only use fixed label positions such as 1.5, 2.5, and 3.5 for training, failing to cover data outside these labels, resulting in poor depth prediction accuracy at these positions. Even when trained on regression tasks, discrete label methods still cannot accurately handle continuous differences between labels, especially at non-label positions, where the model struggles to accurately predict depth information, leading to decreased accuracy. Furthermore, this discrete label training method fails to adequately consider the detector's measurement accuracy and the actual distribution of depth information, limiting the model's ability to learn from depth variations. Therefore, improving label mapping methods to provide more accurate training labels is crucial for improving the accuracy of depth prediction models. Summary of the Invention

[0010] The technical problem addressed in this application is: how to achieve continuity and accuracy of label data in order to improve the prediction accuracy of the model.

[0011] This application provides a PET detector tag data mapping method, the PET detector tag data mapping method comprising:

[0012] The DOI distributions at several depth locations are calculated based on the collected coincidence event dataset. The coincidence event dataset includes coincidence event data of the tested crystal detector at several different depth locations, and the DOI distribution includes depth values ​​and count values ​​of multiple event points.

[0013] For each depth location, calculate the mean depth distribution of the DOI distribution and the deviation of the depth value of each event point from the mean depth distribution;

[0014] For each depth location, a depth mapping value for each event point is calculated based on the various deviation values, the detection width of the test crystal detector, and the label value of the current depth location. This depth mapping value serves as the label value for each event point.

[0015] Optionally, the plurality of different depth positions are a plurality of depth positions distributed at equal intervals, and the distance between two adjacent depth positions is the detection width of the test crystal detector.

[0016] Optionally, methods for calculating the DOI distribution at several depth locations based on the collected coincident event dataset include:

[0017] For each depth location, Gaussian fitting is performed based on the coincidence event data to obtain the original DOI distribution;

[0018] The error data of the original DOI distribution are corrected based on the mean and variance obtained after Gaussian fitting to obtain the corrected DOI distribution.

[0019] Optionally, the mean of the depth distribution of the DOI distribution is the mean obtained after Gaussian fitting.

[0020] Optionally, for each depth location, a depth mapping value for each event point is calculated based on the various deviation values, the detection width of the test crystal detector, and the tag center value at the current depth location, including:

[0021] The maximum absolute value of each deviation value is taken as the maximum deviation value;

[0022] The tag mapping range is determined based on the detection width of the test crystal detector;

[0023] The normalized depth mapping value of each event point within the label mapping range is determined based on the deviation value and the maximum deviation value of each event point;

[0024] The depth mapping value of each event point is calculated based on the normalized depth mapping value of each event point and the label center value of the current depth position.

[0025] Optionally, the label center value at the depth location is the actual detection center position value at the current depth location.

[0026] Optionally, the label mapping range is to w represents the detection width.

[0027] This application also discloses a training method for a PET detection depth prediction model, the training method comprising:

[0028] Acquire coincidence event data at several different depth positions of the test crystal detector, wherein the coincidence event data at each depth position includes data from multiple event points;

[0029] The tag values ​​for each event point are calculated based on the PET detector tag data mapping method described above.

[0030] The data and label values ​​of all event points were used as training samples to train the PET detection depth prediction model.

[0031] Optionally, the PET detection depth prediction model is a prediction model built based on the extreme gradient boosting algorithm, and the data for each event point includes energy, time, and waveform.

[0032] This application also discloses a method for predicting PET detection depth, the method comprising:

[0033] Acquire coincidence event data collected by the test crystal detector;

[0034] The coincident event data is input into the PET detection depth prediction model to obtain the predicted depth. The PET detection depth prediction model is trained according to the above method.

[0035] This application provides a PET detector tag data mapping method, a PET detection depth prediction model training method, and a PET detection depth prediction method, which have the following technical effects:

[0036] The label values ​​of each event point in the event data at each depth location are calculated, and the distribution of each label value conforms to the actual physical detection distribution. This allows us to obtain continuous, accurate, and richer training data with label values, thereby improving the predictive ability of the prediction model. Attached Figure Description

[0037] Figure 1 This is a flowchart of a PET detector tag data mapping method according to one or more embodiments;

[0038] Figure 2 A diagram of a test apparatus for conforming event data according to one or more embodiments;

[0039] Figure 3 The DOI distribution for multiple depth locations according to one or more embodiments;

[0040] Figure 4 A DOI distribution and tag distribution map for one of the depth locations according to one or more embodiments;

[0041] Figure 5 This is a label distribution map at multiple depth locations according to one or more embodiments. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] Before describing the various embodiments of this application in detail, the technical concept of this application is first briefly described: Currently, the training data used in deep learning-based PET detector depth prediction methods is discrete depth label data, which has large gaps and is discontinuous, leading to a decrease in the prediction accuracy of the trained prediction model. Therefore, this application provides a PET detector label data mapping method that calculates the label values ​​of each event point in the coincidence event data at each depth location, and the distribution of each label value conforms to the actual physical detection distribution. This allows for the acquisition of continuous, accurate, and richer training data with label values, thereby improving the prediction capability of the prediction model. The specific principles of the PET detector label data mapping method and its application in this application are described below with reference to more embodiments.

[0044] Specifically, such as Figure 1 As shown, the PET detector tag data mapping method in this embodiment includes the following steps:

[0045] Step S10: Calculate the DOI distribution of several depth locations based on the collected coincidence event dataset. The coincidence event dataset includes coincidence event data of the test crystal detector at several different depth locations. The DOI distribution includes the depth value and count value of all event points at the current depth location.

[0046] Step S20: For each depth location, calculate the mean depth distribution of the DOI distribution and the deviation between the depth value of each event point and the mean depth distribution.

[0047] Step S30: For each depth position, calculate the depth mapping value of each event point based on the deviation values, the detection width of the test crystal detector, and the label value of the current depth position. The depth mapping value is used as the label value of each event point.

[0048] For example, several different depth positions are several depth positions that are equally spaced, and the distance between two adjacent depth positions is the detection width of the test crystal detector.

[0049] The following section describes the process of collecting data from event datasets. For example... Figure 2As shown, the experimental setup includes a reference detector, a radiation source, and a test crystal detector. The reference detector, used to monitor and verify the precise location of the radiation source, is typically placed symmetrically to the test crystal detector. It provides a baseline for subsequent data analysis. The radiation source is usually a radioactive isotope or a gamma-ray source, used to generate and emit gamma photons, which interact with the crystal to produce coincidence events. The detection range is the single-layer DOI detection range of the test crystal detector, which can be obtained based on the reference detector and geometric setup. In this example, the detection range is 0.99 mm. Figure 2 This embodiment uses 22 Taking a Na radioactive source as an example, a test crystal detector is used to detect photons generated after gamma rays interact with the crystal and record the detected signals. The test crystal uses a high-efficiency scintillator material; in this embodiment, LYSO is used as an example, and it is equipped with a photodetector (such as a photodiode, SiPM, etc.) for signal acquisition. This embodiment uses a dual-end readout detector, that is, there are two SiPM array detectors, SiPM1 and SiPM2, which read out the signals at both ends of the crystal array, respectively.

[0050] During data acquisition, the 20mm crystal detector was first calibrated to 18 positions ranging from 1.5mm, 2.5mm, 3.5mm to 18.5mm according to preset locations. Each position corresponds to different depth information. A precision moving platform was used to move the detector to each depth position (e.g., 1.5mm, 2.5mm, 3.5mm, etc.) to ensure the detector was at an accurate depth during data acquisition. The reference point for this depth position is the connection point between the crystal array and the SiPM1.

[0051] At each depth location, the event information of gamma photons emitted by the radiation source and interacting with the test crystal detector is recorded. If the interaction between the reference detector and the test crystal detector is simultaneously detected by the electronic system within the effective time window, then this event is recorded as a valid event. The double-ended SiPM energy signal is recorded for this valid event. The SiPM1 has a total of A detection channels, where channel a has an energy of E. a The total number of detection channels for SiPM2 is B, of which channel b has an energy of E. b .

[0052]

[0053] Finally, for each depth location, the energy outputs E1 and E2 of each channel in SiPM arrays 1 and 2 are calculated to determine the energy ratio measured by the two photodetectors as the DOI value. The integral value of the energy waveform curve, the slope of the waveform's rising and falling edges, and the signal acquisition time waveform are then calculated as features. These features are used to infer the depth location of gamma photon interactions and for tasks such as machine learning or deep learning prediction.

[0054] In one or more embodiments, a method for plotting DOI distributions at several depth locations based on a collected coincidence event dataset includes: for each depth location, performing Gaussian fitting on the coincidence event data to obtain an original DOI distribution; and correcting the error data of the original DOI distribution based on the mean and variance obtained after Gaussian fitting to obtain a corrected DOI distribution. Specifically, data deviating from the main peak due to photon scattering, light transmission deviation, photoelectric conversion error, and electronic device noise are removed to obtain a filtered DOI distribution.

[0055] For example, taking a 4×4 crystal array, each crystal can obtain coincidence events and their corresponding information at depths of 1.5, 2.5, 3.5 to 18.5 mm through acquisition experiments. DOIs can be calculated mathematically, and the DOI distribution at each depth location can be plotted, such as... Figure 3 As shown, the horizontal axis represents the interaction depth (DOI, hereinafter referred to as depth value), and the vertical axis represents the count value. Figure 3 'a' represents the original DOI distribution of a single crystal. Figure 3 'b' represents the DOI distribution of a single crystal after noise reduction. After obtaining the DOI data for each location, Gaussian fitting is performed on each set of data, and a reasonable range of DOI data x is selected. In this example, the 3σ range of the main peak is used for fitting, as shown in the following formula. The data within this range represents the main detection characteristics at the depth of each depth distribution layer and conforms to the distribution trend at that location. Figure 3 b.

[0056] μ-3σ≤x≤μ+3σ

[0057] For each depth location, a Gaussian fit is performed. After the fit is complete, the mean μ and variance σ of the Gaussian fit are calculated. In this example, there are 18 depth locations (e.g., ...). Figure 3 (a) requires fitting 18 normal distributions, and for each normal distribution, a 3σ range is taken with each mean as the center. The purpose of this 3σ range is that it contains 99.73% of the original normal distribution data, thus removing tail noise before label mapping, preserving the most important features to the greatest extent possible, and eliminating the impact of data deviating from the main peak due to photon scattering, light transmission bias, photoelectric conversion errors, and electronic device noise on the learning of normal DOI data.

[0058] In other implementations, if the data distribution characteristics change under different experimental conditions or different detector configurations, other statistical distribution fitting methods (such as Log-Normal distribution, Poisson distribution, etc.) can be used to replace the Gaussian fitting method, or machine learning methods (such as clustering algorithms) can be used to automatically determine the distribution of labels to meet the needs under different detection conditions.

[0059] In traditional methods, the 18 sets of matching event data are each assigned a separate label, and each set uses the same label, assuming they represent data from the same depth location. However, in reality, each set of matching event data contains data within a certain depth range. For example... Figure 3 As shown, the coincidence range is 1mm. For example, coincidence event data obtained at a depth of 1.5mm actually corresponds to a range of 1mm to 2mm. However, traditional methods consider all coincidence event data to correspond to a depth of 1.5mm and use the same label. This reduces the precision of the data samples, leading to label discontinuities. During training, the model struggles to learn more accurate and richer data features, thus reducing the model's prediction accuracy. Therefore, the following method uses label mapping to expand coincidence event data from single-label to label-range coincidence event data; that is, coincidence event data at different depth locations are represented by different labels.

[0060] In one or more embodiments, for each depth location, a depth mapping value for each event point is calculated based on various deviation values, the detection width of the test crystal detector, and the tag center value at the current depth location, including:

[0061] First, the maximum absolute value of each deviation value is taken as the maximum deviation value. Each DOI distribution includes depth and count values ​​for multiple event points, with each event point corresponding to a different depth location. The mean of the depth distribution of the DOI distribution is the mean μ obtained after Gaussian fitting, and the deviation of the depth value of each event point from the mean of the depth distribution is:

[0062] δ i =x i -μ

[0063] In the formula, x i δ represents the depth value of the i-th event point. i This represents the deviation value of the i-th event point. Since the depth values ​​of some event points are less than the mean, the deviation values ​​of these event points are negative. The maximum deviation value can be expressed as:

[0064] δ max d =max(|δ all d |)

[0065] In the formula, δ max d δ represents the maximum deviation value of the depth position d. all d |δ represents the deviation value of each event point at depth d. all d | represents the absolute value of the deviation of each event point at depth position d.

[0066] Next, the tag mapping range is determined based on the detection width of the test crystal detector. Assuming the detection width is denoted as w, the tag mapping range is... to The label mapping range here is a normalized range. For example, w is 1 mm.

[0067] Then, based on the deviation value and the maximum deviation value of each event point, the normalized depth mapping value of each event point within the label mapping range is determined. Here, the normalized depth mapping value of the i-th event point at the current depth position d is δ. norm di :

[0068]

[0069] Finally, the depth mapping value for each event point is calculated based on the normalized depth mapping value and the label center value at the current depth position. For the current depth position d, the depth mapping value V of the i-th event point is... map di for:

[0070] V map di =δ norm di +lable d

[0071] In the formula, label d This represents the center value of the label at the current depth position d. For example, the center value of the label at a depth position is the depth value of that depth position; for instance, if the current depth position is 1.5mm, then the center value of the label is 1.5mm. After the above mapping, the distribution range of each depth mapping value is consistent with the actual physical detection distribution range. For example, for a depth position of 1.5mm, the distribution range of the depth mapping value and the actual physical detection distribution range are both 1mm to 2mm. Figure 4 As shown, Figure 4 In the diagram, 'a' represents the DOI distribution at a depth of 1.5 mm. Figure 4 The 'b' in the diagram represents the label distribution at a depth of 1.5 mm.

[0072] For example, the same label mapping operation was performed on the coincident event data of the above 18 depth locations, and the distribution map of each DOI distribution after label mapping is as follows. Figure 5 As shown, by mapping the label center value at each depth location to the corresponding depth value, and by setting the label distribution range to match the detector's detection accuracy range, the label data not only conforms to the accuracy measurement and actual distribution of depth information, but also retains the continuity and regularity of the DOI distribution. This makes the collected data more refined, avoids the problem of discontinuous label intervals that may occur in traditional label design, and effectively improves the stability and accuracy of model training.

[0073] The above embodiments use LYSO scintillation crystal arrays and SiPM arrays as detector configurations, but this method is also applicable to other scintillation materials (such as BGO, NaI, etc.) or other photodetectors (such as PMT, etc.) to adapt to different experimental needs or performance requirements. Furthermore, this mapping method can also be used as a training and learning method for semi-continuous crystal detectors, continuous crystal detectors, and semiconductor detectors. In semi-continuous and continuous crystal detectors, depth location (DOI) prediction is performed using the geometric characteristics and energy distribution of the detector array through a mapping tag method, ensuring the continuity and uniform distribution of tags, thereby improving depth resolution. Similarly, in semiconductor detectors, by training on energy signals and carrier drift characteristics, this method can also be effectively applied to the training of depth prediction models, meeting the depth localization requirements of different types of detectors.

[0074] In one or more embodiments, a method for training a PET detection depth prediction model includes the following steps:

[0075] Acquire coincidence event data at several different depth positions of the test crystal detector. The coincidence event data at each depth position includes data from multiple event points.

[0076] The tag values ​​for each event point are calculated based on the PET detector tag data mapping method described above.

[0077] The data and label values ​​of all event points were used as training samples to train the PET detection depth prediction model.

[0078] For example, the PET detection depth prediction model is a prediction model built on the Extreme Gradient Boosting (XGBoost) algorithm, and the data for each event point includes energy, time and waveform.

[0079] In other embodiments, other supervised learning models, semi-supervised learning models, or ensemble learning algorithms (such as LightGBM, CatBoost, etc.) can be considered as alternatives to XGBoost. Alternatively, deep learning algorithms, such as models based on convolutional neural networks (CNNs) or recurrent neural networks (RNNs), can be employed, especially when processing time-series data such as time signals, where deep learning algorithms like LSTM are particularly useful.

[0080] XGBoost is an efficient implementation of the Gradient Boosting Trees (GBT) algorithm, widely used in various supervised learning tasks. It improves the predictive ability of a model by constructing a series of decision trees, performing particularly well when dealing with complex data. XGBoost combines multiple simple decision tree models, resulting in a final model that outperforms a single decision tree. Each tree continuously reduces the gap between the predicted value and the label by optimizing the loss function and considering the model's residuals (errors). By training multiple trees, XGBoost can fully learn the features in the data and make more accurate predictions when solving regression and classification problems.

[0081] In supervised learning, the primary purpose of the objective function is to measure the difference between the model's predicted values ​​and the true label values, guiding the model to continuously optimize during training. The objective function plays a guiding role in training, helping the model adjust its parameters to minimize the loss function. Specifically, the objective function typically consists of two parts: a prediction error term, which measures the difference between the model's output and the actual label; and a regularization term, which prevents overfitting. By optimizing the objective function, the model can gradually approach the true label and improve prediction accuracy.

[0082] In XGBoost, each tree is built to reduce the residual between the predicted and labeled values. A new tree is constructed based on the residual of the current model and corrects the prediction error of the previous tree by minimizing the loss function. Each added tree attempts to improve the shortcomings of the previous model, gradually approaching the true label; this is called "gradient boosting." During each tree split, the newly formed leaves reduce prediction error by finding the split point that minimizes the loss function. This process is achieved using the gradient descent algorithm; that is, each tree split is performed "along the direction of the gradient," aiming to minimize the loss function and make the predicted value as close to the true label as possible, avoiding overfitting. Therefore, each new decision tree in the XGBoost algorithm improves model performance by reducing the gradient (residual) based on the previous tree.

[0083] The algorithm flow is as follows:

[0084] Input: Training dataset X i For each event i ∈ N, there is a corresponding label value y. i , tag y i It is calculated using the aforementioned label mapping method. X i Training features include, but are not limited to, information such as energy, time, and waveform.

[0085] XGBoost objective function:

[0086]

[0087] y i Calculate the label value for the mapping of event i. Here, γ represents the model's predicted value, γ is the regularization parameter controlling the weights of the leaf nodes, and k is the number of nodes in the T-tree. For each tree, the more nodes, the higher the model's complexity. K represents the total number of trees, and k is the current tree number. For L2 regularization terms, where It is the weight of the j-th leaf node in tree k.

[0088] Model loss function:

[0089]

[0090] The algorithm calculates the model prediction in round t-1:

[0091]

[0092] The objective function is expanded using Taylor and then transformed into gradient calculation:

[0093]

[0094] Formula for calculating the t-th gradient boosting tree:

[0095] In the formula for calculating the t-th gradient boosting tree, g i In the objective function f t (x i According to the first derivative of the Taylor expansion, h i For the second derivative, w t (x i )=f t (x i The method of summing the values ​​of each sub-leaf first, and then summing all samples of the current sub-leaf, is primarily used during the training of the tree model. Each tree divides the data into multiple sub-leafs, and each node only focuses on the data of its corresponding sub-leaf. By summing the sample values ​​within each sub-leaf first, and then summing the results of the entire tree, XGB effectively reduces unnecessary computation, thereby accelerating the model training process and helping XGB to more efficiently calculate the gradient and loss of each sub-leaf during gradient optimization. Therefore, the formula for calculating the gradient boosting tree t is transformed into the formula for summing the values ​​of each sub-leaf first, and then summing all samples of the current sub-leaf:

[0096] In the formula for calculating the t-th gradient boosting tree, g i In the objective function f t (x i According to the first derivative of the Taylor expansion, h iFor the second derivative, w t (x i )=f t (x i The method of summing the values ​​of each sub-leaf first, and then summing all samples of the current sub-leaf, is primarily used during the training of the tree model. Each tree divides the data into multiple sub-leafs, and each node only focuses on the data of its corresponding sub-leaf. By summing the sample values ​​within each sub-leaf first, and then summing the results of the entire tree, XGB effectively reduces unnecessary computation, thereby accelerating the model training process and helping XGB to more efficiently calculate the gradient and loss of each sub-leaf during gradient optimization. Therefore, the formula for calculating the gradient boosting tree t is transformed into the formula for summing the values ​​of each sub-leaf first, and then summing all samples of the current sub-leaf:

[0097]

[0098] T represents the total number of sub-leaves, and j represents the current sub-leaf number. Since we want to minimize the objective function, we find the minimum value of the loss function. The first partial derivative with respect to w is zero, and we then calculate the weight w of the j-th sub-leaf. j The calculation is as follows:

[0099]

[0100] This leads to the final expression of the current loss function, which is used to measure the quality of the t-th tree structure q.

[0101]

[0102] A greedy algorithm is used to iteratively add branches to the tree, starting from a single cotyledon. Each branch is divided into a left subset (IL) and a right subset (IR). The splitting loss is calculated by subtracting the original loss from the loss of splitting into left and right subsets, and the splitting structure of the cotyledon with the higher loss reduction is selected. Finally, the desired tree structure is obtained.

[0103] Output: XGB regression tree prediction model

[0104]

[0105] The goal of supervised learning algorithms is to continuously adjust the model so that predicted values ​​closely approximate the labeled values. Therefore, label design is crucial. Labels not only need to be representative but also closely related to the actual requirements of the target task. Appropriately designed labels help the model learn patterns in the data more accurately, leading to more precise predictions in practical applications. In this embodiment, a continuous and effective mapping method is used for the labels, avoiding obvious bifurcation in the prediction results. When the labels are designed to have physical meaning and conform to the probe distribution, during model training, the prediction results will naturally converge towards this actual distribution of depth information based on measurement accuracy, avoiding prediction bias caused by discontinuous label intervals. Therefore, reasonable label design plays a vital role in improving the model's generalization ability and accuracy.

[0106] In one or more embodiments, the PET detection depth prediction method includes: acquiring coincidence event data collected by a test crystal detector; inputting the coincidence event data into a PET detection depth prediction model to obtain a predicted depth, wherein the PET detection depth prediction model is trained according to the above method.

[0107] The PET detector tag data mapping method, the PET detection depth prediction model training method, and the PET detection depth prediction method in this embodiment have the following advantages:

[0108] This invention effectively solves the problem of discrete labels in traditional methods by mapping DOI labels to a range of data measurement accuracy and ensuring that the labels are uniformly distributed across all depth locations from 1.0 to 20.0 mm. In existing technologies, discrete labels typically only cover fixed depth locations (e.g., 1.5, 2.5, 3.5 mm), failing to adequately consider the continuity between depth labels, resulting in poor prediction accuracy at these non-label locations (e.g., 2.0, 3.0 mm). In contrast, this example uses uniformly distributed label mapping, ensuring sufficient learning references at each depth location, making the training data more balanced and better reflecting the continuous changes in actual depth. This significantly improves the accuracy of depth prediction, especially for applications requiring refined imaging (e.g., small animal imaging and brain imaging).

[0109] Secondly, this scheme fully considers the data measurement accuracy of the detector and sets labels for each depth position through precise mathematical calculations. This method breaks through the limitations of traditional methods that simply obtain labels based on position, and can better simulate the working conditions and physical limitations of actual detectors, ensuring that the label settings are more in line with actual application scenarios, thereby improving the physical rationality of the training process and the prediction accuracy.

[0110] Furthermore, because this approach provides more accurate label mapping and reduces the label discreteness problem in traditional methods, the model can learn more continuous depth information during training. This is particularly important for high-precision imaging requirements (such as precise scanning of brain, tumor, and other regions). This method not only improves prediction accuracy but also lays the foundation for future improvements in the depth resolution of PET detectors.

[0111] In summary, the advantages of this approach compared to existing technologies are as follows: it improves the accuracy of depth prediction, especially in unlabeled locations where prediction is difficult; it optimizes the label mapping method, making the training process more consistent with physical reality and avoiding the discrete label problem in traditional methods; and it enhances the generalization ability of the model, providing reliable technical support for high-precision imaging requirements.

[0112] The feasibility of this scheme has been experimentally verified, with experiments conducted on two different LYSO detector arrays. Specifically, tests were performed on an 8×8 detector array (1.5mm crystal cell size) and a 4×4 detector array (3mm crystal cell size). Energy information from 8 channels was acquired using a dual-end readout SiPM as the training input for the model. Experimental results show that, based on mapped labels and training with 8-channel energy features, the XGBoost model achieves a predicted depth resolution (DOI) of 1.76mm for 64 crystals in an 8×8 detector array, an improvement of 11.96% compared to the energy ratio (depth resolution of 2.00mm) obtained using only dual-end SiPMs. Furthermore, by acquiring additional data at positions 4.0, 7.0, 12.0, and 15.0, without additional training, the depth DOI was improved from the mathematically calculated 1.97mm resolution to 1.90mm using the model trained at positions 1.5mm to 19.5mm, further validating the model's effectiveness and generalization ability.

[0113] In a 4×4 detector array, 16 crystals were used to train an XGBoost model based on mapped tags and the same 8-channel energy features. Experimental results showed that the predicted depth of DOI reached 3.22 mm, an improvement of 26.76% compared to the traditional method which only used the energy ratio of the two SiPMs (depth resolution of 4.40 mm). Similarly, by acquiring additional data at 8.0 mm and 15.0 mm, the untrained model improved the predicted depth of DOI from the mathematically calculated 4.22 mm resolution to 3.75 mm, an improvement of 15.04%. These results demonstrate that the present invention can effectively improve depth resolution in different detector arrays and also has good model generalization ability, providing a more accurate and stable solution for DOI prediction.

[0114] Other applications of this PET detector tag data mapping method include:

[0115] a. Radiomedical imaging: This tag data mapping method can be widely used in the field of radiomedical imaging, especially in PET (positron emission tomography), SPECT (single photon emission computed tomography), and Compton imaging systems. It improves imaging resolution through precise depth positioning technology, thereby enhancing diagnostic accuracy.

[0116] b. Particle detector design: This tag data mapping method can be used for the design of particle detectors in high-energy physics experiments, especially in applications that require precise detection of the depth of particle-material interaction, such as high-energy accelerator experiments, radiation detection, and environmental monitoring.

[0117] c. Industrial Flaw Detection: In the industrial field, especially in applications such as non-destructive testing of materials and inspection of welded joints, the precise depth measurement technology provided by this label data mapping method can be used in non-destructive testing equipment to improve the accuracy and efficiency of material defect detection.

[0118] d. Nuclear Safety and Radiation Monitoring: The depth detection technology of this tag data mapping method can be applied to the field of nuclear safety, especially for the precise location and monitoring of radiation sources. This is of great significance for the safety monitoring of nuclear facilities and the detection of radiation leaks.

[0119] The specific embodiments of this application have been described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that modifications and improvements can be made to these embodiments without departing from the principles and spirit of this application as defined by the claims and their equivalents, and such modifications and improvements should also be within the protection scope of this application.

Claims

1. A method for mapping PET detector tag data, characterized in that, The PET detector tag data mapping method includes: The DOI distributions at several depth locations are calculated based on the collected coincidence event dataset. The coincidence event dataset includes coincidence event data of the tested crystal detector at several different depth locations, and the DOI distribution includes the depth value and count value of all event points at the current depth location. For each depth location, calculate the mean depth distribution of the DOI distribution and the deviation of the depth value of each event point from the mean depth distribution; For each depth location, a depth mapping value for each event point is calculated based on the various deviation values, the detection width of the test crystal detector, and the label value of the current depth location. This depth mapping value serves as the label value for each event point.

2. The PET detector tag data mapping method according to claim 1, characterized in that, The several different depth positions are several depth positions that are equally spaced, and the distance between two adjacent depth positions is the detection width of the test crystal detector.

3. The PET detector tag data mapping method according to claim 1, characterized in that, Methods for calculating the DOI distribution at several depth locations based on the collected event dataset include: For each depth location, Gaussian fitting is performed based on the coincidence event data to obtain the original DOI distribution; The error data of the original DOI distribution are corrected based on the mean and variance obtained after Gaussian fitting to obtain the corrected DOI distribution.

4. The PET detector tag data mapping method according to claim 3, characterized in that, The mean of the depth distribution of the DOI is the mean obtained after Gaussian fitting.

5. The PET detector tag data mapping method according to claim 2, characterized in that, For each depth location, based on the various deviation values, the detection width of the test crystal detector, and the tag center value at the current depth location, the depth mapping value for each event point is calculated, including: The maximum absolute value of each deviation value is taken as the maximum deviation value; The tag mapping range is determined based on the detection width of the test crystal detector; The normalized depth mapping value of each event point within the label mapping range is determined based on the deviation value and the maximum deviation value of each event point; The depth mapping value of each event point is calculated based on the normalized depth mapping value of each event point and the label center value of the current depth position.

6. The method according to claim 5, characterized in that, The label center value at the depth location is the actual detection center position value at the current depth location.

7. The PET detector tag data mapping method according to claim 5, characterized in that, The label mapping range is to w represents the detection width.

8. A training method for a PET detection depth prediction model, characterized in that, The training method includes: Acquire coincidence event data at several different depth positions of the test crystal detector, wherein the coincidence event data at each depth position includes data from multiple event points; The tag value of each event point is calculated based on the PET detector tag data mapping method according to any one of claims 1 to 7; The data and label values ​​of all event points were used as training samples to train the PET detection depth prediction model.

9. The training method for the PET detection depth prediction model according to claim 7, characterized in that, The PET detection depth prediction model is a prediction model built based on the extreme gradient boosting algorithm. The data for each event point includes energy, time, and waveform.

10. A method for predicting PET detection depth, characterized in that, The PET detection depth prediction method includes: Acquire coincidence event data collected by the test crystal detector; The coincident event data is input into the PET detection depth prediction model to obtain the predicted depth, which is trained by the method according to any one of claims 8 to 9.

Images