A gamma imaging device and method

Abstract

The application provides a kind of gamma imaging device and method, which can be applied to nuclear medical image technology field.The imaging device includes: multiple layer detector array is detected to the gamma photon generated by radioactive source at predetermined position, obtains multiple detection information;Processor obtains single photon data and Compton scattering data according to multiple detection information, predetermined single photon event condition and predetermined Compton scattering event condition;Single photon data and Compton scattering data are respectively processed using predetermined network architecture, obtain the profile reference feature and Compton constraint feature corresponding to the predetermined position;Single photon data and Compton scattering data are fused and processed, obtain the profile physical statistical feature corresponding to the predetermined position;The profile reference feature, profile physical statistical feature and Compton constraint feature are fused, obtain the target image corresponding to the predetermined position.

Description

Technical Field

[0001] This application relates to the field of nuclear medicine imaging technology, and more specifically, to a gamma imaging device and method. Background Technology

[0002] In the field of nuclear medicine, there is a need for three-dimensional imaging of the distribution of radiation sources at predetermined locations. Related technologies employ a gamma-ray "self-collimation" method (CN202010840410.2) using multi-layer detector arrays that mutually block each other. This method performs three-dimensional imaging of the distribution of radiation sources at predetermined locations based on the gamma photons they produce. However, when facing imaging scenarios with medium to high-energy gamma photons, this method experiences significantly enhanced penetration and scattering between detectors, leading to severe projection aliasing and inter-crystal scattering interference. This results in increased artifacts and a sharp decrease in the signal-to-noise ratio (SNR) in the reconstructed image, making it difficult to meet the high-performance imaging requirements for high-energy photons. Summary of the Invention

[0003] In view of this, this application provides a gamma imaging device and method.

[0004] According to one aspect of this application, a gamma imaging apparatus is provided, comprising: a multilayer detector array for detecting gamma photons generated by a radiation source at a predetermined location, obtaining multiple detection information, wherein the detection information includes the time and gamma photon energy of gamma photons incident on detectors in the multilayer detector array; and a processor for: obtaining single-photon data and Compton scattering data based on the multiple detection information, predetermined single-photon event conditions, and predetermined Compton scattering event conditions, wherein the single-photon data includes statistical values ​​of single-photon events at each detector location, and the Compton scattering data includes the location information and gamma photon energy of the detector corresponding to the Compton scattering event; processing the single-photon data and the Compton scattering data respectively using a predetermined network architecture to obtain contour reference features and Compton constraint features corresponding to the predetermined location; fusing the single-photon data and the Compton scattering data to obtain contour physical statistical features corresponding to the predetermined location; and fusing the contour reference features, the contour physical statistical features, and the Compton constraint features to obtain a target image corresponding to the predetermined location.

[0005] According to embodiments of this application, the predetermined network architecture includes a single-photon feature extraction network, a physical-driven unfolding network, a Compton feature extraction network, and a predetermined image segmentation network;

[0006] The processor is configured to: upsample the single-photon data using a single-photon feature extraction network to obtain a contour reference feature corresponding to the predetermined position; iteratively perform image reconstruction and denoising multiple times on the single-photon data and the Compton scattering data using a physical-driven unfolding network to obtain physical statistical features of the contour corresponding to the predetermined position; extract and upsample the physical geometric features of the Compton scattering data using a Compton feature extraction network to obtain Compton constraint features; and fuse the contour reference feature, the contour physical statistical features, and the Compton constraint features using a predetermined image segmentation network to obtain a target image corresponding to the predetermined position.

[0007] According to an embodiment of this application, for each detector array in the first to M-1th layers, each detector array includes multiple crystal units, multiple heavy metal units, and detection units coupled to the crystal units. The crystal units and heavy metal units in each detector array are randomly arranged. The Mth layer detector array includes multiple crystal units and detection units coupled to the crystal units. The crystal units in the Mth layer detector array are arranged in an array, where M is an integer greater than or equal to 3. The number of crystal units in the multilayer detector array increases sequentially along the direction away from the predetermined position, and the first layer detector array is closest to the predetermined position.

[0008] According to embodiments of this application, for each detector array in the first to M-1 layers, each detector array is divided into multiple first rectangular regions of equal area; in the same detector array, the multiple first rectangular regions have the same number of crystal units and the multiple first rectangular regions have the same number of heavy metal units, wherein the crystal units and heavy metal units in each first rectangular region are randomly arranged in each first rectangular region; the multiple second rectangular regions of equal area in the aforementioned M-layer detector array have the same number of crystal units, and the crystal units in each second rectangular region of the M-layer detector array are arranged in an array in each second rectangular region of the M-layer detector array.

[0009] According to embodiments of this application, the number of multiple rectangular regions of equal area in the detector array is 4 to 8; the number of rectangular regions varies depending on the detector array layer.

[0010] According to an embodiment of this application, the single-photon feature extraction network includes a first deep transposed convolution module and a first convolution module; the processor uses the single-photon feature extraction network to upsample the single-photon data to obtain contour reference features corresponding to the predetermined position, including: using the first deep transposed convolution module to iteratively perform multiple transposed convolutions and nonlinear mappings on the single-photon data to obtain initial contour reference features; and using the first convolution module to perform channel data fusion on the initial contour reference features to obtain the contour reference features.

[0011] According to an embodiment of this application, the first deep transposed convolutional module includes a first deep transposed convolutional layer, a second deep transposed convolutional layer, a third deep transposed convolutional layer, and a fourth deep transposed convolutional layer; the first convolutional module includes a first convolutional layer; the processor uses the first deep transposed convolutional module to iteratively perform multiple transposed convolutions and nonlinear mappings on the single-photon data to obtain initial contour reference features, including: using the first deep transposed convolutional layer to perform transposed convolutions and nonlinear mappings on the single-photon data to obtain first transposed convolutional features; using the second deep transposed convolutional layer to process the first transposed convolutional features... The first convolutional layer is used to perform transpose convolution and nonlinear mapping to obtain the second transpose convolutional feature; the third deep transpose convolutional layer is used to perform transpose convolution and nonlinear mapping on the second transpose convolutional feature to obtain the third transpose convolutional feature; the fourth deep transpose convolutional layer is used to perform transpose convolution and nonlinear mapping on the third transpose convolutional feature to obtain the initial contour reference feature; the processor uses the first convolutional module to perform channel data fusion on the initial contour reference feature to obtain the contour reference feature, including: using the first convolutional layer to perform channel data fusion on the initial contour reference feature to obtain the contour reference feature.

[0012] According to an embodiment of this application, the aforementioned physical drive unfolding network includes a first reconstruction and denoising module, a second reconstruction and denoising module, and a third reconstruction and denoising module. The aforementioned contour physical statistical features include a first contour statistical feature, a second contour statistical feature, and a third contour statistical feature. The processor uses the physical drive unfolding network to iteratively perform multiple image reconstructions and denoising operations on the aforementioned single-photon data and the aforementioned Compton scattering data to obtain contour physical statistical features corresponding to the aforementioned predetermined position. This includes: using the aforementioned first reconstruction and denoising module to perform image reconstruction and denoising on the aforementioned single-photon data and the aforementioned Compton scattering data to obtain the aforementioned first contour statistical feature; using the aforementioned second reconstruction and denoising module to perform image reconstruction and denoising on the aforementioned first contour statistical feature, the aforementioned single-photon data, and the aforementioned Compton scattering data to obtain the aforementioned second contour statistical feature; and using the aforementioned third reconstruction and denoising module to perform image reconstruction and denoising on the aforementioned second contour statistical feature, the aforementioned single-photon data, and the aforementioned Compton scattering data to obtain the aforementioned third contour statistical feature.

[0013] According to an embodiment of this application, the first reconstruction and denoising module includes a first reconstruction layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a first fusion layer. The processor uses the first reconstruction and denoising module to perform image reconstruction and denoising on the single-photon data and the Compton scattering data to obtain the first contour statistical feature, which includes: using the first reconstruction layer to perform image reconstruction on the single-photon data and the Compton scattering data according to the maximum likelihood maximum expectation algorithm to obtain a first reconstruction feature; using the second convolutional layer to perform feature extraction and denoising on the first reconstruction feature to obtain a first denoised feature; using the third convolutional layer to perform feature extraction and denoising on the first denoised feature to obtain a second denoised feature; using the fourth convolutional layer to perform feature extraction and denoising on the second denoised feature to obtain a third denoised feature; and using the first fusion layer to perform residual connection on the third denoised feature and the first reconstruction feature to obtain the first contour statistical feature.

[0014] According to an embodiment of this application, the second reconstruction and denoising module includes a second reconstruction layer, a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, and a second fusion layer. The processor uses the second reconstruction and denoising module to perform image reconstruction and denoising on the first contour statistical features, the single-photon data, and the Compton scattering data to obtain the second contour statistical features. This includes: using the second reconstruction layer, performing image reconstruction on the first contour statistical features, the single-photon data, and the Compton scattering data according to the maximum likelihood maximum expectation algorithm to obtain second reconstructed features; using the fifth convolutional layer to perform feature extraction and denoising on the second reconstructed features to obtain fourth denoised features; using the sixth convolutional layer to perform feature extraction and denoising on the fourth denoised features to obtain fifth denoised features; using the seventh convolutional layer to perform feature extraction and denoising on the fifth denoised features to obtain sixth denoised features; and using the second fusion layer to perform residual connection between the sixth denoised features and the second reconstructed features to obtain the second contour statistical features.

[0015] According to an embodiment of this application, the third reconstruction and denoising module includes a third reconstruction layer, an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, and a third fusion layer. The processor uses the third reconstruction and denoising module to perform image reconstruction and denoising on the second contour statistical features, the single-photon data, and the Compton scattering data to obtain the third contour statistical features, which includes: using the third reconstruction layer to perform image reconstruction on the second contour statistical features, the single-photon data, and the Compton scattering data according to the maximum likelihood maximum expectation algorithm to obtain the third reconstruction feature; using the eighth convolutional layer to perform feature extraction and denoising on the third reconstruction feature to obtain the seventh denoising feature; using the ninth convolutional layer to perform feature extraction and denoising on the seventh denoising feature to obtain the eighth denoising feature; using the tenth convolutional layer to perform feature extraction and denoising on the eighth denoising feature to obtain the ninth denoising feature; and using the third fusion layer to perform residual connection on the ninth denoising feature and the third reconstruction feature to obtain the third contour statistical feature.

[0016] According to an embodiment of this application, the Compton feature extraction network includes a point cloud network encoder, which includes a Compton scattering feature extraction module and a Compton absorption feature extraction module.

[0017] According to an embodiment of this application, the processor uses a Compton feature extraction network to extract physical geometric features and upsample the Compton scattering data to obtain Compton constraint features. This includes: using the Compton scattering feature extraction module, based on the same first predetermined neural network weights, performing nonlinear mapping on the position information of the first detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data to obtain first nonlinear mapping features; and performing channel data fusion on the first nonlinear mapping features to obtain Compton scattering features, wherein the first detector is capable of partially absorbing gamma photons. The detector simultaneously performs partial scattering of gamma photons; using the aforementioned Compton absorption feature extraction module, based on the same second predetermined neural network weights, nonlinearly maps the position information of the second detector and the gamma photon energy corresponding to each Compton scattering event included in the aforementioned Compton scattering data to obtain a second nonlinear mapping feature, and performs channel data fusion on the aforementioned second nonlinear mapping feature to obtain a Compton absorption feature, wherein the aforementioned second detector is a detector that absorbs the gamma photons scattered by the aforementioned first detector; the aforementioned Compton scattering feature and the aforementioned Compton absorption feature are spliced ​​together to obtain a physical geometric feature.

[0018] According to an embodiment of this application, the Compton feature extraction network further includes a second deep transposed convolution module and a second convolution module; the above-mentioned extraction and upsampling of physical geometric features from the Compton scattering data using the Compton feature extraction network to obtain Compton constraint features further includes: using the second deep transposed convolution module to iteratively perform multiple transposed convolutions and nonlinear mappings on the physical geometric features to obtain initial Compton constraint features; and using the second convolution module to perform channel data fusion on the initial Compton constraint features to obtain the Compton constraint features.

[0019] According to an embodiment of this application, the Compton scattering feature extraction module includes a multilayer first perceptron with shared weights, a multilayer second perceptron with shared weights, and a first max pooling layer.

[0020] According to an embodiment of this application, the aforementioned Compton scattering feature extraction module performs nonlinear mapping on the position information of the first detector corresponding to each Compton scattering event in the Compton scattering data and the gamma photon energy based on the same first predetermined neural network weights to obtain first nonlinear mapping features, and performs channel data fusion on the first nonlinear mapping features to obtain Compton scattering features. This includes: using the aforementioned multilayer first perceptron with shared weights to perform nonlinear mapping on the position information of the first detector corresponding to each Compton scattering event in the Compton scattering data and the gamma photon energy to obtain multiple first nonlinear features; using the aforementioned multilayer second perceptron with shared weights to perform nonlinear mapping on the multiple first nonlinear features to obtain multiple second nonlinear features; and using the aforementioned first max pooling layer to perform channel data fusion on the channel splicing features obtained from the multiple second nonlinear features to obtain the aforementioned Compton scattering features.

[0021] According to an embodiment of this application, the Compton absorption feature extraction module includes a multilayer third perceptron with shared weights, a multilayer fourth perceptron with shared weights, and a second max pooling layer.

[0022] According to an embodiment of this application, the processor utilizes the Compton absorption feature extraction module to perform nonlinear mapping on the position information of the second detector corresponding to each Compton scattering event in the Compton scattering data and the gamma photon energy based on the same second predetermined neural network weights, to obtain second nonlinear mapping features. Channel data fusion is then performed on the second nonlinear mapping features to obtain Compton absorption features, including: using a multilayer third perceptron with shared weights to perform nonlinear mapping on the position information of the second detector corresponding to each Compton scattering event in the Compton scattering data and the gamma photon energy, to obtain multiple third nonlinear features; using a multilayer fourth perceptron with shared weights to perform nonlinear mapping on the multiple third nonlinear features, to obtain multiple fourth nonlinear features; and using a second max-pooling layer to perform channel data fusion on the channel splicing features obtained from the multiple fourth nonlinear features to obtain the Compton absorption features.

[0023] According to an embodiment of this application, the second deep transposed convolutional module includes a fifth deep transposed convolutional layer, a sixth deep transposed convolutional layer, a seventh deep transposed convolutional layer, and an eighth deep transposed convolutional layer; the second convolutional module includes an eleventh convolutional layer.

[0024] According to an embodiment of this application, the processor uses the second deep transposed convolution module to iteratively perform multiple transposed convolutions and nonlinear mappings on the physical geometric features to obtain the initial Compton constraint features, including: using the fifth deep transposed convolution layer to perform transposed convolutions and nonlinear mappings on the Compton absorption features to obtain the fifth transposed convolution feature; using the sixth deep transposed convolution layer to perform transposed convolutions and nonlinear mappings on the fifth transposed convolution feature to obtain the sixth transposed convolution feature; using the seventh deep transposed convolution layer to perform transposed convolutions and nonlinear mappings on the sixth transposed convolution feature to obtain the seventh transposed convolution feature; and using the eighth deep transposed convolution layer to perform transposed convolutions and nonlinear mappings on the seventh transposed convolution feature to obtain the initial Compton constraint features.

[0025] According to an embodiment of this application, the processor uses the second convolutional module to perform channel data fusion on the initial Compton constraint features to obtain the Compton constraint features, including: using the eleventh convolutional layer to perform channel data fusion on the initial Compton constraint features to obtain the Compton constraint features.

[0026] According to an embodiment of this application, the predetermined image segmentation network is a Unet network.

[0027] According to an embodiment of this application, the predetermined single-photon event condition includes: the energy of the gamma photon detected by the detector is within a predetermined photon energy threshold range; the predetermined Compton scattering event condition includes: the time difference between the time when the gamma photon is incident on the first detector and the second detector of two adjacent layers is less than a predetermined duration, and the sum of the energies of the gamma photons detected by the first detector and the second detector is within a predetermined photon energy threshold range, wherein the first detector is a detector capable of partially absorbing gamma photons and partially scattering gamma photons, and the second detector is a detector capable of absorbing the gamma photons scattered by the first detector.

[0028] According to another aspect of this application, a gamma imaging method is provided, applied to the aforementioned gamma imaging apparatus. The gamma imaging method includes: detecting gamma photons generated by a radiation source at a predetermined location using a multilayer detector array to obtain multiple detection information, wherein the detection information includes the time and energy of the gamma photons incident on detectors in the multilayer detector array; and using a processor to obtain single-photon data and Compton scattering data based on the multiple detection information, predetermined single-photon event conditions, and predetermined Compton scattering event conditions, wherein the single-photon data includes the time and energy of the gamma photons incident on detectors at each detector location. The statistical values ​​of single-photon events are obtained, and the Compton scattering data includes the detector position information and gamma photon energy corresponding to the Compton scattering event. The single-photon data and the Compton scattering data are processed separately using a predetermined network architecture to obtain contour reference features and Compton constraint features corresponding to the predetermined position. The single-photon data and the Compton scattering data are fused to obtain contour physical statistical features corresponding to the predetermined position. The contour reference features, contour physical statistical features, and Compton constraint features are fused to obtain the target image corresponding to the predetermined position.

[0029] According to embodiments of this application, after obtaining single-photon data including statistical values ​​of single-photon events at each detector location and Compton scattering data including detector location information and gamma photon energy corresponding to Compton scattering events, the single-photon data and Compton scattering data are processed separately using a predetermined network architecture. This process extracts spatial correlation information from the single-photon data and extracts physical geometric features from the Compton scattering data, resulting in contour reference features and Compton constraint features corresponding to predetermined locations that reflect the distribution of radioactive sources. The single-photon data and Compton scattering data are then fused to obtain contour physical statistical features corresponding to predetermined locations that reflect the distribution of radioactive sources. Finally, the contour reference features, contour physical statistical features, and Compton constraint features are fused to maintain the system's high intrinsic spatial resolution using high-statistic single-photon events and significantly suppress background noise using the additional information provided by Compton scattering events, resulting in a high-resolution, high signal-to-noise ratio target image for medium- and high-energy gamma photons. Attached Figure Description

[0030] The above and other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0031] Figure 1 A schematic diagram of the structure of a gamma imaging device according to an embodiment of this application is shown.

[0032] Figure 2 The illustration shows a schematic diagram of a gamma imaging apparatus according to an embodiment of the present application performing self-collimation-Compton combined imaging of incident gamma photons.

[0033] Figure 3 A schematic diagram of the structure of a multilayer detector array according to an embodiment of this application is shown.

[0034] Figure 4 This schematically illustrates a network structure used by the processor in a gamma imaging apparatus according to an embodiment of this application when processing detection information; and

[0035] Figure 5 A flowchart illustrating a gamma imaging method provided according to an embodiment of this application is shown schematically. Detailed Implementation

[0036] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0037] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0038] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0039] In the embodiments of this application, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security and network security.

[0040] In the embodiments of this application, the user's authorization or consent was obtained before obtaining or collecting the user's personal information.

[0041] In the field of nuclear medicine, there is a need for three-dimensional imaging of the distribution of radiation sources at predetermined locations. Related technologies employ mechanical collimation methods to perform three-dimensional imaging of the distribution of radiation sources at predetermined locations based on gamma photons generated by these sources.

[0042] For example, related technologies employ single-photon emission computed tomography (SPECT) to create a three-dimensional image of the distribution of a radiation source at a predetermined location, based on gamma photons emitted from that source. SPECT is a key technology in nuclear medicine and is widely used in the diagnosis of tumors, cardiovascular diseases, and neurological disorders. Currently, it primarily involves medium- and high-energy nuclides (e.g.,... 177 Lu、 225 Ac、 18The integrated diagnosis and treatment technology and intraoperative real-time monitoring technology of F) have become the forefront of application.

[0043] The mechanical collimation methods employed in related technologies face inherent limitations in sensitivity and resolution. Furthermore, high-energy photons (>300 keV) cause image quality degradation due to collimator penetration effects. The "self-collimation" method in these mechanical collimation techniques utilizes the mutual blocking of multiple detector arrays to achieve collimation. While this collimation method can significantly improve spatial resolution without sacrificing detection efficiency, it significantly enhances penetration and scattering between detectors when imaging medium- to high-energy gamma photons. This leads to severe projection aliasing and inter-crystal scattering interference, resulting in increased artifacts and a sharp decrease in signal-to-noise ratio (SNR) in the reconstructed image, making it difficult to meet the high-performance imaging requirements for high-energy photons.

[0044] Figure 1 A schematic diagram of the structure of a gamma imaging device according to an embodiment of this application is shown.

[0045] like Figure 1 As shown, the gamma imaging device may include a multi-layer detector array 110 and a processor 120.

[0046] The multilayer detector array 110 can be used to detect gamma photons generated by a radiation source at a predetermined location, obtaining multiple detection information. Among them, the detection information includes the time and energy of the gamma photons incident on the detectors in the multilayer detector array.

[0047] For example, a detector array may include multiple crystal units and detection units coupled to each crystal unit. The number of crystal units included in a multilayer detector array varies. Some detector arrays may also include multiple heavy metal units. The distances of the multilayer detector arrays from the predetermined location are different for each array.

[0048] The radiation source at the predetermined location can be a radiation source located at a predetermined site on the target object. Alternatively, the radiation source at the predetermined location can be a radiation source at any location that can penetrate the target object. The target object can be a human or animal. The predetermined site can be a target organ or diseased area of ​​the target object. The radiation source can be a radionuclide. 177 Lu、 225 Ac、 18 F, etc.

[0049] Figure 1 The four layers in the multilayer detector array 110 shown are merely illustrative of its arrangement. In practical applications, the number of layers in the multilayer detector array 110 can be selected according to the specific circumstances.

[0050] The processor 120 can be used to: obtain single-photon data and Compton scattering data based on multiple detection information, predetermined single-photon event conditions, and predetermined Compton scattering event conditions; wherein the single-photon data includes statistical values ​​of single-photon events at each detector location, and the Compton scattering data includes the detector location information and gamma photon energy corresponding to the Compton scattering event; process the single-photon data and Compton scattering data separately using a predetermined network architecture to obtain contour reference features and Compton constraint features corresponding to the predetermined location; fuse the single-photon data and Compton scattering data to obtain contour physical statistical features corresponding to the predetermined location; and fuse the contour reference features, contour physical statistical features, and Compton constraint features to obtain a target image corresponding to the predetermined location.

[0051] For example, a predetermined network architecture can be used to extract features from single-photon data to obtain contour reference features corresponding to a predetermined location. A predetermined network architecture can also be used to extract features from Compton scattering data to obtain Compton constraint features corresponding to a predetermined location. Furthermore, a predetermined reconstruction algorithm can be used to iteratively reconstruct the single-photon data and Compton scattering data to obtain contour physical statistical features corresponding to a predetermined location.

[0052] According to embodiments of this application, the processor 120 can determine the location, time, and deposited energy of the incident gamma photon in each detector based on the location, time, and signal intensity information of the scintillation photon signal corresponding to the incident gamma photon. The incident gamma photon can interact with the crystal unit in the detector to generate a scintillation photon signal.

[0053] According to embodiments of this application, the gamma imaging device has sufficient energy and temporal resolution to identify and classify gamma photons of multiple energies, thereby further meeting the need for simultaneous imaging of gamma photons of multiple energies.

[0054] According to embodiments of this application, the decay of a radionuclide (i.e., a radioactive source) emits a single gamma photon. If this gamma photon is completely absorbed in the detector (i.e., all energy is deposited in a single crystal cell without Compton scattering escape), it is called a "single-photon event." Therefore, a single-photon event is an event in which all energy is deposited in a single detector.

[0055] For example, processor 120 can count the number of single-photon events at the location of any detector within a predetermined time period, obtaining a count of single-photon events corresponding to any detector. Based on the number of single-photon events at each detector, single-photon data is obtained. This single-photon data includes the number of single-photon events corresponding to the location of each detector.

[0056] According to an embodiment of this application, for a Compton scattering event, the processor 120 can determine the coincidence time window based on the time resolution of the gamma imaging device, and identify an event in which the time difference between multiple scintillation photon signals incident on the detector is less than the coincidence time window and the sum of their energies is equal to the energy of the incident gamma photon as a Compton scattering event. The processor 120 can determine the signal sequence order based on the time sequence of the scintillation photon signals incident on the gamma imaging device.

[0057] For example, processor 120 can determine whether a Compton scattering event has occurred between two adjacent detector layers based on the time difference between the time when a gamma photon is incident on the detectors of two adjacent layers (i.e., the time difference between the time when two scintillation photon signals are incident on the detectors of two adjacent layers), a predetermined duration, and the gamma photon energy obtained by the detectors of the two adjacent layers. Compton scattering data is obtained based on the location of the detectors of the two adjacent layers where the Compton scattering event occurred and the gamma photon energy detected by the detectors of the two adjacent layers where the Compton scattering event occurred.

[0058] According to an embodiment of this application, the processor 120 obtains single-photon data and Compton scattering data based on multiple detection information, predetermined single-photon event conditions, and predetermined Compton scattering event conditions. The single-photon data includes statistical values ​​of single-photon events at each detector location, and the Compton scattering data includes the location information of the detector corresponding to the Compton scattering event and the gamma photon energy. This technology enables effective differentiation and processing of single-photon events and Compton scattering events based on energy windows and time windows.

[0059] The detector's position information can include the coordinates of the detector's location in the X, Y, and Z directions.

[0060] According to embodiments of this application, after obtaining single-photon data including statistical values ​​of single-photon events at each detector location and Compton scattering data including detector location information and gamma photon energy corresponding to Compton scattering events, the single-photon data and Compton scattering data are processed separately using a predetermined network architecture. This process extracts spatial correlation information from the single-photon data and extracts physical geometric features from the Compton scattering data, resulting in contour reference features and Compton constraint features corresponding to predetermined locations that reflect the distribution of radioactive sources. The single-photon data and Compton scattering data are then fused to obtain contour physical statistical features corresponding to predetermined locations that reflect the distribution of radioactive sources. Finally, the contour reference features, contour physical statistical features, and Compton constraint features are fused to maintain the system's high intrinsic spatial resolution using high-statistic single-photon events and significantly suppress background noise using the additional information provided by Compton scattering events, resulting in a high-resolution, high signal-to-noise ratio target image for medium- and high-energy gamma photons.

[0061] According to embodiments of this application, the predetermined network architecture may include a single-photon feature extraction network, a physical-driven unfolding network, a Compton feature extraction network, and a predetermined image segmentation network.

[0062] According to embodiments of this application, the processor can also be used to: upsample single-photon data using a single-photon feature extraction network to obtain contour reference features corresponding to a predetermined position; iteratively perform multiple image reconstructions and denoising on single-photon data and Compton scattering data using a physical-driven unfolding network to obtain contour physical statistical features corresponding to the predetermined position; extract and upsample physical geometric features of Compton scattering data using a Compton feature extraction network to obtain Compton constraint features; and fuse contour reference features, contour physical statistical features, and Compton constraint features using a predetermined image segmentation network to obtain a target image corresponding to the predetermined position.

[0063] According to embodiments of this application, after obtaining single-photon data including statistical values ​​of single-photon events at various detector locations and Compton scattering data including detector location information and gamma photon energy corresponding to Compton scattering events, the single-photon data is upsampled using a single-photon feature extraction network to extract spatial correlation information, resulting in contour reference features reflecting the distribution of radioactive sources corresponding to predetermined locations. A physics-driven unfolding network iteratively performs multiple image reconstructions and denoising on the single-photon and Compton scattering data, constraining the feature extraction process using physical statistical laws to obtain contour physical statistical features reflecting the distribution of radioactive sources corresponding to predetermined locations. A Compton feature extraction network is used to extract physical geometric features and upsample the Compton scattering data, treating each Compton scattering event as an independent point in a high-dimensional space for analysis, resulting in Compton-constrained features. By using a predefined image segmentation network to fuse contour reference features, contour physical statistical features, and Compton constraint features, the system maintains high intrinsic spatial resolution using high statistical single-photon events and significantly suppresses background noise using additional information provided by Compton scattering events, resulting in high-resolution, high signal-to-noise ratio target images for medium- and high-energy gamma photons.

[0064] According to embodiments of this application, the predetermined single-photon event condition may include: the energy of the gamma photon detected by the detector being within a predetermined photon energy threshold range.

[0065] For example, the processor 120 can compare the gamma photon energy of a gamma photon incident on any detector with a predetermined energy, and determine that a single photon event has occurred at any detector location if the gamma photon energy is within the predetermined photon energy threshold range.

[0066] According to an embodiment of this application, the predetermined Compton scattering event condition may include: the time difference between the time when a gamma photon is incident on a first detector and a second detector in two adjacent layers is less than a predetermined duration, and the sum of the energies of the gamma photons detected by the first detector and the second detector is within a predetermined photon energy threshold range, wherein the first detector is a detector capable of partially absorbing gamma photons and partially scattering gamma photons, and the second detector is a detector capable of absorbing gamma photons scattered by the first detector.

[0067] For example, the processor 120 can determine that a Compton scattering event occurs between the first and second detectors if the time difference between the time when a gamma photon is incident on the first detector and the second detector of two adjacent layers is less than a predetermined duration and the sum of the energies of the gamma photons detected by the first detector and the second detector is within a predetermined photon energy threshold range.

[0068] According to an embodiment of this application, during the process where the processor 120 uses a physics-driven unfolding network to iteratively perform multiple image reconstructions and denoising operations on the single-photon data and the Compton scattering data to obtain the contour physical statistical features corresponding to the predetermined position: the processor 120 can determine the possible incident path of the gamma photon based on the self-collimation effect formed by the blocking effect of the detector array on the gamma photon; for Compton scattering events, the possible gamma photon path can first be determined based on the self-collimation effect, and then the scattering angle can be calculated using the kinematic characteristics of the Compton scattering events occurring between detector layers to determine the Compton cone surface from which the gamma photon originates; finally, by calculating the intersection of the incident path of the gamma photon and the Compton cone surface in three-dimensional space, the incident path of the gamma photon is constrained and filtered, thereby eliminating projection aliasing and achieving an accurate solution for the position of the radiation source. Simultaneously, by comprehensively utilizing single-photon events and Compton scattering events, the imaging effect of the gamma camera for medium- and high-energy photons can be significantly improved, increasing detection efficiency and signal-to-noise ratio while maintaining the same resolution.

[0069] Figure 2 The illustration shows a schematic diagram of a gamma imaging apparatus according to an embodiment of the present application performing self-collimation-Compton combined imaging of incident gamma photons.

[0070] like Figure 2 As shown in (a), the response of a multilayer detector array to a single incident high-energy gamma photon has multiple interfering photon paths.

[0071] exist Figure 2In (a) of the diagram, after determining that a single-photon event occurs at any detector position in the multilayer detector array, the incident path of the gamma photon that can be incident on that detector position can be inferred based on the structure of the multilayer detector array, and there are multiple incident paths. For example, the incident path of the gamma photon that can be incident on that detector position can be the path corresponding to the red beam or the path corresponding to the green beam. The path corresponding to the green beam can be the true incident path, and the path corresponding to the red beam is the interference path. The imaging plane is the image plane where the target image is located.

[0072] like Figure 2 As shown in (b), high-energy gamma photons incident on a multilayer detector array undergo Compton scattering in the detectors of two adjacent layers. The first detector of the two detectors undergoing Compton scattering partially absorbs the incident gamma photons and partially scatters them (the incident path is the path corresponding to the green beam, and the scattered path is the path corresponding to the blue beam), while the second detector absorbs the gamma photons scattered by the first detector.

[0073] like Figure 2 As shown in (c), the correct gamma photon incident path can be obtained by superimposing the Compton cone obtained by Compton scattering and the incident path of the gamma photon corresponding to the single photon event determined based on the self-collimation response.

[0074] like Figure 1 and Figure 2 As shown, the gamma imaging device may include a mosaic array 110 of multiple layers arranged sequentially along the incident direction of gamma photons. Each layer of the multilayer detector array 110, from layer 1 to layer M-1, is composed of crystal units and high atomic number heavy metal units (e.g., tungsten) arranged in a specific mosaic pattern. The M-th layer detector array includes multiple crystal units. These crystal units can be scintillation crystal units, for example, GAGG(Ce) crystals. Each yellow region represents one crystal unit, and each black region represents one heavy metal unit.

[0075] For each detector array in layers 1 to M-1, each detector array may include multiple crystal units, multiple heavy metal units, and detection units coupled to the crystal units. The crystal units and heavy metal units in each detector array are randomly arranged. The M-th detector array includes multiple crystal units and detection units coupled to the crystal units. The crystal units in the M-th detector array are arranged in an array, where M is an integer greater than or equal to 3. The number of crystal units in the multi-layer detector array increases sequentially away from the predetermined position, with the first-layer detector array being closest to the predetermined position and the M-th detector array being farthest from the predetermined position.

[0076] For example, M can be 3, 4, 5, 6, 7, 8, 9, or 10, etc.

[0077] For example, each crystal unit is coupled one-to-one with a silicon photomultiplier tube (SiPM), and the signal output by the silicon photomultiplier tube (SiPM) is weighted by an ASIC chip to achieve accurate calculation of the detector position affected by the incident gamma photon and the energy deposited in the detector at the corresponding position.

[0078] According to an embodiment of this application, in order to balance detection efficiency and self-collimation performance, the multilayer detector array 110 adopts a non-uniform interlayer design, that is, from the front layer closest to the radiation source to the rear layer farthest away, the number of crystal units gradually increases until full row is filled to improve detection efficiency, and the number of heavy metal units gradually decreases until 0 to increase the blocking ability of high-energy gamma photons.

[0079] For each detector array layer from layer 1 to layer M-1, each detector array layer can be divided into multiple first rectangular regions of equal area. Within the same detector array layer, multiple first rectangular regions contain the same number of crystal units and the same number of heavy metal units. The crystal units and heavy metal units within each first rectangular region are randomly arranged within that region.

[0080] The M-th layer detector array contains multiple equal-area second rectangular regions with the same number of crystal cells. The crystal cells within each second rectangular region of the M-th layer detector array are arranged in an array within that region. The arrangement of the crystal cells within the multiple second rectangular regions of the M-th layer detector array differs.

[0081] The number of equal-area rectangular regions in a detector array can be four to eight. The number of rectangular regions corresponding to different layers of the detector array can be different.

[0082] According to embodiments of this application, the number of rectangular regions corresponding to detector arrays in different layers can be completely different or partially different.

[0083] According to the embodiments of this application, by gradually increasing the number of scintillation crystals in each layer of the detector array, the scintillation crystal units and heavy metal units in each layer of the detector array from layer 1 to layer M-1 are randomly distributed in blocks. For example, each layer is divided into 4 to 8 equally sized rectangular regions, with the same number of scintillation crystals and heavy metal units in multiple rectangular regions, and they are randomly arranged in the region. The scintillation crystal units in the detector array of layer M are arranged in a block array to ensure the uniformity of sampling between rectangular regions of each layer and between different layers, and to ensure that the detection efficiency of each layer remains approximately the same.

[0084] According to embodiments of this application, in order to improve Compton performance while maintaining self-collimation performance, the size of the crystal unit has been optimized so that it has a certain blocking ability for medium and high energy gamma photons, while ensuring that most of the scattered photons can leave the crystal unit after Compton scattering occurs within the crystal unit. That is, the optimized size of the crystal unit can ensure that enough Compton scattering events can occur after gamma photons are incident on the multilayer detector array, so as to provide more reference information for constructing the target image.

[0085] According to embodiments of this application, during the optimization of a gamma imaging device, the contrast recovery coefficient (CRC)-noise level (VAR) curve of the system can be obtained by calculating the system transfer matrix under various random arrangement schemes, and the optimal detector array arrangement scheme can be selected. For example, after the gamma imaging device is optimized, for gamma photons of 300-700 keV, the system can utilize... Figure 3 The structure of the multilayer detector array shown in Table 1 is used for detection.

[0086] Figure 3 A schematic diagram of the structure of a multilayer detector array according to an embodiment of this application is shown.

[0087] like Figure 3 As shown, the number of multilayer detector arrays can be four. The multilayer detector array consists of a first-layer detector array, a second-layer detector array, a third-layer detector array, and a fourth-layer detector array. Each yellow region represents a crystal unit, and each black region represents a heavy metal unit.

[0088] From the first layer of detector array closest to the radiation source to the fourth layer of detector array furthest from the radiation source, the number of crystal units gradually increases until the array is full to improve detection efficiency, while the number of heavy metal units gradually decreases until it is 0 to increase the blocking ability of high-energy gamma photons.

[0089] Figure 3 The size and number of crystal units and heavy metal units in each detector array layer are designed as shown in Table 1.

[0090] Table 1

[0091]

[0092] As shown in Table 1, the size of the crystal units and heavy metal units in the detector arrays from the first layer to the third layer can all be 3×3×3 mm. 3 The crystal units and heavy metal units in the fourth-layer detector array can both be 2×2×6 mm in size. 3 .

[0093] The first layer detector array contains 256 crystal units and 256 heavy metal units. The second layer detector array contains 384 crystal units and 168 heavy metal units. The third layer detector array contains 512 crystal units and 88 heavy metal units. The fourth layer detector array contains 4096 crystal units and 0 heavy metal units.

[0094] The first layer detector can be divided into four equal-area first rectangular regions, the second layer detector can be divided into six equal-area first rectangular regions, the third layer detector can be divided into six equal-area first rectangular regions, and the fourth layer detector can be divided into eight equal-area second rectangular regions.

[0095] Can be used Figure 3 The multilayer detector array shown in Table 1 detects gamma photons in the range of 300-700 keV, obtaining multiple detection information. Then, the processor 120 processes the multiple detection information to obtain the target image.

[0096] The following will be based on Figure 4 right Figure 1 The network structure used by the processor in the gamma imaging device to process the detection information is explained.

[0097] Figure 4 The diagram illustrates a schematic representation of the network structure used by the processor in a gamma imaging apparatus according to an embodiment of this application when processing detection information.

[0098] like Figure 4 As shown, the network used by the processor in the gamma imaging device to process the detection information includes a single-photon feature extraction network (in... Figure 4 (Represented by Single-Photon Net in Chinese) 410, Physics-Driven Unfolding Network (in...) Figure 4(Represented by Physicsal Net in Chinese) 420, Compton Feature Extraction Network (in...) Figure 4 (represented by Compton Net) 430 and a predefined image segmentation network (in Figure 4 (Unet is used to represent this in Chinese) 440.

[0099] For example, the predetermined image segmentation network 440 can be a fully convolutional encoding and decoding symmetric semantic segmentation network (U-Net: Convolutional Networks for Biomedical Image Segmentation, Unet).

[0100] For example, the Unet network can be used to fuse image features obtained from various network paths and generate an image to obtain a reconstructed image, i.e., the target image (in...). Figure 4 (Represented by Output Image in Chinese).

[0101] like Figure 4 As shown, the single-photon feature extraction network 410 may include a first deep transposed convolutional module and a first convolutional module. The first deep transposed convolutional module may include a first deep transposed convolutional layer 4111, a second deep transposed convolutional layer 4112, a third deep transposed convolutional layer 4113, and a fourth deep transposed convolutional layer 4114. The first convolutional module may include a first convolutional layer 4121.

[0102] exist Figure 4 In this context, DeconV represents a deep transposed convolutional layer, and ConV represents a convolutional layer.

[0103] The processor 120 uses the single-photon feature extraction network 410 to upsample the single-photon data to obtain the contour reference features corresponding to the predetermined position. This can include: using the first deep transposed convolution module to iteratively perform multiple transposed convolutions and nonlinear mappings on the single-photon data to obtain the initial contour reference features; and using the first convolution module to perform channel data fusion on the initial contour reference features to obtain the contour reference features.

[0104] The processor 120 utilizes a first deep transposed convolutional module to iteratively perform multiple transposed convolutions and nonlinear mappings on the single-photon data to obtain initial contour reference features, which may include: using the first deep transposed convolutional layer 4111 to process the single-photon data (in... Figure 4The first transposed convolution feature is obtained by performing transposed convolution and nonlinear mapping on the first transposed convolution feature (represented by Single-Photon Data). The second transposed convolution feature is obtained by performing transposed convolution and nonlinear mapping on the first transposed convolution feature using a second deep transposed convolution layer 4112. The third transposed convolution feature is obtained by performing transposed convolution and nonlinear mapping on the second transposed convolution feature using a third deep transposed convolution layer 4113. The third transposed convolution feature is obtained by performing transposed convolution and nonlinear mapping on the third transposed convolution feature using a fourth deep transposed convolution layer 4114. The initial contour reference feature is obtained.

[0105] Single-photon data (in) Figure 4 The dimension of the transposed convolutional feature (represented by Single-Photon Data) is 1*5248. The dimension of the first transposed convolutional feature is 5×5×512, the dimension of the second transposed convolutional feature is 25×25×256, the dimension of the third transposed convolutional feature is 50×25×128, and the dimension of the fourth transposed convolutional feature is 100×100×64.

[0106] The processor 120 uses a first convolutional module to perform channel data fusion on the initial contour reference features to obtain contour reference features. This can include: using a first convolutional layer 4121 to perform channel data fusion on the initial contour reference features to obtain contour reference features. The contour reference features have a dimension of 100×100.

[0107] For example, the first convolutional layer 4121 can be used to calculate the average or maximum value of the data from multiple channels at each pixel location of the initial contour reference feature to obtain the contour reference feature.

[0108] According to embodiments of this application, for fixed-length, high-dimensional sparse counting vectors (i.e., single-photon data) obtained by a gamma imaging device, embodiments of this application construct a feature mapping path based on deep transposed convolution (Deconv), namely a single-photon feature extraction network 410. This single-photon feature extraction network 410 can learn a nonlinear inverse mapping relationship from the sensor domain to the image domain. Through multi-level upsampling and nonlinear activation operations, the single-photon feature extraction network 410 can extract implicit spatial correlation information from the original Poisson statistical distribution, directly recovering the macroscopic anatomical contour and global topological structure of the radionuclide distribution, and obtaining contour reference features corresponding to a predetermined location that reflect the distribution of the radioactive source.

[0109] like Figure 4 As shown, the physics-driven unfolded network 420 may include a first reconstruction and denoising module, a second reconstruction and denoising module, and a third reconstruction and denoising module.

[0110] According to embodiments of this application, the number of convolutional layers connected to each reconstruction layer can be selected based on actual conditions. For example, the number of convolutional layers connected to each reconstruction layer can be 3, 4, 5, or 6, etc.

[0111] For example, the number of convolutional layers connected to each reconstruction layer can be three. The first reconstruction and denoising module may include a first reconstruction layer 4211, a second convolutional layer 4212, a third convolutional layer 4213, a fourth convolutional layer 4214, and a first fusion layer (i.e., the functional layer implemented by the plus sign "+" between the fourth convolutional layer 4214 and the second reconstruction layer 4221). The second reconstruction and denoising module may include a second reconstruction layer 4221, a fifth convolutional layer 4222, a sixth convolutional layer 4223, a seventh convolutional layer 4224, and a second fusion layer (i.e., the functional layer implemented by the plus sign "+" between the seventh convolutional layer 4224 and the third reconstruction layer 4231). The third reconstruction and denoising module may include a third reconstruction layer 4231, an eighth convolutional layer 4232, a ninth convolutional layer 4233, a tenth convolutional layer 4234, and a third fusion layer (i.e., the functional layer implemented by the plus sign "+" between the tenth convolutional layer 4234 and the predetermined image segmentation network 440).

[0112] The physical statistical features of a contour may include a first contour statistical feature, a second contour statistical feature, and a third contour statistical feature.

[0113] Processor 120 uses a physics-driven unfolded network to iteratively perform image reconstruction and denoising multiple times on single-photon data and Compton scattering data to obtain contour physical statistical features corresponding to a predetermined position. This may include: using a first reconstruction and denoising module to reconstruct and denoise single-photon data and Compton scattering data (in... Figure 4 The first contour statistical feature is obtained by performing image reconstruction and denoising on the first contour statistical feature, single-photon data and Compton scattering data using the second reconstruction and denoising module; the second contour statistical feature is obtained by performing image reconstruction and denoising on the second contour statistical feature, single-photon data and Compton scattering data using the third reconstruction and denoising module.

[0114] The Compton scattering data has a dimension of N×8. Here, N represents the number of Compton scattering events, which is an integer greater than 1 and needs to be determined based on the actual situation. 8 represents the amount of data corresponding to each Compton scattering event. Four of the eight data points are the position information and gamma photon energy of the first detector corresponding to the Compton scattering event, and the other four are the position information and gamma photon energy of the second detector corresponding to each Compton scattering event. The first detector is capable of partially absorbing and partially scattering gamma photons, and the second detector absorbs the gamma photons scattered by the first detector.

[0115] exist Figure 4 In this context, the multiplication sign “×” indicates that the single-photon data and Compton scattering data are input as a single data set into subsequent modules.

[0116] The processor 120 uses a first reconstruction and denoising module to perform image reconstruction and denoising on single-photon data and Compton scattering data to obtain a first contour statistical feature, which may include: using a first reconstruction layer 4211 to perform image reconstruction on single-photon data and Compton scattering data according to the Maximum Likelihood Expectation Maximization (MLEM) algorithm to obtain a first reconstructed feature; using a second convolutional layer 4212 to extract and denoise the first reconstructed feature to obtain a first denoised feature; using a third convolutional layer 4213 to extract and denoise the first denoised feature to obtain a second denoised feature; using a fourth convolutional layer 4214 to extract and denoise the second denoised feature to obtain a third denoised feature; and using a first fusion layer to perform residual connection between the third denoised feature and the first reconstructed feature to obtain the first contour statistical feature.

[0117] exist Figure 4 In this context, the MLEM Layer represents the reconstruction layer.

[0118] For example, a lightweight convolutional neural network (CNN) corresponding to the convolutional layers in a physically driven unfolded network can be a residual neural network (ResNet).

[0119] The first reconstruction feature has a dimension of 100×100, the third denoising feature has a dimension of 100×100, and the first contour statistical feature has a dimension of 100×100.

[0120] The processor 120 uses a second reconstruction and denoising module to perform image reconstruction and denoising on the first contour statistical features, single-photon data, and Compton scattering data to obtain the second contour statistical features. This can include: using a second reconstruction layer 4221 to perform image reconstruction on the first contour statistical features, single-photon data, and Compton scattering data according to the maximum likelihood maximum expectation algorithm to obtain the second reconstructed features; using a fifth convolutional layer 4222 to perform feature extraction and denoising on the second reconstructed features to obtain the fourth denoised features; using a sixth convolutional layer 4223 to perform feature extraction and denoising on the fourth denoised features to obtain the fifth denoised features; using a seventh convolutional layer 4224 to perform feature extraction and denoising on the fifth denoised features to obtain the sixth denoised features; and using a second fusion layer to perform residual connection between the sixth denoised features and the second reconstructed features to obtain the second contour statistical features.

[0121] The second reconstruction feature has a dimension of 100×100, the sixth denoising feature has a dimension of 100×100, and the second contour statistical feature has a dimension of 100×100.

[0122] The processor 120 uses a third reconstruction and denoising module to perform image reconstruction and denoising on the second contour statistical features, single-photon data, and Compton scattering data to obtain the third contour statistical features. This can be achieved by: using a third reconstruction layer 4231 to perform image reconstruction on the second contour statistical features, single-photon data, and Compton scattering data according to the maximum likelihood maximum expectation algorithm to obtain the third reconstructed features; using an eighth convolutional layer 4232 to perform feature extraction and denoising on the third reconstructed features to obtain the seventh denoised features; using a ninth convolutional layer 4233 to perform feature extraction and denoising on the seventh denoised features to obtain the eighth denoised features; using a tenth convolutional layer 4234 to perform feature extraction and denoising on the eighth denoised features to obtain the ninth denoised features; and using a third fusion layer to perform residual connection between the ninth denoised features and the third reconstructed features to obtain the third contour statistical features.

[0123] The third reconstruction feature has a dimension of 100×100, the ninth denoising feature has a dimension of 100×100, and the third contour statistical feature has a dimension of 100×100.

[0124] According to the embodiments of this application, the MLEM algorithm provided in the embodiments of this application can accurately model and calculate the response of Compton scattering events. The relevant modeling and calculation are shown in the following formulas (1) to (4).

[0125] After obtaining the Compton scattering data, each reconstruction layer can calculate the scattering angle corresponding to each Compton scattering event according to formula (1), calculate the scattering angle uncertainty of Compton scattering event n at voxel j according to formula (2), and calculate the response of Compton scattering event n at voxel j according to formula (3).

[0126] (1);

[0127] (2);

[0128] (3);

[0129] in, Compton scattering event The scattering angle, For electronic quality, At the speed of light, The energy deposited in the detector unit (detector unit 1) where Compton scattering occurs. The energy deposited in the detector unit (detector unit 2) to absorb scattered photons, Compton scattering event In voxels The scattering angle uncertainty, , and These represent the scattering angle uncertainties caused by detector position resolution, energy resolution, and the Doppler effect, respectively. Compton scattering event In voxels The response voxels —The supplementary angle of the angle formed by detector unit 1 and detector unit 2 The system transfer matrix corresponds to detector element 1 and voxel. The elements. Among them, detector unit 1 can be the first detector, and detector unit 2 can be the second detector.

[0130] After obtaining single-photon data, each reconstruction layer can calculate the voxel value at each voxel location of the reconstructed image according to the MLEM iterative algorithm shown in formula (4).

[0131] (4);

[0132] in, Let be the voxel value of voxel j obtained in the k-th iteration. For the system transmission matrix corresponding to the elements of probe unit i and voxel j, y i For the single-photon event projection count of detection unit i, Let be the single-photon event detection efficiency of voxel j. The Compton scattering event detection efficiency of voxel j.

[0133] When the first reconstruction layer 4211 calculates the voxel value at each voxel position of the first contour statistical feature according to formula (4), the initial value of the voxel value at each voxel position is a predetermined value. When the second reconstruction layer 4221 calculates the voxel value at each voxel position of the second contour statistical feature according to formula (4), the initial value of the voxel value at each voxel position is the voxel value at the corresponding voxel position of the first contour statistical feature. When the third reconstruction layer 4231 calculates the voxel value at each voxel position of the third contour statistical feature according to formula (4), the initial value of the voxel value at each voxel position is the voxel value at the corresponding voxel position of the second contour statistical feature.

[0134] According to the embodiments of this application, for scenarios where gamma photons of multiple energies are imaged simultaneously, the MLEM iterative algorithm in formula (4) can be replaced with the MLEM iterative algorithm shown in formula (5) to calculate the voxel value at each voxel position of the reconstructed image.

[0135] (5);

[0136] In this context, the subscript 'e' indicates data corresponding to gamma photons of different energies.

[0137] According to embodiments of this application, image reconstruction can also be performed using a single physical-driven unfolding network (e.g., a first reconstruction and denoising module) or solely using the MLEM algorithm provided in this application, to accelerate the image reconstruction process or reduce computational overhead. Furthermore, alternating layers of MLEM-CNN modules (i.e., reconstruction and denoising modules) can be added to the physical-driven unfolding network to extract deeper levels of image information.

[0138] For example, the number of reconstruction and denoising modules can be 1 to 10.

[0139] According to embodiments of this application, the number of reconstruction and convolutional layers in the physics-driven unfolding network can be selected according to actual conditions, and is not limited herein.

[0140] According to embodiments of this application, the early stages of the first, second, and third reconstruction and denoising modules in the physics-driven unfolded network can provide low-frequency, global structural information, while the later stages can provide high-frequency, detailed information.

[0141] According to embodiments of this application, a Physicsal Net can "unfold" the traditional MLEM iterative algorithm into a cascaded neural network structure. In each stage, the network alternately executes MLEM iterative layers containing responses to self-collimation and Compton scattering events, as well as convolutional neural network regularization layers for denoising and artifact suppression, allowing the Physicsal Net to follow physical statistics during the learning process. Simultaneously, the network inputs the images obtained from each stage into the final fusion network to provide layer-by-layer information from low to high frequencies in the reconstructed image.

[0142] like Figure 4 As shown, the Compton feature extraction network 430 may include a Point Cloud Network Encoder (PointNet), a second deep transposed convolutional module, and a second convolutional module. The Point Cloud Network Encoder may include a Compton scattering feature extraction module and a Compton absorption feature extraction module. The Compton scattering feature extraction module may include a multilayer first perceptron 4311 with shared weights, a multilayer second perceptron 4312 with shared weights, and a first max pooling layer (in... Figure 4 (Represented by Max Pooling in Chinese) 4313. The Compton Absorption feature extraction module may include a multi-layer third perceptron with shared weights 4321, a multi-layer fourth perceptron with shared weights 4322, and a second max pooling layer 4323. The second deep transposed convolution module may include a fifth deep transposed convolution layer 4331, a sixth deep transposed convolution layer 4332, a seventh deep transposed convolution layer 4333, and an eighth deep transposed convolution layer 4334. The second convolution module may include an eleventh convolution layer 4341.

[0143] exist Figure 4 In this context, Shsred is used to denote a multi-layer perceptron with shared weights (Shared MLP).

[0144] Processor 120 uses a Compton feature extraction network to extract and upsample the physical geometric features of the Compton scattering data, obtaining Compton constraint features that may include:

[0145] Using the Compton scattering feature extraction module, the position information of the first detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data are nonlinearly mapped based on the same first predetermined neural network weights to obtain the first nonlinear mapping feature. Then, the first nonlinear mapping feature is fused with channel data to obtain the Compton scattering feature. The first detector is a detector that can partially absorb gamma photons and partially scatter gamma photons.

[0146] Using the Compton absorption feature extraction module, the position information of the second detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data are nonlinearly mapped based on the same second predetermined neural network weights to obtain the second nonlinear mapping feature. The second nonlinear mapping feature is then fused with channel data to obtain the Compton absorption feature, where the second detector is the detector that absorbs the gamma photons scattered by the first detector. The Compton scattering feature and the Compton absorption feature are then concatenated to obtain the physical geometric feature.

[0147] The physical geometric features are iteratively transposed and nonlinearly mapped multiple times using the second deep transposed convolution module to obtain the initial Compton constraint features.

[0148] The initial Compton constraint features are obtained by channel data fusion using the second convolution module.

[0149] exist Figure 4 In this diagram, the position information and gamma photon energy of the first detector corresponding to each Compton scattering event are represented by Scatter, which has a dimension of N×4. The position information and gamma photon energy of the second detector corresponding to each Compton scattering event are represented by Absorber, which has a dimension of N×4.

[0150] Using the Compton scattering feature extraction module, the position information of the first detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data are nonlinearly mapped based on the same first predetermined neural network weights to obtain the first nonlinear mapping feature. Channel data fusion is then performed on the first nonlinear mapping feature to obtain the Compton scattering feature, which may include:

[0151] The first multilayer perceptron 4311 with shared weights performs nonlinear mapping on the position information of the first detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data to obtain multiple first nonlinear features. The second multilayer perceptron 4312 with shared weights performs nonlinear mapping on the multiple first nonlinear features to obtain multiple second nonlinear features. The first max pooling layer 4313 performs channel data fusion on the channel splicing features obtained from the multiple second nonlinear features to obtain the Compton scattering features.

[0152] For example, the first max pooling layer 4313 can be used to find the maximum value of multiple channel data at each pixel position of the channel splicing feature obtained from multiple second nonlinear features, thereby obtaining the Compton scattering feature.

[0153] The processor 120 utilizes the Compton absorption feature extraction module to perform nonlinear mapping on the position information of the second detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data, based on the same second predetermined neural network weights, to obtain second nonlinear mapping features. Channel data fusion is then performed on the second nonlinear mapping features to obtain Compton absorption features, which may include:

[0154] A multilayer third perceptron 4321 with shared weights is used to perform nonlinear mapping on the position information of the second detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data to obtain multiple third nonlinear features. A multilayer fourth perceptron 4322 with shared weights is used to perform nonlinear mapping on the multiple third nonlinear features to obtain multiple fourth nonlinear features. A second max-pooling layer 4323 is used to perform channel data fusion on the channel splicing features obtained from the multiple fourth nonlinear features to obtain the Compton absorption features.

[0155] The weights of the first perceptron 4311, the second perceptron 4312, the third perceptron 4321, and the fourth perceptron 4322 are all different.

[0156] The processor 120 uses a second deep transposed convolutional module to iteratively perform multiple transposed convolutions and nonlinear mappings on the physical geometric features to obtain initial Compton constraint features. This can include: using a fifth deep transposed convolutional layer 4331 to perform transposed convolutions and nonlinear mappings on the Compton absorption features to obtain fifth transposed convolutional features; using a sixth deep transposed convolutional layer 4332 to perform transposed convolutions and nonlinear mappings on the fifth transposed convolutional features to obtain sixth transposed convolutional features; using a seventh deep transposed convolutional layer 4333 to perform transposed convolutions and nonlinear mappings on the sixth transposed convolutional features to obtain seventh transposed convolutional features; and using an eighth deep transposed convolutional layer 4334 to perform transposed convolutions and nonlinear mappings on the seventh transposed convolutional features to obtain initial Compton constraint features.

[0157] The processor 120 uses the second convolutional module to perform channel data fusion on the initial Compton constraint features to obtain Compton constraint features, which may include: using the eleventh convolutional layer 4341 to perform channel data fusion on the initial Compton constraint features to obtain Compton constraint features.

[0158] For example, the eleventh convolutional layer 4341 can be used to calculate the average or maximum value of the data from multiple channels at each pixel location of the initial Compton constraint feature to obtain the Compton constraint feature.

[0159] According to embodiments of this application, the Compton feature extraction network provided by these embodiments can effectively extract useful features from a large number of Compton scattering events and convert them into fixed-dimensional global feature vectors, which can then be used for subsequent image reconstruction, improving imaging quality and diagnostic accuracy.

[0160] According to embodiments of this application, the Compton feature extraction network (ComptonNet) provided in this application addresses the unstructured, discretized, and dynamically changing event count characteristics of Compton scattering data by introducing the PointNet architecture as the core encoder. This Compton feature extraction network (ComptonNet) treats each Compton scattering event as an independent point in a high-dimensional space, utilizes a shared-weight multilayer perceptron (Shared MLP) to deeply extract physical geometric features including scattering energy deposition and interaction locations, and aggregates them into a fixed-dimensional global feature vector, i.e., Compton-constrained features, through max pooling.

[0161] According to embodiments of this application, the gamma imaging device and method provided by this application can solve the problem of noise divergence in traditional iterative algorithms under low count conditions, while overcoming the limitation of pure deep learning methods lacking physical interpretability. While maintaining high resolution, it can effectively suppress aliasing effects and utilize scattering events to improve the signal-to-noise ratio, thereby achieving high-resolution, high-detection-efficiency, and high-signal-to-noise-ratio imaging of medium- and high-energy photons.

[0162] According to embodiments of this application, the gamma imaging method provided is a joint imaging method based on self-collimation and Compton scattering. It utilizes the conical surface determined by Compton scattering events as a physical constraint to accurately filter and eliminate aliasing effects caused by path blurring in self-collimation imaging. This mechanism effectively solves the serious artifact problem faced by existing self-collimation techniques in high-energy photon imaging, significantly reducing background noise. Simultaneously, by combining deep learning networks for regularization optimization of the physical reconstruction process, the image detail recovery capability is further enhanced, enabling the reconstructed image (i.e., the target image) to maintain an extremely high signal-to-noise ratio even under low count conditions.

[0163] Based on the above-mentioned gamma imaging device, this application also provides a gamma imaging method.

[0164] Figure 5 A flowchart illustrating a gamma imaging method provided according to an embodiment of this application is shown schematically.

[0165] like Figure 5 As shown, the gamma imaging method of this embodiment includes operations S510 to S520.

[0166] In operation S510, a multi-layer detector array is used to detect gamma photons generated by a radiation source at a predetermined location, obtaining multiple detection information. This detection information includes the time and energy of the gamma photons incident on the detectors in the multi-layer detector array.

[0167] In operation of S520, the processor obtains single-photon data and Compton scattering data based on multiple detection information, predetermined single-photon event conditions, and predetermined Compton scattering event conditions. A predetermined network architecture is used to process the single-photon data and Compton scattering data separately to obtain contour reference features and Compton constraint features corresponding to predetermined locations. The single-photon data and Compton scattering data are then fused to obtain contour physical statistical features corresponding to predetermined locations. Finally, the contour reference features, contour physical statistical features, and Compton constraint features are fused to obtain the target image corresponding to the predetermined location. The single-photon data includes statistical values ​​of single-photon events at each detector location, and the Compton scattering data includes the detector location information and gamma photon energy corresponding to the Compton scattering events.

[0168] According to embodiments of this application, the predetermined network architecture includes a single-photon feature extraction network, a physically driven unfolding network, a Compton feature extraction network, and a predetermined image segmentation network. The processor is configured to: upsample single-photon data using the single-photon feature extraction network to obtain contour reference features corresponding to a predetermined location; iteratively perform multiple image reconstructions and denoising operations on the single-photon data and Compton scattering data using the physically driven unfolding network to obtain contour physical statistical features corresponding to the predetermined location; extract and upsample physical geometric features from the Compton scattering data using the Compton feature extraction network to obtain Compton constraint features; and fuse the contour reference features, contour physical statistical features, and Compton constraint features using the predetermined image segmentation network to obtain a target image corresponding to the predetermined location.

[0169] According to an embodiment of this application, the single-photon feature extraction network includes a first deep transposed convolution module and a first convolution module; the processor uses the single-photon feature extraction network to upsample the single-photon data to obtain contour reference features corresponding to a predetermined position, including: using the first deep transposed convolution module to iteratively perform multiple transposed convolutions and nonlinear mappings on the single-photon data to obtain initial contour reference features; and using the first convolution module to perform channel data fusion on the initial contour reference features to obtain contour reference features.

[0170] According to an embodiment of this application, the first deep transposed convolutional module includes a first deep transposed convolutional layer, a second deep transposed convolutional layer, a third deep transposed convolutional layer, and a fourth deep transposed convolutional layer; the first convolutional module includes a first convolutional layer.

[0171] According to an embodiment of this application, the processor uses a first deep transposed convolutional module to iteratively perform multiple transposed convolutions and nonlinear mappings on single-photon data to obtain initial contour reference features, including: performing transposed convolutions and nonlinear mappings on single-photon data using a first deep transposed convolutional layer to obtain first transposed convolution features; performing transposed convolutions and nonlinear mappings on the first transposed convolution features using a second deep transposed convolutional layer to obtain second transposed convolution features; performing transposed convolutions and nonlinear mappings on the second transposed convolution features using a third deep transposed convolutional layer to obtain third transposed convolution features; and performing transposed convolutions and nonlinear mappings on the third transposed convolution features using a fourth deep transposed convolutional layer to obtain initial contour reference features. The processor uses a first convolutional module to perform channel data fusion on the initial contour reference features to obtain contour reference features, including: performing channel data fusion on the initial contour reference features using a first convolutional layer to obtain contour reference features.

[0172] According to an embodiment of this application, the physical drive unfolding network includes a first reconstruction and denoising module, a second reconstruction and denoising module, and a third reconstruction and denoising module, and the contour physical statistical features include a first contour statistical feature, a second contour statistical feature, and a third contour statistical feature.

[0173] According to embodiments of this application, the processor uses a physical drive unfolding network to iteratively perform image reconstruction and denoising multiple times on single-photon data and Compton scattering data to obtain contour physical statistical features corresponding to a predetermined position. This includes: using a first reconstruction and denoising module to perform image reconstruction and denoising on single-photon data and Compton scattering data to obtain a first contour statistical feature; using a second reconstruction and denoising module to perform image reconstruction and denoising on the first contour statistical feature, single-photon data, and Compton scattering data to obtain a second contour statistical feature; and using a third reconstruction and denoising module to perform image reconstruction and denoising on the second contour statistical feature, single-photon data, and Compton scattering data to obtain a third contour statistical feature.

[0174] According to an embodiment of this application, the Compton feature extraction network includes a point cloud network encoder, which includes a Compton scattering feature extraction module and a Compton absorption feature extraction module.

[0175] According to an embodiment of this application, the processor uses a Compton feature extraction network to extract physical geometric features and upsample Compton scattering data to obtain Compton constraint features. This includes: using a Compton scattering feature extraction module, based on the same first predetermined neural network weights, performing nonlinear mapping on the position information of the first detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data to obtain first nonlinear mapping features; and performing channel data fusion on the first nonlinear mapping features to obtain Compton scattering features. The first detector is capable of partially absorbing gamma photons. The detector receives and partially scatters gamma photons. Using the Compton absorption feature extraction module, the position information of the second detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data are nonlinearly mapped based on the same second predetermined neural network weights to obtain the second nonlinear mapping feature. The second nonlinear mapping feature is then fused with channel data to obtain the Compton absorption feature, where the second detector is the detector that absorbs the gamma photons scattered by the first detector. The Compton scattering feature and the Compton absorption feature are then spliced ​​together to obtain the physical geometric feature.

[0176] According to embodiments of this application, the Compton feature extraction network further includes a second deep transposed convolution module and a second convolution module. The process of extracting and upsampling physical geometric features from Compton scattering data using the Compton feature extraction network to obtain Compton constraint features further includes: iteratively performing multiple transposed convolutions and nonlinear mappings on the physical geometric features using the second deep transposed convolution module to obtain initial Compton constraint features; and then using the second convolution module to perform channel data fusion on the initial Compton constraint features to obtain the final Compton constraint features.

[0177] It should be noted that the gamma imaging method part in the embodiments of this application corresponds to the gamma imaging device part in the embodiments of this application. The description of the gamma imaging method part is specifically referred to in the gamma imaging device part, and will not be repeated here.

[0178] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of this application is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this application, and all such substitutions and modifications should fall within the scope of this application.

Claims

1. A gamma imaging device, comprising: A multilayer detector array is used to detect gamma photons generated by a radiation source at a predetermined location and obtain multiple detection information, wherein the detection information includes the time and energy of the gamma photons incident on the detectors in the multilayer detector array. Processor, used for: Based on the multiple detection information, the predetermined single-photon event conditions, and the predetermined Compton scattering event conditions, single-photon data and Compton scattering data are obtained. The single-photon data includes the statistical values ​​of single-photon events at each detector location, and the Compton scattering data includes the detector location information and gamma photon energy corresponding to the Compton scattering event. The single-photon data is upsampled using a single-photon feature extraction network to obtain contour reference features corresponding to the predetermined position; The physical drive unfolding network is used to iteratively perform image reconstruction and denoising multiple times on the single photon data and the Compton scattering data to obtain the contour physical statistical features corresponding to the predetermined position; The Compton scattering data is subjected to physical geometric feature extraction and upsampling using a Compton feature extraction network to obtain Compton constraint features. The contour reference features, the contour physical statistical features, and the Compton constraint features are fused using a predetermined image segmentation network to obtain a target image corresponding to the predetermined position. The predetermined network architecture includes a single-photon feature extraction network, a physical-driven unfolding network, a Compton feature extraction network, and a predetermined image segmentation network.

2. The gamma imaging device according to claim 1, wherein, For each detector array in the detector array from the first layer to the (M-1)th layer, each detector array includes multiple crystal units, multiple heavy metal units, and detection units coupled to the crystal units. The crystal units and heavy metal units in each detector array are randomly arranged in each detector array. The M-th layer detector array includes multiple crystal units and detection units coupled to the crystal units. The crystal units in the M-th layer detector array are arranged in an array, where M is an integer greater than or equal to 3. The number of crystal units in the multilayer detector array increases sequentially along the direction away from the predetermined position, with the first layer detector array being closest to the predetermined position.

3. The gamma imaging device according to claim 2, wherein, For each detector array in the detector array from layer 1 to layer M-1, each detector array is divided into multiple first rectangular regions of equal area; In the same layer of the detector array, multiple first rectangular regions have the same number of crystal units and multiple first rectangular regions have the same number of heavy metal units, wherein the crystal units and heavy metal units in each first rectangular region are randomly arranged in each first rectangular region; The multiple equal-area second rectangular regions in the M-th layer detector array have the same number of crystal units, and the crystal units in each second rectangular region of the M-th layer detector array are arranged in an array within each second rectangular region of the M-th layer detector array.

4. The gamma imaging device according to claim 3, wherein, The detector array contains 4 to 8 rectangular regions of equal area. The number of rectangular regions corresponding to detector arrays in different layers is different.

5. The gamma imaging apparatus according to any one of claims 1 to 4, wherein, The single-photon feature extraction network includes a first deep transposed convolutional module and a first convolutional module; The processor uses a single-photon feature extraction network to upsample the single-photon data to obtain contour reference features corresponding to the predetermined position, including: The first deep transposed convolution module is used to iteratively perform multiple transposed convolutions and nonlinear mappings on the single photon data to obtain initial contour reference features. The initial contour reference features are obtained by fusing channel data using the first convolution module.

6. The gamma imaging device according to claim 5, wherein, The first deep transposed convolutional module includes a first deep transposed convolutional layer, a second deep transposed convolutional layer, a third deep transposed convolutional layer, and a fourth deep transposed convolutional layer; the first convolutional module includes a first convolutional layer; The processor uses the first deep transposed convolution module to iteratively perform multiple transposed convolutions and nonlinear mappings on the single-photon data to obtain initial contour reference features, including: The first deep transposed convolutional layer is used to perform transposed convolution and nonlinear mapping on the single photon data to obtain the first transposed convolutional feature; The second transposed convolutional feature is obtained by transposing and nonlinearly mapping the first transposed convolutional feature using the second deep transposed convolutional layer. The third deep transposed convolutional layer is used to perform transposed convolution and nonlinear mapping on the second transposed convolutional features to obtain the third transposed convolutional features; The initial contour reference features are obtained by performing transpose convolution and nonlinear mapping on the third transpose convolution features using the fourth deep transpose convolution layer. The processor uses the first convolution module to perform channel data fusion on the initial contour reference features to obtain the contour reference features, including: The initial contour reference features are obtained by fusing channel data using the first convolutional layer.

7. The gamma imaging apparatus according to any one of claims 1 to 4, wherein, The physical drive unfolding network includes a first reconstruction and denoising module, a second reconstruction and denoising module, and a third reconstruction and denoising module. The contour physical statistical features include a first contour statistical feature, a second contour statistical feature, and a third contour statistical feature. The processor uses a physics-driven unfolding network to iteratively perform image reconstruction and denoising multiple times on the single-photon data and the Compton scattering data to obtain the contour physical statistical features corresponding to the predetermined position, including: The first reconstruction and denoising module is used to perform image reconstruction and denoising on the single-photon data and the Compton scattering data to obtain the first contour statistical features. The second reconstruction and denoising module is used to perform image reconstruction and denoising on the first contour statistical features, the single-photon data and the Compton scattering data to obtain the second contour statistical features. The third contour statistical feature is obtained by using the third reconstruction and denoising module to perform image reconstruction and denoising on the second contour statistical feature, the single photon data and the Compton scattering data.

8. The gamma imaging device according to claim 7, wherein, The first reconstruction and denoising module includes a first reconstruction layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a first fusion layer; The processor uses the first reconstruction and denoising module to perform image reconstruction and denoising on the single-photon data and the Compton scattering data, obtaining the first contour statistical features, including: Using the first reconstruction layer, the image is reconstructed from the single-photon data and the Compton scattering data according to the maximum likelihood maximum expectation algorithm to obtain the first reconstruction feature; The first reconstructed feature is extracted and denoised using the second convolutional layer to obtain the first denoised feature; The third convolutional layer is used to extract and denoise the first denoised feature to obtain the second denoised feature. The fourth convolutional layer is used to extract and denoise the second denoised feature to obtain the third denoised feature. The first contour statistical feature is obtained by performing a residual connection between the third denoising feature and the first reconstructed feature using the first fusion layer.

9. The gamma imaging device according to claim 7, wherein, The second reconstruction and denoising module includes a second reconstruction layer, a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, and a second fusion layer; The processor utilizes the second reconstruction and denoising module to perform image reconstruction and denoising on the first contour statistical features, the single-photon data, and the Compton scattering data, obtaining the second contour statistical features, which include: Using the second reconstruction layer, the first contour statistical features, the single-photon data, and the Compton scattering data are reconstructed according to the maximum likelihood maximum expectation algorithm to obtain the second reconstruction features; The fifth convolutional layer is used to extract and denoise the second reconstructed feature to obtain the fourth denoised feature. The sixth convolutional layer is used to extract and denoise the fourth denoising feature to obtain the fifth denoising feature. The seventh convolutional layer is used to extract and denoise the fifth denoising feature to obtain the sixth denoising feature. The second contour statistical feature is obtained by performing a residual connection between the sixth denoising feature and the second reconstructed feature using the second fusion layer.

10. The gamma imaging device according to claim 7, wherein, The third reconstruction and denoising module includes a third reconstruction layer, an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, and a third fusion layer; The processor utilizes the third reconstruction and denoising module to perform image reconstruction and denoising on the second contour statistical features, the single-photon data, and the Compton scattering data to obtain the third contour statistical features, which include: Using the third reconstruction layer, the second contour statistical features, the single-photon data, and the Compton scattering data are reconstructed using the maximum likelihood maximum expectation algorithm to obtain the third reconstruction features; The eighth convolutional layer is used to extract and denoise the third reconstructed feature to obtain the seventh denoised feature. The seventh denoising feature is extracted and denoised using the nine convolutional layers to obtain the eighth denoising feature. The tenth convolutional layer is used to extract and denoise the eighth denoised feature to obtain the ninth denoised feature. The third contour statistical feature is obtained by performing residual connection between the ninth denoising feature and the third reconstruction feature using the third fusion layer.

11. The gamma imaging apparatus according to any one of claims 1 to 4, wherein, The Compton feature extraction network includes a point cloud network encoder, which includes a Compton scattering feature extraction module and a Compton absorption feature extraction module. The processor uses a Compton feature extraction network to extract physical geometric features and upsample the Compton scattering data, obtaining Compton constraint features including: Using the Compton scattering feature extraction module, the position information of the first detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data are nonlinearly mapped based on the same first predetermined neural network weights to obtain the first nonlinear mapping feature. Then, the first nonlinear mapping feature is fused with channel data to obtain the Compton scattering feature. The first detector is a detector that can partially absorb gamma photons and partially scatter gamma photons. Using the Compton absorption feature extraction module, the position information of the second detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data are nonlinearly mapped based on the same second predetermined neural network weights to obtain the second nonlinear mapping feature. Then, the second nonlinear mapping feature is fused with channel data to obtain the Compton absorption feature, wherein the second detector is a detector that absorbs the gamma photons scattered by the first detector. The physical geometric features are obtained by splicing the Compton scattering features and the Compton absorption features.

12. The gamma imaging device according to claim 11, wherein, The Compton feature extraction network also includes a second deep transposed convolutional module and a second convolutional module; The step of using a Compton feature extraction network to extract physical geometric features and upsample the Compton scattering data to obtain Compton constraint features also includes: The physical geometric features are iteratively transposed and nonlinearly mapped multiple times using the second deep transposed convolution module to obtain the initial Compton constraint features. The initial Compton constraint features are obtained by channel data fusion using the second convolution module.

13. The gamma imaging apparatus according to claim 11, wherein, The Compton scattering feature extraction module includes a multilayer first perceptron with shared weights, a multilayer second perceptron with shared weights, and a first max pooling layer. The Compton scattering feature extraction module uses the same first predetermined neural network weights to perform nonlinear mapping on the position information of the first detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data, respectively, to obtain a first nonlinear mapping feature. Then, channel data fusion is performed on the first nonlinear mapping feature to obtain the Compton scattering features, including: The shared-weight multilayer first sensor is used to perform nonlinear mapping on the position information of the first detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data, respectively, to obtain multiple first nonlinear features. The shared weights of the multilayer second perceptron are used to perform nonlinear mapping on the plurality of first nonlinear features to obtain a plurality of second nonlinear features. The first max pooling layer is used to perform channel data fusion on the channel splicing features obtained from the multiple second nonlinear features to obtain the Compton scattering features.

14. The gamma imaging apparatus according to claim 11, wherein, The Compton absorption feature extraction module includes a multilayer third perceptron with shared weights, a multilayer fourth perceptron with shared weights, and a second max pooling layer. The processor utilizes the Compton absorption feature extraction module to perform nonlinear mapping on the position information of the second detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data, based on the same second predetermined neural network weights, to obtain second nonlinear mapping features. Then, channel data fusion is performed on the second nonlinear mapping features to obtain Compton absorption features including: The shared-weight multilayer third perceptron is used to perform nonlinear mapping on the position information of the second detector and the gamma photon energy corresponding to each Compton scattering event in the Compton scattering data, respectively, to obtain multiple third nonlinear features. The shared weights of the multilayer fourth perceptron are used to perform nonlinear mapping on the multiple third nonlinear features to obtain multiple fourth nonlinear features. The channel data is fused using the second max pooling layer on the channel splicing features obtained from the multiple fourth nonlinear features to obtain the Compton absorption features.

15. The gamma imaging apparatus according to claim 12, wherein, The second deep transposed convolutional module includes a fifth deep transposed convolutional layer, a sixth deep transposed convolutional layer, a seventh deep transposed convolutional layer, and an eighth deep transposed convolutional layer; the second convolutional module includes an eleventh convolutional layer; The processor uses the second deep transposed convolution module to iteratively perform multiple transposed convolutions and nonlinear mappings on the physical geometric features to obtain the initial Compton constraint features, including: The fifth deep transposed convolutional layer is used to perform transposed convolution and nonlinear mapping on the Compton absorption features to obtain the fifth transposed convolutional features; The sixth transposed convolutional feature is obtained by transposing and nonlinearly mapping the fifth transposed convolutional feature using the sixth deep transposed convolutional layer. The seventh transposed convolutional feature is obtained by transposing and nonlinearly mapping the sixth transposed convolutional feature using the seventh deep transposed convolutional layer. The initial Compton-constrained features are obtained by performing transpose convolution and nonlinear mapping on the seventh transpose convolution features using the eighth deep transpose convolution layer. The processor uses the second convolution module to perform channel data fusion on the initial Compton constraint features to obtain the Compton constraint features, which include: The initial Compton constraint features are fused using the eleventh convolutional layer to obtain the Compton constraint features.

16. The gamma imaging apparatus according to any one of claims 1 to 4, wherein, The predetermined image segmentation network is a Unet network.

17. The gamma imaging apparatus according to any one of claims 1 to 4, wherein, The predetermined single-photon event conditions include: the energy of the gamma photon detected by the detector is within a predetermined photon energy threshold range; The predetermined Compton scattering event conditions include: the time difference between the time when a gamma photon is incident on the first detector and the second detector of two adjacent layers is less than a predetermined duration, and the sum of the energies of the gamma photons detected by the first detector and the second detector is within a predetermined photon energy threshold range, wherein the first detector is a detector capable of partially absorbing gamma photons and partially scattering gamma photons, and the second detector is a detector capable of absorbing gamma photons scattered by the first detector.

18. A gamma imaging method, applied to the gamma imaging apparatus according to any one of claims 1 to 17, the gamma imaging method comprising: The gamma photons generated by a radiation source at a predetermined location are detected using a multi-layer detector array, and multiple detection information is obtained. The detection information includes the time and energy of the gamma photons incident on the detectors in the multi-layer detector array. The processor uses the multiple detection information, predetermined single-photon event conditions, and predetermined Compton scattering event conditions to obtain single-photon data and Compton scattering data. The single-photon data includes statistical values ​​of single-photon events at each detector location, and the Compton scattering data includes the detector location information and gamma photon energy corresponding to the Compton scattering event. The single-photon data and the Compton scattering data are processed separately using a predetermined network architecture to obtain contour reference features and Compton constraint features corresponding to the predetermined position; the single-photon data and the Compton scattering data are fused to obtain contour physical statistical features corresponding to the predetermined position; the contour reference features, the contour physical statistical features, and the Compton constraint features are fused to obtain the target image corresponding to the predetermined position.

Images