Method, apparatus, device, medium and program product for determining physical parameters of a projection domain tissue

Abstract

The present disclosure provides a projection domain tissue physical parameter determination method, device, equipment, medium and program product, which can be applied to the field of medical image processing. The method comprises: acquiring original projection data of single-energy CT scanning of a to-be-measured object; inputting the original projection data into a first target network to obtain monochrome projection sinograms at a plurality of preset energies output by the first target network; inputting the monochrome projection sinograms at the plurality of preset energies into a second target network, so that the second target network maps the monochrome projection sinograms at the plurality of preset energies into a relative electron density line integral sinogram and a stopping power ratio line integral sinogram of a projection domain and outputs the relative electron density line integral sinogram and the stopping power ratio line integral sinogram; and performing filtered back-projection reconstruction or iterative reconstruction on the relative electron density line integral sinogram and the stopping power ratio line integral sinogram respectively to obtain a relative electron density spatial distribution and a stopping power ratio spatial distribution of the to-be-measured object.

Description

Technical Field

[0001] This disclosure relates to the field of medical image processing technology, and more specifically to a method, apparatus, device, medium, and program product for determining the physical parameters of tissue in the projection domain. Background Technology

[0002] Proton therapy and heavy ion therapy are both types of particle therapy. Particle beams exhibit distinct depth-dose distribution characteristics (Bragg peak) in human tissues, with energy deposition primarily concentrated near the end of the beam's range. Therefore, particle therapy planning requires high accuracy in the tissue's relative electron density (RED) and stopping power ratio (SPR). SPR directly determines the particle range and is a key physical quantity in treatment plan dose calculation and range assessment.

[0003] Currently, clinical practice typically uses Huntsfield units (HU) from computed tomography (CT) images to establish conversion curves between HU and RED or SPR using tissue equivalent phantoms (such as Gammex). This method has the following limitations: HU is not only related to material density but is also affected by effective atomic number, X-ray energy spectrum, scanning protocol, reconstruction algorithm, beam hardening, scattering, noise, and metal artifacts; tissue equivalent phantoms can only discretize approximate real human tissue; recalibration is required under different equipment / protocols / energy spectrum conditions; and the conversion from HU to SPR may lead to range uncertainty. Summary of the Invention

[0004] This disclosure provides a method, apparatus, device, medium, and program product for determining the physical parameters of a projection domain.

[0005] According to a first aspect of this disclosure, a method for determining tissue physical parameters in a projection domain is provided, comprising: acquiring raw projection data of a single-energy CT scan of a test object; inputting the raw projection data into a first target network to obtain a plurality of monochromatic projection sine maps at preset energies output by the first target network; inputting the plurality of monochromatic projection sine maps at preset energies into a second target network, such that the second target network maps the plurality of monochromatic projection sine maps at preset energies into a relative electron density line integral sine map and a stopping power ratio line integral sine map in the projection domain and outputs the relative electron density line integral sine map and the stopping power ratio line integral sine map; performing filtered back-projection reconstruction or iterative reconstruction on the relative electron density line integral sine map and the stopping power ratio line integral sine map respectively to obtain the spatial distribution of the relative electron density and the spatial distribution of the stopping power ratio of the test object.

[0006] According to embodiments of this disclosure, the method further includes: constructing a gold standard distribution map of relative electron density and a gold standard distribution map of stopping power ratio in the image domain; forward projecting the gold standard distribution map of relative electron density and the gold standard distribution map of stopping power ratio to generate labels for a linear integral sine curve of relative electron density and a linear integral sine curve of stopping power ratio; using a virtual CT system, generating single-energy CT projection training samples based on the elemental composition and density distribution used to construct the gold standard distribution map, and generating monochromatic projection sine curve training labels corresponding to multiple preset energies; training a first initial network with the single-energy CT projection training samples and the monochromatic projection sine curve training labels to obtain a first target network; and training a second initial network with the monochromatic projection sine curve or monochromatic projection sine curve training labels output by the first target network as input, and with the labels for a linear integral sine curve of relative electron density and the linear integral sine curve of stopping power ratio as supervision, to obtain a second target network.

[0007] According to embodiments of this disclosure, the gold standard distribution map of relative electron density and the gold standard distribution map of stopping power ratio in the image domain are constructed by one of the following methods: directly using a digital model with known elemental composition and density distribution to calculate the relative electron density value and stopping power ratio on a voxel-by-voxel basis; acquiring high-energy and low-energy images from real dual-energy CT scans, performing voxel-level elemental decomposition using a material decomposition method to obtain the elemental mass fraction and density distribution of each voxel, and then calculating the relative electron density and stopping power ratio based on physical formulas.

[0008] According to embodiments of this disclosure, during the training process, the first initial network and the second initial network adopt a cascaded training strategy, including: inputting single-energy CT projection training samples into the first initial network for independent training until the first initial network converges; freezing the network parameters of the converged first initial network; inputting single-energy CT projection training samples into the frozen first initial network; and using the monochromatic projection sine curve output by the network as a fixed input feature for training the second initial network.

[0009] According to embodiments of this disclosure, when training the second initial network, the loss function includes at least: a numerical consistency loss term, used to constrain the relative electron density line integral sine and the stopping ability ratio line integral sine output by the second initial network to remain consistent with the corresponding relative electron density line integral sine label and the stopping ability ratio line integral sine label on a pixel-by-pixel basis; and a gradient consistency loss term, used to constrain the relative electron density line integral sine and the stopping ability ratio line integral sine output by the second initial network to remain consistent with the gradient changes of the corresponding relative electron density line integral sine label and the stopping ability ratio line integral sine label in the projection angle direction and the detector channel direction.

[0010] According to embodiments of this disclosure, the first target network is an encoder-decoder structure with an energy conditional modulation module. The encoder is used to extract multi-scale features from the input conventional monoenergetic CT projection sine map and output feature maps at different scales. The energy conditional modulation module is connected between the encoder and the decoder and is used to receive the encoding vector of the target energy parameter and perform channel-level affine transformation on the feature map output by the encoder according to the encoding vector to achieve feature modulation under different energies. The decoder is used to progressively upsample the modulated feature map and restore it to a monochromatic projection sine map under a specified energy.

[0011] According to embodiments of this disclosure, the second target network includes an input layer, a feature extraction layer, and an output layer. The input layer receives multiple monochromatic projection sine waves at preset energies and stacks the monochromatic projection sine waves of different energies in the channel dimension to form a multi-channel two-dimensional tensor with energy and spatial dimensions. The feature extraction layer performs convolution operations on the multi-channel two-dimensional tensor in the projection angle direction and the detector channel direction to extract attenuation features across energies. The output layer outputs sine waves of two independent channels, corresponding to the relative electron density line integral sine wave and the stopping power ratio line integral sine wave, respectively.

[0012] According to embodiments of this disclosure, after obtaining the relative electron density line integral sine map and the stopping power ratio line integral sine map output by the second target network, the method further includes: directly outputting the obtained stopping power ratio line integral sine map as a water equivalent path length map; acquiring measured proton radiography data, which includes the measured water equivalent path length measured along the same scanning geometry; registering and comparing the stopping power ratio line integral sine map with the measured proton radiography data, and determining the proton range offset based on the comparison result.

[0013] According to a second aspect of this disclosure, a device for determining the physical parameters of tissue in the projection domain is provided, comprising: an acquisition module for acquiring raw projection data of a single-energy CT scan of a test object; a first processing module for inputting the raw projection data into a first target network to obtain a plurality of monochromatic projection sine maps at preset energies output by the first target network; a second processing module for inputting the plurality of monochromatic projection sine maps at preset energies into a second target network, such that the second target network maps the plurality of monochromatic projection sine maps at preset energies into a relative electron density line integral sine map and a blocking ability ratio line integral sine map in the projection domain and outputs the relative electron density line integral sine map and the blocking ability ratio line integral sine map; and a reconstruction module for performing filtered back-projection reconstruction or iterative reconstruction on the relative electron density line integral sine map and the blocking ability ratio line integral sine map, respectively, to obtain the spatial distribution of the relative electron density and the spatial distribution of the blocking ability ratio of the test object.

[0014] According to a third aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods described above.

[0015] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided that stores a computer program or instructions thereon, which, when executed by a processor, implement the steps of the method described above.

[0016] According to a fifth aspect of this disclosure, a computer program product is also provided, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0017] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0018] Figure 1 A flowchart illustrating a method for determining the physical parameters of a projection domain according to an embodiment of the present disclosure is shown schematically.

[0019] Figure 2 A schematic diagram illustrating the principle of a method for determining the physical parameters of a projection domain according to an embodiment of the present disclosure is provided.

[0020] Figure 3 A schematic diagram illustrating the principle of the training phase according to an embodiment of the present disclosure is shown.

[0021] Figure 4 A schematic block diagram of a device for determining the physical parameters of a projection domain organization according to an embodiment of the present disclosure is shown.

[0022] Figure 5 A block diagram of an electronic device for determining the physical parameters of a projection domain organization according to an embodiment of the present disclosure is shown schematically. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments and accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0024] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of a feature, step, operation, and / or component, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0025] In this disclosure, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances.

[0026] In the description of this disclosure, it should be understood that the terms "longitudinal", "length", "circumferential", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this disclosure and simplifying the description, and do not indicate or imply that the subsystem or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this disclosure.

[0027] Throughout the accompanying drawings, identical elements are represented by the same or similar reference numerals. Conventional structures or configurations have been omitted where they may cause confusion in understanding this disclosure. Furthermore, the shapes, dimensions, and positional relationships of the components in the drawings do not reflect actual size, scale, or actual positional relationships. Additionally, any reference symbols enclosed in parentheses should not be construed as limiting.

[0028] Similarly, to simplify this disclosure and aid in understanding one or more of the various aspects of the disclosure, in the above description of exemplary embodiments of the present disclosure, various features of the present disclosure are sometimes grouped together in a single embodiment, figure, or description thereof. The use of terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refers to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the present disclosure. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0029] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this disclosure, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0030] In the technical solution disclosed herein, the collection, storage, use, processing, transmission, provision, disclosure, and application of user personal information all comply with relevant laws and regulations, necessary confidentiality measures have been taken, and they do not violate public order and good morals. In the technical solution disclosed herein, user authorization or consent has been obtained before acquiring or collecting user personal information.

[0031] This disclosure provides a method for determining the physical parameters of a projection domain. Before introducing the technical solution provided by this disclosure, the relevant technologies involved in this disclosure will be explained.

[0032] Currently commonly used methods include:

[0033] 1. Dual-energy CT (DECT): Provides attenuation information under two energy conditions, and calculates SPR using physical formulas after material / element decomposition. However, DECT is highly dependent on equipment and requires actual clinical acquisition of dual-energy data.

[0034] 2. Image Domain Deep Learning Prediction Methods: For example, end-to-end prediction of mass density and RSP / SPR from single-energy CT images, or synthesis of dual-energy CT from single-energy CT and subsequent SPR calculation. These methods all perform RED / SPR prediction in the reconstructed image domain, which has the following problems: the network needs to implicitly cope with beam hardening, noise, and ring / strip artifacts introduced during the reconstruction process; the prediction object is a voxel-by-voxel RED / SPR map, which cannot directly give the water equivalent path length (WEPL) for each ray, and therefore cannot be directly used for in vivo range verification in proton radiography; the network needs to implicitly learn from attenuation to tissue physical quantities in the data, while simultaneously tolerating the full complexity of the reconstruction inverse problem.

[0035] 3. Projection Domain Monochromatic Method: This method maps multicolor projections to monochromatic projections of a specified energy using a neural network, for beam hardening correction and virtual monochromatic imaging. However, this type of method stops at monochromatic projection / monochromatic images and does not directly predict RED and SPR in the projection domain.

[0036] Based on this, this disclosure proposes a method for determining the physical parameters of the projection domain, which is used to solve at least the following problems:

[0037] (1) How to enable the network to directly predict tissue physical parameters using CT projection data instead of predicting in the reconstructed image domain? Thus, the network mainly learns the mapping between attenuated projection and the line integral of tissue physical parameters, while the inverse Radon reconstruction process from projection to image is completed by a deterministic reconstruction algorithm.

[0038] (2) How to obtain an output with direct physical meaning that can be used for in vivo range verification, namely the water equivalent path length (WEPL) of each ray / proton ray radiograph result;

[0039] (3) How to get rid of the dependence on dual-energy CT acquisition in the practical application stage and predict RED and SPR using only conventional single-energy CT projection;

[0040] (4) How to reduce the dependence on Gammex empirical calibration curves and make the supervision labels based on elemental composition and physical formulas.

[0041] Figure 1 A flowchart illustrating a method for determining the physical parameters of a projection domain according to an embodiment of the present disclosure is shown.

[0042] like Figure 1 As shown, the method for determining the physical parameters of the projection domain in this embodiment includes operations S110 to S140.

[0043] In operation S110, the raw projection data of the single-energy CT scan of the object under test is acquired.

[0044] In some embodiments, a conventional single-energy CT scan is performed on the object under test to obtain raw projection data. The raw projection data is a sine wave format data formed by logarithmic transformation of the X-ray attenuation signal received by the detector, and scanning geometric parameters are recorded, including the distance from the source to the isocenter, the distance from the isocenter to the detector, the number of detector channels, and the number of projection angles.

[0045] In operation S120, the original projection data is input into the first target network to obtain a monochromatic projection sine curve with multiple preset energies output by the first target network.

[0046] The acquired raw single-energy CT projection data is input into the first target network. This network has been pre-trained to map conventional single-energy CT projections into monochromatic projection sine waves at multiple preset energies. The input to the first target network is a single-channel conventional single-energy CT projection sine wave, and the output is a monochromatic projection sine wave at multiple energy points. The preset energies can be, for example, 70 keV, 120 keV, and 200 keV.

[0047] For example, the first target network can adopt an encoder-decoder structure. The encoder performs multi-scale feature extraction on the input sine wave, capturing global and local structural information in the projection data layer by layer. The decoder gradually restores the extracted features to the spatial resolution of the sine wave and generates a monochromatic projection at a specified energy at the output. The first target network completes all operations in the projection domain, and both the input and output are in sine wave format.

[0048] In operation S130, a number of monochromatic projection sine diagrams at preset energies are input into the second target network, so that the second target network maps the number of monochromatic projection sine diagrams at preset energies into a relative electron density line integral sine diagram and a blocking ability ratio line integral sine diagram in the projection domain, and outputs the relative electron density line integral sine diagram and the blocking ability ratio line integral sine diagram.

[0049] In some embodiments, the monochromatic projected sine waves at multiple preset energies output by the first target network are input into the second target network. The second target network has been pre-trained and is used to map the multi-energy monochromatic projected sine waves to RED line integral sine waves and SPR line integral sine waves in the projection domain. The processing of the second target network is entirely performed in the projection domain; its operations are purely projection domain operations and do not involve image reconstruction or feature extraction.

[0050] The RED line integral sine curve is defined as the line integral of the relative electron density along the scanning ray. The SPR line integral sine curve is defined as the line integral of the stopping power ratio along the scanning ray. The physical meaning of the SPR line integral sine curve is the water equivalent path length (WEPL) of each ray path under that scanning geometry.

[0051] For example, monochromatic projection sine waves of different energies can be stacked along the channel dimension to form multi-channel input data. The second target network extracts features from the multi-channel data in the projection angle direction and the detector channel direction, learning the mapping relationship between multi-energy attenuation information and RED / SPR line integrals. Finally, it outputs two independent channel sine waves, corresponding to the RED line integral sine wave and the SPR line integral sine wave, respectively. The second target network is a pure projection domain operation network, and its entire forward propagation process is completed within the sine wave domain, without including the operation of reconstructing the sine wave into image domain data before feature extraction.

[0052] In operation S140, filtered back-projection reconstruction or iterative reconstruction is performed on the relative electron density line integral sine diagram and the stopping power ratio line integral sine diagram, respectively, to obtain the spatial distribution of the relative electron density and the spatial distribution of the stopping power ratio of the object under test.

[0053] In some embodiments, filtered backprojection reconstruction is performed on the RED line integral sine map and SPR line integral sine map output by the second target network, respectively. Filtered backprojection first filters the sine map (e.g., using a ramp filter or band-limited filter), then backprojects the filtered projection data along the scan path into the image space to obtain a spatial distribution map. Since the RED and SPR sine maps are defined as the line integrals of the corresponding physical quantities and satisfy the definition of the Radon transform, standard filtered backprojection reconstruction algorithms can be used to obtain the three-dimensional or two-dimensional spatial distribution.

[0054] This embodiment of the invention moves the prediction step of RED / SPR to the projection domain. The second target network directly takes a sine wave as input and outputs a sine wave, and all calculations are performed within the projection domain. The estimation results of RED / SPR are not affected by image reconstruction artifacts, thus improving the accuracy and precision of physical parameter determination. The method provided by this embodiment only requires the raw projection data from a conventional single-energy CT scan to obtain the distribution maps of RED and SPR tissue physical parameters, without the need for dual-energy CT scans or special acquisition protocols. This lowers the threshold for clinical promotion and reduces equipment investment costs, making it highly universal.

[0055] According to embodiments of this disclosure, the first target network is an encoder-decoder structure with an energy conditional modulation module. The encoder is used to extract multi-scale features from the input conventional monoenergetic CT projection sine wave and output feature maps at different scales. The energy conditional modulation module is connected between the encoder and the decoder, and is used to receive the encoded vector of the target energy parameter, and perform channel-level affine transformation on the feature map output by the encoder according to the encoded vector to achieve feature modulation at different energies. The decoder is used to progressively upsample the modulated feature map and restore it to a monochromatic projection sine wave at a specified energy.

[0056] In some embodiments, a single-energy CT projection sine wave (e.g., a 120 kVp multicolor projection) is input into the encoder. The encoder consists of multiple stacked convolutional and downsampling layers, each performing a convolution operation on the input feature map followed by downsampling. As the network depth increases, the spatial resolution of the feature map gradually decreases (from the original sine wave size), while the number of channels gradually increases (from the initial number of channels to hundreds of channels in deeper layers). Shallow feature maps retain local details and high-frequency information (such as noise distribution and local attenuation variations) of the original projection, while deep feature maps capture global contextual information and low-frequency structures (such as the overall anatomical contours and large-scale trends in projection values). Finally, the encoder outputs multiple feature maps at different scales, forming a pyramid-structured feature representation from fine to coarse.

[0057] The deep feature map output from the encoder is input into the energy conditional modulation module. The target energy parameter (such as 70 keV, 120 keV, or 200 keV, or other arbitrarily specified energy value) is encoded into a vector. For example, discrete energy values ​​can be mapped to continuous vectors through an embedding layer, or continuous energy values ​​can be encoded into vectors through sinusoidal positional encoding, or the energy values ​​can be normalized and then input into a fully connected layer to be mapped into vectors. For this energy encoding vector, scaling parameters and bias parameters are generated through a fully connected layer. The dimensions of these two parameter vectors are the same as the number of channels in the feature map output by the encoder. The modulation module performs a channel-level affine transformation on the feature map output by the encoder, multiplying each channel of the feature map by the corresponding scaling parameter and adding the corresponding bias parameter. Both the scaling parameter and the bias parameter are generated from the energy encoding vector. The affine transformation preserves the semantic information of the original feature map and injects the target energy condition, effectively coupling the two types of information.

[0058] The modulated feature map is input into the decoder. The decoder consists of multiple stacked upsampling and convolutional layers, forming a symmetrical structure with the encoder. The upsampling layers gradually restore the spatial resolution of the feature map through transposed convolution or interpolation operations, and simultaneously concatenate or add the feature map output from the corresponding layer of the encoder through skip connections to supplement spatial detail information. After successive upsampling and feature fusion, the spatial resolution of the final feature map is restored to the same size as the input sine map. At the output, a convolutional layer maps the feature map to a monochromatic projected sine map at a specified energy; this sine map is the final output.

[0059] The architecture of the first target network is not limited to the encoder-decoder structure mentioned above; it can also adopt other architectures such as residual network, U-Net, residual U-Net, Transformer projection domain network, or conditional diffusion model.

[0060] This embodiment uses an energy condition modulation module to take the target energy as an input parameter, enabling dynamic generation of modulation parameters from the energy encoding vector, so that the feature response of the first target network changes with the input energy condition. The user can input any specified energy value so that the first target network outputs a monochromatic projection of the corresponding energy. This effectively realizes the interpolation capability of a single network for continuous energy values. The modulation module uses channel-level affine transformation (scaling and biasing) instead of directly concatenating the energy vector onto the feature map, which effectively improves computational efficiency. The encoder-decoder structure directly transmits shallow detail information to the decoder through skip connections, effectively avoiding information loss during downsampling, ensuring local details (such as attenuation changes at tissue boundaries) while maintaining overall numerical accuracy and structural consistency. All network operations are performed in the sine wave domain; the input is the original projected sine wave, and the output is a monochromatic projected sine wave, without involving reconstructing the sine wave into a CT image for further processing. The first target network learns the mapping relationship of attenuated signals in the projection domain, rather than image features contaminated by reconstruction artifacts (such as beam hardening artifacts and ring artifacts). This avoids interference from image domain operations on network learning from the source, thus improving the fidelity of predicting monochromatic projections.

[0061] According to embodiments of this disclosure, the second target network includes an input layer, a feature extraction layer, and an output layer. The input layer receives multiple monochromatic projected sine waves at preset energies and stacks these energies along the channel dimension to form a multi-channel two-dimensional tensor with energy and spatial dimensions. The feature extraction layer performs convolution operations on this multi-channel two-dimensional tensor along the projection angle direction and the detector channel direction to extract attenuation features across energy levels. The output layer outputs two independent channel sine waves, corresponding to the RED line integral sine wave and the SPR line integral sine wave, respectively.

[0062] In some embodiments, monochromatic projected sine waves at multiple preset energies (e.g., 70 keV, 120 keV, and 200 keV) output by the first target network are input into the input layer of the second target network. The input layer stacks the monochromatic projected sine waves of different energies along the channel dimension to form a multi-channel two-dimensional tensor. The channel dimension corresponds to the energy (e.g., the three channels correspond to 70 keV, 120 keV, and 200 keV respectively), and the two spatial dimensions correspond to the projection angle direction and the detector channel direction, respectively. This tensor contains information from both the energy and spatial dimensions; that is, the data at each spatial location is a feature vector composed of multiple energy attenuation values.

[0063] A multi-channel two-dimensional tensor is input into the feature extraction layer. The feature extraction layer consists of multiple stacked convolutional layers, each of which uses a two-dimensional convolutional kernel that slides along the projection angle direction and the detector channel direction to perform convolution operations.

[0064] Convolution operations act simultaneously on two spatial dimensions: the projection angle direction and the detector channel direction. This allows them to capture spatial correlation features between adjacent projection angles and adjacent detector channels. Since the input is a multi-channel tensor, and the depth of the convolution kernel is the same as the number of input channels, a single convolution operation can fuse attenuation information at different energies, extracting attenuation features across energy levels. As the network depth increases, the feature extraction layers gradually learn features from simple to complex. Shallow layers extract local attenuation change patterns, while deeper layers extract global structural distribution patterns.

[0065] In this process, all operations of the feature extraction layer are performed within the sinusoidal map domain. The two spatial dimensions always correspond to the projection angle direction and the detector channel direction, and do not use spatial cross-sectional coordinates as the operation dimensions. During convolution, padding methods can be set as needed to ensure that the feature map maintains the same resolution as the input sinusoidal map in the spatial dimensions, or the spatial size can be gradually compressed through downsampling to expand the receptive field.

[0066] The final feature map output from the feature extraction layer is input into the output layer. The output layer maps the feature map into a sine map with two independent channels through a convolution layer. These two channels correspond to the RED line integral sine map and the SPR line integral sine map, respectively. The RED line integral sine map is defined as the line integral of the relative electron density along the ray; the SPR line integral sine map is defined as the line integral of the stopping power ratio along the ray, and its physical meaning is the water-equivalent path length (WEPL) on each ray path.

[0067] The second target network is a pure projection domain operation network, which uses a sine wave as input and output, and all feature extraction is performed in the projection angle direction and the detector channel direction.

[0068] According to embodiments of this disclosure, the second target network is a pure projection domain operation network, and all operations are performed in the sinusoidal domain. The input is a monochrome projection sinusoid, and the output is a RED / SPR line integral sinusoid. The entire forward propagation process does not involve reconstructing the sinusoid into image domain data and then performing feature extraction, which can effectively avoid the contamination of RED / SPR estimation by image reconstruction artifacts.

[0069] Figure 2 A schematic diagram illustrates the principle of a method for determining the physical parameters of a projection domain according to an embodiment of the present disclosure.

[0070] like Figure 2As shown, after obtaining the relative electron density line integral sine map and the stopping power ratio line integral sine map output by the second target network, the method further includes: directly outputting the obtained SPR line integral sine map as a water equivalent path length map; acquiring measured proton radiography data, which includes the measured water equivalent path length measured along the same scanning geometry; registering and comparing the SPR line integral sine map with the measured proton radiography data, and determining the proton range offset based on the comparison result.

[0071] In some embodiments, after the second target network outputs the RED line integral sine map and the SPR line integral sine map, if only range verification is required, the subsequent filtering back projection reconstruction step is skipped, and the SPR line integral sine map is directly output as the water equivalent path length (WEPL) map.

[0072] The physical meaning of the SPR line integral sine plot is ∫SPR·dl for each ray path, where SPR is the stopping power ratio and dl is the integral infinitesimal element along the ray path. This value is physically equal to the total water equivalent thickness of the tissue through which the ray passes, in millimeters of water. Since no reconstruction is performed, this WEPL plot retains the original sine plot's data format, i.e., rows of the two-dimensional matrix correspond to projection angles, and columns correspond to detector channels.

[0073] During treatment, a proton radiography system was used to collect data on the same patient and under the same scanning geometry. The proton radiography system directly obtains the measured water equivalent path length value for each beam path by measuring the residual energy or positional shift of high-energy protons after they pass through the patient. This measured data is geometrically consistent with the predicted WEPL map (same number of projection angles and detector channels), ensuring that both sets of data have the same spatial sampling grid.

[0074] Spatially register the SPR line integral sine plot (predicted WEPL) with the acquired measured proton beam radiograph data (measured WEPL) to ensure that the two sets of data correspond one-to-one in projection angle and detector channel.

[0075] After registration, the difference between the two sets of data at corresponding pixels is calculated: Difference = Predicted WEPL - Measured WEPL. The difference map reflects the range deviation along each ray path: a positive value indicates that the predicted equivalent water thickness is greater than the measured value, meaning the actual range may be shorter than the planned range; a negative value indicates that the predicted equivalent water thickness is less than the measured value, meaning the actual range may be longer than the planned range.

[0076] Statistical analysis of the difference map was performed to calculate the overall average offset, maximum offset, and spatial distribution of the offset. If the WEPL deviation on a certain ray path exceeds the clinically tolerable threshold (e.g., ±3 mm water equivalent thickness), the spatial location corresponding to that ray is marked as having a significant range offset. Combined with the spatial distribution of the difference map, the specific anatomical region where the range offset occurred can be located (e.g., patient positioning errors causing discrepancies between the actual and planned puncture paths in some areas, or changes in anatomical structures during treatment leading to alterations in the water equivalent thickness), thus providing a basis for clinical decision-making.

[0077] In the method of this embodiment, the SPR line integral sine graph output by the second target network is directly used as the WEPL graph. This eliminates the need for FBP reconstruction and allows for comparison with measured data, saving the time and computational resources required for reconstruction and resulting in faster inference speed. This is particularly crucial in scenarios with high real-time requirements during treatment. Directly outputting the WEPL in the projection domain, which is in the same data domain as the measured proton beam radiography data, avoids the additional errors introduced by the reconstruction-reintegration stage during the comparison process, and the comparison results more accurately reflect the true range deviation.

[0078] The method for determining the physical parameters of the projection domain provided in this embodiment further includes: training a first target model and a second target model.

[0079] Figure 3 A schematic diagram illustrating the training phase according to an embodiment of the present disclosure is shown.

[0080] According to embodiments of this disclosure, such as Figure 3 As shown, the training process can be as follows: constructing the RED gold standard distribution map and SPR gold standard distribution map in the image domain; performing forward projection on the RED gold standard distribution map and SPR gold standard distribution map to generate RED line integral sine map labels and SPR line integral sine map labels; using a virtual CT system, generating single-energy CT projection training samples based on the element composition and density distribution used to construct the gold standard distribution map, and generating monochrome projection sine map training labels corresponding to multiple preset energies; training a first initial network with the single-energy CT projection training samples and monochrome projection sine map training labels to obtain a first target network; using the monochrome projection sine map output by the first target network or the monochrome projection sine map training labels as input, and using the RED line integral sine map labels and SPR line integral sine map labels as supervision, training a second initial network to obtain a second target network.

[0081] The gold standard is the true value used as a supervisory signal during training, i.e., the target label that the network learns. The predicted value output by the network needs to be compared with the gold standard, and the network parameters are updated by minimizing the difference between the two.

[0082] In some embodiments, the RED gold standard distribution map and SPR gold standard distribution map in the image domain can be constructed in any of the following ways: directly using a digital model with known elemental composition and density distribution to calculate the RED and SPR values ​​on a voxel-by-voxel basis; acquiring high-energy and low-energy images from real dual-energy CT scans, performing voxel-level elemental decomposition using a material decomposition method to obtain the elemental mass fraction and density distribution of each voxel, and then calculating the RED and SPR values ​​based on physical formulas.

[0083] For example, a digital model with known elemental composition and density distribution can be used to calculate the RED and SPR values ​​on a voxel-by-voxel basis. This includes using digital human body models with known elemental composition and density distribution (such as the XCAT model, ICRP reference human body model, etc.). These models have built-in information on the anatomical structure, elemental mass fractions (hydrogen, carbon, nitrogen, oxygen, phosphorus, calcium, etc.), and mass density of various human tissues and organs. For each voxel in the model, the electron density of that voxel is calculated based on its elemental mass fraction and density. The electron density is equal to the weighted sum of the number of electrons of each element and their mass fraction multiplied by the mass density. The calculated electron density is normalized relative to the electron density of water to obtain the RED value of that voxel. Combining the material average excitation energy I and the incident particle energy correlation term of that voxel, the mass fraction and density of each element are substituted into the Bethe formula to calculate the stopping power ratio of that voxel relative to water, thus obtaining the SPR value of that voxel. By traversing the entire digital human body model voxel by voxel, a three-dimensional gold standard RED distribution map and a gold standard SPR distribution map are finally obtained.

[0084] For example, high-energy and low-energy images from real dual-energy CT scans are acquired. Voxel-level elemental decomposition is performed using a material decomposition method to obtain the elemental mass fraction and density distribution of each voxel. Then, the RED and SPR values ​​are calculated based on physical formulas. This process includes: acquiring real dual-energy CT scan data of the object to be modeled, including low-energy and high-energy images. CT images at two different energies provide information on the differential attenuation of tissue to X-rays of different energies, which is the basis for distinguishing different elemental compositions. The low-energy and high-energy images are input into a material decomposition network (such as a fully connected network, convolutional network, or physically constrained network), which has learned the mapping from dual-energy attenuation information to elemental mass fractions. The network performs elemental decomposition on each pixel or voxel, outputting the mass fraction distribution of the six main elements (hydrogen, carbon, nitrogen, oxygen, phosphorus, and calcium) and the tissue's mass density distribution.

[0085] After obtaining the elemental mass fraction and density of each voxel, the RED and SPR values ​​are calculated based on physical formulas. The electron density is calculated from the elemental mass fraction and density and normalized to obtain the RED; the elemental composition, density, and average excitation energy are substituted into the Bethe formula to calculate the SPR. After voxel-by-voxel calculations, three-dimensional gold standard distribution maps of RED and SPR are obtained. It should be noted that in this method, dual-energy CT scanning and elemental decomposition are only used for label generation during the training phase and are not necessary inputs for the application phase.

[0086] In some embodiments, the RED map and SPR map obtained in step 1 can be forward-projected (i.e., Radon transform) under the same scanning geometry conditions as the actual CT scan (including source-to-isocenter distance, isocenter-to-detector distance, number of detector channels, number of projection angles, etc.).

[0087] For each projection angle and each detector channel, calculate the line integral value of the RED image along the ray path to obtain the RED line integral sine map label; calculate the line integral value of the SPR image along the ray path to obtain the SPR line integral sine map label (i.e., the water equivalent path length WEPL of the ray).

[0088] These two sets of projection domain labels correspond strictly to the output of the subsequent second network in terms of scanning geometry, and serve as the training targets for supervised learning of the second network.

[0089] In some embodiments, the acquired elemental composition and density distribution are input into a virtual CT system. The virtual CT system calculates the linear attenuation coefficient distribution based on the mass attenuation coefficient of each element at different energies and simulates the complete physical process of a CT scan. Conventional single-energy CT projection training samples, such as 120 kVp multicolor projection, are generated, and physical effects such as Poisson photon statistical noise, scattering, and beam hardening can be incorporated, serving as input training samples for the first initial network. Additionally, monochromatic projection sine curve training labels corresponding to multiple preset energies (e.g., 70 keV, 120 keV, 200 keV) are generated as supervision targets for the first initial network. All projection data are generated under the same virtual scan geometry, ensuring complete geometric correspondence between the input samples and the supervision labels.

[0090] In some embodiments, the generated conventional single-energy CT projection training samples are used as input, and the corresponding monochromatic projection sine curves at multiple preset energies are used as supervision to train a first initial network. The first initial network learns the mapping relationship from noisy, artifact-laden conventional single-energy CT projections to ideal monochromatic projections. By minimizing the difference between the network output and the labels, the network gradually acquires the ability to remove physical interferences such as beam hardening and noise and output ideal monochromatic projections at a specified energy. After training, the first target network is obtained.

[0091] In some embodiments, the monochrome projected sine curve output by the first target model can be used as the input feature of the second initial network to train the second initial network, or the monochrome projected sine curve label can be directly used as the input feature of the second initial network to train the second initial network.

[0092] For example, conventional monoenergetic CT projection training samples are input into a pre-trained first target network. The monochromatic projection sine curve output by the first target network is used as the input feature of the second network. The labels of the RED line integral sine curve and the SPR line integral sine curve are used as supervision to train the second initial network so that the second initial network can adapt to the feature distribution of the actual output of the first target network.

[0093] For example, the labels trained using monochrome projected sine waves are used as input features for the second network, while the labels of RED line integral sine waves and SPR line integral sine waves are used as supervision to train the second initial network, resulting in the second target network. This approach can achieve a more desirable upper limit performance when there is sufficient training data.

[0094] In this embodiment, the RED and SPR labels are directly calculated from the elemental composition and density of the digital human body model or the elemental decomposition results of real dual-energy CT, combined with the Bethe physics formula. This calculation does not rely on calibration using the empirical HU-electron density curve of the Gammex phantom, effectively avoiding systematic errors caused by differences between the tissue-equivalent phantom and actual human tissue. It can be extended to any elemental composition and any tissue type. By forward-projecting the RED / SPR gold standard image to generate projection domain labels, the training labels and the output of the second target network strictly correspond under the same scanning geometry. The second target network directly uses the line integral sine curve as the supervision target in the sine wave domain, avoiding errors caused by additional registration or resampling of the image domain labels.

[0095] According to embodiments of this disclosure, during the training process, the first initial network and the second initial network adopt a cascaded training strategy, including: inputting single-energy CT projection training samples into the first initial network for independent training until the first initial network converges; freezing the network parameters of the converged first initial network; inputting single-energy CT projection training samples into the frozen first initial network; and using the monochromatic projection sine curve output by the network as a fixed input feature for training the second initial network.

[0096] In some embodiments, a first initial network is independently trained until convergence to obtain a first target network. During training, the network updates its parameters by minimizing the difference between its output (the predicted monochromatic projection sine) and the supervision labels. An optimizer (such as Adam) gradually adjusts the network weights, and the loss function value continuously decreases. When the loss value stabilizes and no longer decreases significantly, the network is considered to have converged. At this point, the first initial network has mastered the mapping ability from noisy multicolor projection to ideal monochromatic projection, thus obtaining the first target network.

[0097] Once the initial network converges, all its network parameters (weights and biases) are fixed so that they are no longer updated in subsequent operations. "Freezing" means that during subsequent training, these parameters remain in their pre-trained optimal state, and gradient information is not backpropagated into the first network.

[0098] The first initial network, whose parameters have been frozen, takes the same monoenergetic CT projection training samples as input. This network then outputs a corresponding monochromatic projection sine wave during forward propagation. Since the parameters of the first network are fixed, its output is deterministic and invariant for the same input.

[0099] These output monochromatic projected sine waves are used as input features of the second initial network (remaining constant). Simultaneously, the RED line integral sine wave labels obtained by forward projection from the RED / SPR gold standard image and the SPR line integral sine wave labels are used as supervision to train the second initial network. During training, only the parameters of the second initial network are updated, ensuring that its output RED / SPR line integral sine waves continuously approximate the supervision labels.

[0100] This embodiment employs a cascaded training strategy. The second network receives the "monochrome projection actually output by the first target network" as input features, rather than ideal labels. This ensures that the input distribution during the training of the second target network is completely consistent with that during application, resulting in superior overall performance after cascading. Cascaded training decouples the training of the two networks. The first network is trained independently until stable, and then its parameters are frozen to train the second network independently. Each network undertakes only a specific subtask (the first network is responsible for artifact removal and generating the monochrome projection, and the second network is responsible for predicting the RED / SPR line integral), making the training process stable and controllable. This training process avoids adaptive drift caused by the first network trying to adapt to the second network, preventing the degradation of the mapping capabilities originally learned by the first network.

[0101] During the training of the second target network, in addition to using the numerical losses of the RED sine curve and the SPR sine curve as the main supervision, a projection domain consistency constraint can also be applied to the sine curve output by the second network.

[0102] According to embodiments of this disclosure, when training the second initial network, the loss function includes at least: a numerical consistency loss term, used to constrain the RED line integral sine and SPR line integral sine output by the second initial network to maintain consistency with the corresponding RED line integral sine label and SPR line integral sine label on a pixel-by-pixel basis; and a gradient consistency loss term, used to constrain the gradient changes of the RED line integral sine and SPR line integral sine output by the second initial network to maintain consistency with the corresponding RED line integral sine label and SPR line integral sine label on the gradient changes in the projection angle direction and the detector channel direction.

[0103] In some embodiments, when calculating the numerical consistency loss, the RED line integral sine map output by the second initial network is compared pixel-by-pixel with the RED line integral sine map label (obtained by forward projection of the RED gold standard map), and the difference between the two (such as mean absolute error or mean squared error) is calculated. Additionally, the output SPR line integral sine map is compared pixel-by-pixel with the SPR line integral sine map label, and the difference is calculated. The differences between the RED and SPR parts are summed or weighted summed to obtain the numerical consistency loss term. This loss term is used to constrain the network's output value at each pixel to be as close as possible to the label value, ensuring the overall numerical accuracy of the predicted sine map.

[0104] For the RED and SPR line integral sine curves output by the second initial network, their gradients in the projection direction and detector channel direction are calculated, respectively. For example, the gradient in the projection angle direction can be calculated along the projection angle index direction, representing the rate of change of the line integral value between adjacent angles; a large gradient value indicates a drastic signal change near that angle. The gradient in the detector channel direction can be calculated along the detector channel index direction, representing the rate of change of the line integral value between adjacent channels; a large gradient value indicates a structural abrupt change near that spatial location. Similarly, the gradients in the projection angle direction and detector channel direction are also calculated for the RED and SPR line integral sine curve labels, respectively.

[0105] After the calculation is complete, the gradient of the predicted sine wave is compared pixel by pixel with the gradient of the label sine wave, and the difference between the two is calculated. The gradient differences of RED and SPR in the two directions are combined to obtain the gradient consistency loss term.

[0106] The numerical consistency loss term and the gradient consistency loss term are weighted and summed to form the overall loss function for training the second initial network. The parameters of the second initial network are updated through backpropagation, causing the overall loss to continuously decrease.

[0107] This embodiment uses both numerical consistency loss and gradient consistency loss as the loss function, which ensures both the accuracy and spatial structure rationality of the predicted values, resulting in accurate numerical values ​​and a clear structure in the output data of the second target network. The projection domain consistency constraint acts directly on the line integral data itself, rather than on the reconstructed image domain RED / SPR distribution. This effectively avoids the indirectness of image constraints, accurately suppresses anomalous projection points, improves constraint efficiency, and prevents the difficulty in accurately locating and correcting local anomalies in the sine curve due to image domain constraints.

[0108] Figure 4 A schematic block diagram of a device for determining the physical parameters of a projection domain according to an embodiment of the present disclosure is shown.

[0109] like Figure 4 As shown, the projection domain tissue physical parameter determination device 200 of this embodiment includes an acquisition module 210, a first processing module 220, a second processing module 230, and a reconstruction module 240.

[0110] The acquisition module 210 is used to acquire the raw projection data of the single-energy CT scan of the object under test. In one embodiment, the acquisition module 210 can be used to perform the operation S110 described above, which will not be repeated here.

[0111] The first processing module 220 is used to input the original projection data into the first target network to obtain a number of monochromatic projection sine waves at preset energies output by the first target network. In one embodiment, the fitting module 220 can be used to perform the operation S120 described above, which will not be repeated here.

[0112] The second processing module 230 is used to input monochromatic projection sine waves at multiple preset energies into the second target network, so that the second target network maps the monochromatic projection sine waves at multiple preset energies into a relative electron density line integral sine wave and a blocking ability ratio line integral sine wave in the projection domain, and outputs the relative electron density line integral sine wave and the blocking ability ratio line integral sine wave. In one embodiment, the second processing module 230 can be used to perform the operation S130 described above, which will not be repeated here.

[0113] The reconstruction module 240 is used to perform filtered back-projection reconstruction or iterative reconstruction on the relative electron density line integral sine diagram and the stopping power ratio line integral sine diagram, respectively, to obtain the spatial distribution of the relative electron density and the spatial distribution of the stopping power ratio of the object under test. In one embodiment, the reconstruction module 240 can be used to perform the operation S140 described above, which will not be repeated here.

[0114] According to embodiments of this disclosure, any plurality of modules among the acquisition module 210, the first processing module 220, the second processing module 230, and the reconstruction module 240 may be combined into one module, or any one of these modules may be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of other modules and implemented in one module. According to embodiments of this disclosure, at least one of the acquisition module 210, the first processing module 220, the second processing module 230, and the reconstruction module 240 may be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in any one of the three implementation methods of software, hardware, and firmware, or in a suitable combination of any of these. Alternatively, at least one of the acquisition module 210, the first processing module 220, the second processing module 230, and the reconstruction module 240 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0115] Figure 5 A block diagram of an electronic device for determining the physical parameters of a projection domain organization according to an embodiment of the present disclosure is shown schematically.

[0116] like Figure 5 As shown, an electronic device 300 according to an embodiment of this disclosure includes a processor 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage portion 308 into a random access memory (RAM) 303. The processor 301 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 301 may also include onboard memory for caching purposes. The processor 301 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this disclosure.

[0117] RAM 303 stores various programs and data required for the operation of electronic device 300. Processor 301, ROM 302, and RAM 303 are interconnected via bus 304. Processor 301 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 302 and / or RAM 303. It should be noted that the program may also be stored in one or more memories other than ROM 302 and RAM 303. Processor 301 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in one or more memories.

[0118] According to embodiments of this disclosure, the electronic device 300 may further include an input / output (I / O) interface 305, which is also connected to a bus 304. The electronic device 300 may also include one or more of the following components connected to the input / output (I / O) interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card, modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the input / output (I / O) interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 310 as needed so that computer programs read from it can be installed into the storage section 308 as needed.

[0119] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0120] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0121] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this disclosure, those skilled in the art can make various substitutions and modifications, all of which should be included within the protection scope of this disclosure.

Claims

1. A method of determining a physical parameter of a tissue in a projection domain, characterized in that, include: Obtain the raw projection data of the single-energy CT scan of the object under test; The original projection data is input into the first target network to obtain a number of monochromatic projection sine curves under preset energies output by the first target network. The monochromatic projection sine diagrams at multiple preset energies are input into the second target network, so that the second target network maps the monochromatic projection sine diagrams at multiple preset energies into a relative electron density line integral sine diagram and a blocking ability ratio line integral sine diagram in the projection domain, and outputs the relative electron density line integral sine diagram and the blocking ability ratio line integral sine diagram. The relative electron density line integral sine map and the stopping power ratio line integral sine map are subjected to filtered back projection reconstruction or iterative reconstruction, respectively, to obtain the spatial distribution of the relative electron density and the spatial distribution of the stopping power ratio of the object under test.

2. The method of claim 1, wherein, Also includes: Construct gold standard distribution maps of relative electron density and gold standard distribution maps of stopping power ratio in the image domain; The gold standard distribution map of relative electron density and the gold standard distribution map of stopping power ratio are forward-projected to generate labels for the line integral sine of relative electron density and the line integral sine of stopping power ratio. Using a virtual CT system, based on the elemental composition and density distribution used to construct the gold standard distribution map, monoenergetic CT projection training samples are generated, and monochromatic projection sine wave training labels corresponding to multiple preset energies are generated. The first initial network is trained using the conventional single-energy CT projection training samples and the monochrome projection sine curve training labels to obtain the first target network. Using the monochromatic projected sine curve output by the first target network or the training label of the monochromatic projected sine curve as input, and the label of the relative electron density line integral sine curve and the label of the stopping power ratio line integral sine curve as supervision, the second initial network is trained to obtain the second target network.

3. The method according to claim 2, wherein the construction of the gold standard distribution map of relative electron density and the gold standard distribution map of stopping power ratio in the image domain is carried out in one of the following ways: The relative electron density and stopping power ratio are calculated on a voxel-by-voxel basis using a digital model with known elemental composition and density distribution. We acquire high-energy and low-energy images from real dual-energy CT scans, perform voxel-level elemental decomposition using material decomposition methods, obtain the elemental mass fraction and density distribution of each voxel, and then calculate the relative electron density and stopping power ratio based on physical formulas.

4. The method of claim 2, wherein, During training, the first initial network and the second initial network employ a cascaded training strategy, including: The single-energy CT projection training samples are input into the first initial network for independent training until the first initial network converges. The network parameters of the first initial network after convergence are frozen. The monoenergetic CT projection training samples are input into the frozen first initial network. The monochromatic projection sine curve output by the network is used as a fixed input feature to train the second initial network.

5. The method of claim 4, wherein, When training the second initial network, the loss function includes at least: The numerical consistency loss term is used to constrain the relative electron density line integral sine and the stopping power ratio line integral sine output by the second initial network to maintain consistency with the corresponding relative electron density line integral sine and stopping power ratio line integral sine labels on a pixel-by-pixel basis. The gradient consistency loss term is used to constrain the relative electron density line integral sine and the stopping ability ratio line integral sine output by the second initial network to maintain consistency with the gradient changes of the corresponding relative electron density line integral sine label and the stopping ability ratio line integral sine label in the projection angle direction and the detector channel direction, respectively.

6. The method according to claim 1, wherein the first target network is an encoder-decoder structure with an energy conditional modulation module. The encoder is used to extract multi-scale features from the input conventional monoenergetic CT projection sinusoidal image and output feature maps at different scales. The energy condition modulation module is connected between the encoder and the decoder. It is used to receive the encoding vector of the target energy parameter and perform channel-level affine transformation on the feature map output by the encoder according to the encoding vector to achieve feature modulation under different energies. The decoder is used to progressively upsample the modulated feature map and restore it to a monochromatic projected sine map at a specified energy.

7. The method according to claim 1, wherein the second target network comprises an input layer, a feature extraction layer, and an output layer. The input layer is used to receive monochromatic projected sine waves at multiple preset energies and stack the monochromatic projected sine waves of different energies in the channel dimension to form a multi-channel two-dimensional tensor with energy and spatial dimensions. The feature extraction layer is used to perform convolution operations on the multi-channel two-dimensional tensor in the projection angle direction and the detector channel direction to extract the attenuation features across energy. The output layer is used to output sine curves for two independent channels, corresponding to the relative electron density line integral sine curve and the stopping power ratio line integral sine curve, respectively.

8. The method according to claim 1, further comprising, after obtaining the relative electron density line integral sine and the blocking power ratio line integral sine output by the second target network: The obtained line integral sine graph of the blocking ability ratio is directly output as the water equivalent path length graph; Acquire measured proton radiograph data, which includes the measured water equivalent path length measured along the same scanning geometry; The integral sine curve of the stopping power ratio is registered and the measured proton radiation data is compared with the difference. The proton range offset is determined based on the comparison results.

9. A device for determining the physical parameters of a projection domain, characterized in that, include: The acquisition module is used to acquire the raw projection data of the single-energy CT scan of the object under test; The first processing module is used to input the original projection data into the first target network to obtain a number of monochromatic projection sine waves under preset energies output by the first target network. The second processing module is used to input the monochromatic projection sine diagrams at multiple preset energies into the second target network, so that the second target network maps the monochromatic projection sine diagrams at multiple preset energies into the relative electron density line integral sine diagram and the blocking ability ratio line integral sine diagram of the projection domain and outputs the relative electron density line integral sine diagram and the blocking ability ratio line integral sine diagram. The reconstruction module is used to perform filtered back-projection reconstruction or iterative reconstruction on the relative electron density line integral sine diagram and the stopping power ratio line integral sine diagram, respectively, to obtain the spatial distribution of the relative electron density and the spatial distribution of the stopping power ratio of the object under test.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

Images