Proton imaging scattering calibration method based on Monte Carlo and deep learning

Abstract

The invention discloses a proton imaging scattering calibration method based on Monte Carlo and deep learning, and relates to the technical field of medical imaging, proton imaging, deep learning and signal processing, and the method comprises the steps: carrying out the modeling of a proton accelerator in Monte Carlo software; the Monte Carlo software is expanded and developed to realize the control of proton angle deflection on / off, so that the control of the proton in the scattering transport process in the medium is realized; utilizing the expanded and developed Monte Carlo software to construct an energy analysis dose function library, and respectively obtaining energy analysis dose function libraries before and after scattering removal; performing proton imaging by utilizing the expanded and developed Monte Carlo software and the energy analysis dose function libraries before and after scattering removal to respectively obtain proton images before and after scattering removal; collecting a plurality of groups of proton imaging results, and constructing a sample set; training a deep learning network by using the sample set, inputting the proton image before scattering removal, and outputting the proton image after scattering removal; the trained deep learning network is used for performing scattering calibration on proton imaging.

Description

Technical Field

[0001] The present invention relates to technical fields such as medical imaging, proton imaging, deep learning and signal processing, and in particular to a proton imaging scattering calibration method based on Monte Carlo and deep learning. Background Art

[0002] The challenge facing modern radiotherapy is to increase the radiation dose to the tumor while reducing the dose to surrounding critical organs. Proton radiotherapy offers significant advantages in tumor dosimetry compared to traditional photon radiotherapy, as the Bragg peak produced by proton injection allows protons to precisely target the tumor, offering significant potential for improved target coverage and organ protection. Furthermore, protons have a higher biological effect than photons, making them more effective in killing tumor cells.

[0003] However, the uncertainty of proton range, for example, due to positioning errors during treatment, anatomical changes in the patient, organ movement caused by breathing, etc., may cause insufficient target dose coverage and excessive irradiation of critical organs. The HU (Honnig unit, a standard unit for describing tissue density) used for dose calculation in radiotherapy planning systems can cause a 3.5% error in the relative tissue power curve (RSP curve), and this error can reach 5% in lung tissue. Therefore, in order to fully realize the potential of proton radiotherapy, it is necessary to accurately evaluate the range of protons. Better assessment of the uncertainty of proton range can help reduce the boundary range required for robust optimization of the target during treatment planning and improve the conformality of the dose.

[0004] As early as the 1960s, researchers proposed imaging protons by measuring the energy lost along their path through the patient. Proton imaging directly provides information about the proton's range, enabling iterative optimization of the HU-relative tissue power curve (RSP curve) in clinical practice. Proton imaging also provides anatomical information about the patient, and its soft tissue contrast surpasses that of X-rays, making it extremely valuable for adaptive proton radiotherapy. Studies have shown that proton imaging can be used for pre-treatment positioning verification.

[0005] A single flat-panel detector offers a large readout area and a spatially resolved two-dimensional signal, making it ideal for proton imaging. To evaluate water-equivalent thickness using a flat-panel detector, a library of energy-resolved dose functions for different water-equivalent thicknesses is required as a calibration reference. When an object is irradiated with an imaging field of varying energies, the signal obtained by each pixel on the flat-panel detector is represented by an energy-resolved dose function. The detector signal is then determined by comparing the curves in the energy-resolved dose function library. Because protons are charged particles, their emission direction is deflected by a series of physical processes when they interact with matter. When protons pass through the gap between two materials, noise is generated in the energy-resolved dose function on the detector. This effect is known as range aliasing. Range aliasing significantly impacts proton image quality, introducing noise and reducing image resolution. Rapidly obtaining high-quality proton images and addressing range aliasing are current research priorities.

[0006] In recent years, deep learning has become increasingly important in the field of medical imaging. It's also widely used in radiotherapy, for example in deep learning-based target delineation and automated radiotherapy planning. Deep learning algorithms in medical image processing can be broadly categorized as supervised and unsupervised. Supervised learning primarily uses paired data as samples, but paired data is often difficult to obtain. Unsupervised learning, on the other hand, doesn't require paired data input, making it suitable for proton image denoising.

[0007] Existing scatter correction methods cannot achieve good scatter correction. Some researchers remove the dose signal generated by protons with a deflection direction greater than 2.5° from the original signal as a scatter signal during simulation, but this method can only partially alleviate the impact of the scatter signal and cannot solve the range aliasing problem; some researchers correct the range aliasing problem by forging noise signals on the detector and combining deep learning, but this method can only take into account some situations and cannot take into account the complex situation of detector signal noise in human tissue. Summary of the Invention

[0008] To overcome the above-mentioned deficiencies in the prior art, the present invention provides a proton imaging scattering calibration method based on Monte Carlo and deep learning. This method innovatively modifies the physical process in the Monte Carlo software to obtain labeled images that can be used for training. Furthermore, a deep learning method is used to perform end-to-end training on proton image data pairs before and after descattering. The resulting model can be used to perform scattering correction on proton images actually measured.

[0009] To achieve the above object, the present invention adopts the following technical solutions, including:

[0010] A Monte Carlo and deep learning-based proton imaging scattering calibration method includes the following:

[0011] Modeling a proton accelerator in Monte Carlo software;

[0012] Expand and develop Monte Carlo software to enable on / off control of proton angle deflection during proton imaging, thereby controlling the proton transport process with or without scattering in the medium;

[0013] The energy-resolved dose function library was constructed using the expanded Monte Carlo software, and the energy-resolved dose function library before and after de-scattering was obtained respectively.

[0014] Proton imaging was performed using the developed Monte Carlo software and the energy-resolved dose function library before and after descattering, and the proton images before and after descattering were obtained respectively.

[0015] Collect several sets of proton imaging results to construct a sample set;

[0016] The deep learning network is trained using a sample set, with the input being the proton image before de-scattering and the output being the proton image after de-scattering. The trained deep learning network is used to perform scattering calibration for proton imaging.

[0017] Preferably, the energy-analyzed dose function library is constructed using the expanded Monte Carlo software, as shown below:

[0018] Setting a set of monoenergetic proton beams of different energies for imaging an object;

[0019] A first solid water with a set thickness is placed in sequence in the beam direction as a range shifter and a second solid water with a set thickness is placed as a detector, and a third solid water with an increasing thickness is placed at the isocenter position between the range shifter and the detector;

[0020] In the expanded and developed Monte Carlo software, proton angle deflection is enabled, that is, a simulation is performed for the case without descattering. Using a set of monoenergetic proton beams with different energies, an energy-resolved dose function curve is established for each thickness of the third solid water, thus forming an energy-resolved dose function library before descattering.

[0021] In the expanded and developed Monte Carlo software, the proton angle deflection is turned off, that is, the descattering situation is simulated, and a set of monoenergetic proton beams with different energies are used to establish an energy-resolved dose function curve for each thickness of the third solid water, thereby forming an energy-resolved dose function library after descattering.

[0022] Preferably, proton imaging is performed using the expanded Monte Carlo software and energy-resolved dose function library, as shown below:

[0023] Place the object to be imaged at the isocenter position;

[0024] A set of monoenergetic proton beams with different energies are used to image the object to be imaged. An energy-resolved dose function curve is obtained at each pixel of the object to be imaged. The energy-resolved dose function curve at each pixel is compared with the curve in the energy-resolved dose function library, and the closest curve is found from the energy-resolved dose function library. The thickness corresponding to the closest curve is the equivalent thickness value at the pixel.

[0025] Preferably, the patient's CT image is input into the expanded and developed Monte Carlo software, and the human body material, ie, the object to be imaged, is obtained using the CT image and placed at the isocenter position.

[0026] Preferably, the CT images of the lung tumor patients are input into the expanded and developed Monte Carlo software to perform proton imaging, thereby obtaining several sets of proton imaging results of the lung tumor patients;

[0027] The deep learning network is trained using the results of several sets of proton imaging of lung tumor patients. The trained deep learning network is then used to perform scattering calibration for proton imaging.

[0028] Preferably, during proton imaging, protons interact with matter and are deflected by four physical processes: protons interact with electrons outside the nucleus of atoms, protons interact with the Coulomb force of the nucleus, protons collide with the nucleus elastically, and protons collide with the nucleus inelasticly.

[0029] In the Monte Carlo software, the codes of the four physical processes are rewritten so that after the protons pass through the four physical processes in the medium, only the energy transfer in the proton transport process is retained and the angle change is cancelled.

[0030] A computer program product comprising a computer program / instruction, which, when executed by a processor, implements the Monte Carlo and deep learning-based proton imaging scatter calibration method.

[0031] A readable storage medium stores a computer program thereon, wherein when the computer program is executed, the proton imaging scattering calibration method based on Monte Carlo and deep learning is implemented.

[0032] An electronic device comprises a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a Monte Carlo and deep learning-based proton imaging scattering calibration method is implemented.

[0033] The advantages of the present invention are:

[0034] (1) The present invention provides a proton imaging scattering calibration method based on Monte Carlo and deep learning. The method innovatively obtains labeled images that can be used for training by modifying the physical process in the Monte Carlo software, and performs end-to-end training on proton image data before and after descattering through a deep learning method. The obtained model can be used to perform scattering correction on the proton images obtained by actual measurement.

[0035] (2) This paper proposes a method for modifying the proton image signal by modifying the physical process during proton simulation. This method expands upon existing Monte Carlo software, retaining the proton-related physical process switches during the simulation process. This method further expands existing Monte Carlo simulation software and develops a universal Monte Carlo program that can be used for proton imaging scattering research. This lays a solid foundation for achieving proton imaging descattering combined with deep learning.

[0036] (3) The present invention solves the range aliasing problem in the proton imaging detector signal by changing the proton-related physical process in the Monte Carlo software.

[0037] (4) This invention uses a deep learning network to perform end-to-end training on proton images of lung diseases, and obtains a model that can be used for scatter correction of proton images of lung diseases. This allows for more accurate monitoring of the radiation range of lung tumor patients before radiotherapy.

[0038] (5) This paper focuses on the scattering problem in proton imaging. By generating reference images through Monte Carlo simulation and combining deep learning techniques, we propose an end-to-end proton image scattering correction deep learning framework for lung diseases. This framework corrects the scattering of the measured images, mitigating the impact of range aliasing on proton imaging quality.

[0039] (6) The deep learning-based proton imaging scatter correction method described in the present invention innovatively modifies the physical process in Monte Carlo software to obtain labeled images that can be used for training. The deep learning method is then used to perform end-to-end training on lung proton image data, both before and after descattering. The resulting model can be used to perform scatter correction on actual measured lung proton images.

[0040] (7) The de-scattered image of the present invention can be used to verify the proton range before proton radiotherapy, making proton radiotherapy more accurate. The de-scattered proton image more accurately reflects the patient's equivalent water depth information. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A brief diagram of the Monte Carlo model.

[0042] Figure 2A brief schematic diagram of the virtual source location.

[0043] Figure 3 A comparison chart of the simulation and measurement results of eight groups of different monoenergetic proton beams.

[0044] Figure 4 A comparison chart of the simulation and measurement results of the extended Bragg peak at different depths.

[0045] Figure 5 Schematic diagram of the compilation and linking between Topas, Geant4 and other libraries.

[0046] Figure 6 This is a compilation-linking diagram between the expanded and developed Topas and Geant4 and other libraries.

[0047] Figure 7 Flowchart of the Monte Carlo software developed for simulating a single proton transport process.

[0048] Figure 8 Schematic diagram for setting simulation conditions for the energy-resolved dose function library.

[0049] Figure 9 Energy resolved dose function library before de-scattering.

[0050] Figure 10 A library of energy-resolved dose functions after descattering.

[0051] Figure 11 This is a proton image of the lung before de-scattering.

[0052] Figure 12 This is the proton image of the lung after descattering.

[0053] Figure 13 This is a flow chart of a proton imaging scattering calibration method based on Monte Carlo and deep learning in the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] Depend on Figure 13 As shown, the present invention provides a proton imaging scattering calibration method based on Monte Carlo and deep learning, comprising the following steps:

[0056] S1, Modeling of a proton accelerator in Monte Carlo software.

[0057] Depend on Figure 1 and Figure 2 As shown, in this embodiment, the Varian Probeam proton accelerator is modeled in Monte Carlo software. Monte Carlo software is a software that simulates particle collisions in the laboratory. The Varian Probeam proton accelerator mainly provides point scanning as the proton scanning method. In the point scanning model, the proton beam is mainly affected by the magnet. After leaving the head, it is only affected by the range shifter and the air. Therefore, the simulation of the treatment head and the scanning magnet is omitted.

[0058] The data needed for actual modeling are two parts. The first part is the phase space parameters of the initial beam, and the other part is the position x, y and divergence angle θx, θy of the proton beam in the source plane.

[0059] The phase space parameters of the initial beam mainly include the beam spot size σ in the X and Y directions, the beam spot divergence angle θ, and the correlation term σθ between the beam spot size and the divergence angle. These values can be obtained by measuring the beam spot flux at five different positions from the isocenter using the detector and the least squares method using formula (1):

[0060] σ(Z) 2 =σ(0) 2 +2σ(0)θ(0)Z+θ(0) 2 Z 2 ; (1)

[0061] Where σ(Z) is the beam spot size at a distance of Z cm from the source plane, σ(0) is the beam spot size at the source plane, θ(0) is the beam spot divergence angle at the source plane, and Z is the distance from the source plane.

[0062] The position x, y and divergence angle θx, θy of the proton beam in the source plane can be obtained by combining the planning information in the radiotherapy plan with the virtual source axis distance of the proton accelerator and according to formulas (2) and (3):

[0063] x=x iso ×(f x -d / f x ), y=y iso ×(f y -d / f y ); (2)

[0064] θx=tan -1 (y iso / f y )×(180 / π),θy=-tan -1 (x iso / f x )×(180 / π); (3)

[0065] Among them, x iso ,y iso is the planned beam spot position in the isocenter plane; f x ,f y is the virtual source wheelbase in the X and Y directions. The virtual source wheelbase refers to the distance from the virtual source to the isocenter point; d is the distance from the source plane to the isocenter plane.

[0066] The Monte Carlo modeling results can be verified by simulating the depth-integrated curves of protons of different energies in water in the software and by comparing the extended Bragg peak with the curves obtained by actual detector measurements. Figure 3 and Figure 4 shown.

[0067] Figure 3 The following is a comparison chart of the simulation and measurement results of eight different monoenergetic proton beams (70MeV, 90MeV, 110MeV, 130MeV, 150MeV, 170MeV, 190MeV, and 210MeV). Figure 3 In the figure, the blue solid line represents the measurement results of the multi-layer ionization chamber, the red dotted line represents the Monte Carlo simulation results, the horizontal axis is the depth (cm), and the vertical axis is the relative dose.

[0068] Figure 4 Comparison of simulation and measurement results of extended Bragg peaks at different depths. Figure 4 In the middle, (a) is 10×10cm 2 Comparison of the simulated and measured results of the modulation depth with a field of view of 15 cm and a range of 10 cm. (b) is 10×10 cm 2 Comparison of the simulated and measured results of the modulation depth with a field of view of 25 cm and a range of 10 cm. (c) is 10×10 cm 2 Comparison chart of the simulated results and the measured results of the 30cm field of view and 10cm modulation depth. Figure 4 The horizontal axis is depth (cm), the vertical axis is relative dose, the blue solid line represents the multi-layer ionization chamber measurement results, and the red dotted line represents the Monte Carlo simulation results.

[0069] S2, the expansion and development of Monte Carlo software, realizes the control of proton angle deflection on / off during proton imaging, thereby realizing the control of proton transport process with / without scattering in the medium.

[0070] During proton imaging, protons interact with matter in four primary ways: 1. Interaction between protons and electrons; 2. Coulomb interaction between protons and nuclei; 3. Elastic collisions between protons and nuclei; and 4. Inelastic collisions between protons and nuclei. In the first process, the energy lost by the proton is transferred to the electrons. If the energy gained by the electrons exceeds their binding energy, they break free from the nucleus and escape, a process called ionization. If the transferred energy is insufficient to free the electrons, they may transition to higher energy levels, a process called excitation. Although the proton's deflection angle is small during this single process, it is a continuous process, and the proton continues to lose energy and change angle as it passes through the medium. The second process, called multiple Coulomb scattering, is generally elastic scattering, with minimal energy loss per shot, but with large deflection angles, significantly impacting proton imaging quality. For the third and fourth physical processes, namely the elastic and inelastic collisions between protons and atomic nuclei, the probability of occurrence is very small and can be almost ignored, and the actual impact on proton imaging is very small.

[0071] According to the above physical process, theoretically speaking, when protons are transported in a medium, if there is no angular deflection, the quality of proton imaging can be greatly improved. Based on this idea, the present invention expands and develops the existing open source Monte Carlo software Topas, with the aim of achieving a transport process in which protons are not scattering in a medium (thereby not causing proton angular deflection), and ultimately obtaining high-quality proton imaging (low-noise imaging) without scattering effects, as the label value for subsequent machine learning. Specifically, the present invention rewrites the code of the physical process related to protons on the existing open source Monte Carlo software Topas, so that after the protons pass through the above four physical processes in the medium, only the energy transfer in the proton transport process is retained, and its angle change is cancelled. At the same time, in order to subsequently perform multi-physical process verification, for each of the above physical processes, the artificial setting of the above-mentioned angle deflection is provided with parameter control in the form of a script during the expansion and development process. The basic principles of the relevant expansion and development will be briefly described below.

[0072] The Monte Carlo software Topas is essentially an extension of the open source particle transport program Geant4, which is a multi-core parallel program package used to simulate various types of particle transport processes, and is primarily written in C++. To use Geant4 for particle transport simulation, users need to write C++ programs to implement interfaces for simulation settings such as particle sources, events, counts, geometry, and dielectric materials, and compile them into executable programs together with the Geant4 kernel library. The Geant4 kernel will call the relevant interfaces at specific times to implement the user-specified simulation and calculation processes. Based on Geant4, Topas encapsulates settings such as particle sources, events, counts, geometry, and dielectric materials. Users only need to write text-type parameter values to complete the relevant settings, simplifying the difficulty of Geant4 simulation settings.

[0073] Figure 5 This is a diagram of the compilation and linking of Topas with the underlying Geant4 and other major libraries. Figure 5 As shown, the Topas source code is at the top of the compilation chain. To implement the above technical approach, either the Geant4 source code or the Topas source code can be extended and developed. In this embodiment, the Topas source code is extended and developed. This way, when the underlying Geant4 version changes, there is no need to redevelop the program, ensuring program stability. Figure 6 To expand the compilation and linking diagram of Topas with the underlying Geant4 and other major libraries, Figure 6 As shown, in this embodiment, script control extension code and physical process extension code are additionally added to the Topas source code to realize the on / off control of the proton angle deflection in the above-mentioned specific physical process.

[0074] The following describes the details of the Geant4 physical process extension and Topas script extension code of the present invention. Figure 1 shows the UML diagram of the main C++ classes of the Geant4 physical process extension and Topas script extension in this embodiment.

[0075] The TsParamterManager class is the main class used by Topas to read script control files and parameters. Add the extended control parameter "ProtonDescattering" to the Topas script. This parameter has four possible values: "de_Hadronic_Elastic_Scattering," "de_Hadronic_InElastic_Scattering," "de_MultiCoulomb_Scattering," and "de_Hadronic_Ionisation_Scattering."

[0076] The TsModularPhysicsList class is the core class for building the physical process list. At the same time, in the constructor of the TsModularPhysicsList class, by calling the GetStringVector and GetVectorLength functions of the TsParamterManager class, the four combination values of the extended control parameter "ProtonDescattering", "de_Hadronic_Elastic_Scattering", "de_Hadronic_InElastic_Scattering", "de_MultiCoulomb_Scattering", and "de_Hadronic_Ionisation_Scattering" are read.

[0077] According to the value set by the extended control parameter "ProtonDescattering", the corresponding values of the Boolean type member variables de_Hadronic_Elastic_Scattering, de_Hadronic_InElastic_Scattering, de_MultiCoulomb_Scattering, and de_Hadronic_Ionisation_Scattering are assigned by calling the resolve() method in the custom class G4PhysicListExtParam.

[0078] Afterwards, when constructing the physical process list, the object of type G4PhysicListExtParam is passed to realize the on / off of the proton angle deflection in the above four physical processes.

[0079] Specifically, in the constructor of the TsModularPhysicsList class, the physical process is registered into the member variable fPhysicsTable. The member variable fPhysicsTable is a table consisting of multiple pairs of <physical process name, physical process construction class>. In this embodiment, the physical process construction class Creator <t>Expand to become a class Creator_WithParam that can accept G4PhysicListExtParam type parameters during class construction <t>. Due to Creator_WithParam <t>The class overloads the operator() operator, so in the subsequent instantiation of the physics process class, the G4PhysicListExtParam type parameter can be passed in.

[0080] Specifically, in the extended class G4EmBuilder_Ext, the value of the member variable de_MultiCoulomb_Scattering of the G4PhysicListExtParam type is used to determine whether or not to change the proton deflection angle when the proton performs multi-Coulomb scattering (but the proton energy will definitely change); in the extended classes G4BraggModel_Ext, G4ICRU73QOModel_Ext, and G4BetheBlochModel_Ext, the value of the member variable de_Hadronic_Ionisation_Scattering of the G4PhysicListExtParam type is used to determine whether or not to change the proton deflection angle when the proton interacts with the extranuclear electrons of the atom (but the proton energy will definitely change); in the extended class G4Had ronElasticPhysics_Ext and the extended class G4HadronElasticPhysicsHP_Ext, the value of the G4PhysicListExtParam type member variable de_Hadronic_Elastic_Scattering is True / False to determine whether to not change the proton deflection angle when elastic scattering occurs between the proton and the nucleus (but the proton energy will definitely change); in the extended class G4HadronPhysicsQGSP_BIC_Ext and the extended class G4HadronPhysicsQGSP_BIC_HP_Ext, the value of the G4PhysicListExtParam type member variable de_Hadronic_InElastic_Scattering is True / False to determine whether to turn off the inelastic scattering of protons.

[0081] Figure 7 The flow chart of simulating a single proton transport process after the expansion and development of Topas / Geant4 in this invention is as follows: Figure 7 As shown, the process is divided into two parts. The first part is the process from reading the control script to injecting the expanded physical process list of the present invention. The second part is the process of calling the expanded physical process list of the present invention in the proton transport calculation.

[0082] S3, Simulation method for paired proton images.

[0083] S31, using the expanded and developed Monte Carlo software to build an energy-resolved dose function library, and obtain the energy-resolved dose function library before and after descattering respectively.

[0084] Under the condition of fixed beam current and fixed direction, the dose generated by a monoenergetic proton beam at a given depth in a given medium is unique. Therefore, the dose sequence generated by a group of monoenergetic proton beams at the same depth in the same medium is also unique. Set up a group of monoenergetic proton beams with different energies, namely E1, E2, ..., E n , used to image an object, the energy-resolved dose function is unique at a given depth in a medium. Based on this characteristic, by placing solid water of varying thickness at the isocenter and imaging it with monoenergetic proton beams of varying energies, energy-resolved dose function curves for equivalent water depths of varying thicknesses are generated, thus forming a ruler (building an energy-resolved dose function library). Subsequently, the energy-resolved dose curve at each pixel measured by the detector is matched with the curves in the energy-resolved dose function library using the minimum mean square error method to obtain the projection value.

[0085] The energy-resolved dose function library for imaging is obtained through Monte Carlo simulation. Figure 8 Schematic diagram of the simulation conditions for the energy-resolved dose function library. To save simulation time, all simulations used a single beam spot. The proton beam was directed perpendicular to the solid water phantom. A first solid water layer with a thickness of 4 cm was placed in the beam direction as a range shifter, and a second solid water layer with a length × width × thickness of 10 × 10 × 0.5 cm was placed as a detector. A third solid water layer with increasing thickness was placed at the isocenter between the range shifter and the detector. The thickness increased at 5 mm intervals, with thicknesses of 0.5 mm, 1 mm, 1.5 mm, and so on, placed at the isocenter. The simulated energy range was 70 MeV to 244 MeV, with an energy interval of 3 MeV. A total of 59 monoenergetic proton beams with different energies, i.e., 59 imaging energy slices, were used.

[0086] In the expanded and developed Monte Carlo software, simulations were performed for both the pre-descattering and de-scattering situations. During the simulation, a set of monoenergetic proton beams with different energies were used to establish an energy-resolved dose function curve for each thickness of the third solid water. Each curve corresponds to an equivalent thickness, thus forming an energy-resolved dose function library before and after de-scattering. Figure 9 and Figure 10 These are two sets of energy-analyzed dose function libraries obtained by simulation before and after de-scattering, Figure 9 and Figure 10 Curves of different colors correspond to different thicknesses.

[0087] Subsequently, the object to be imaged is placed at the isocenter position, and the best matching curve is found according to the energy dose analysis curve at each pixel measured by the detector, so as to obtain the equivalent thickness value corresponding to each pixel.

[0088] S32, using the expanded Monte Carlo software and the energy-resolved dose function library before and after descattering to perform proton imaging, and obtain proton images before and after descattering respectively.

[0089] The CT images of lung tumor patients were input into the developed Monte Carlo software. The human body material (the object to be imaged) was obtained from the CT images and placed at the isocenter position to simulate proton imaging before and after descattering. The proton imaging field size was 30×30cm. 2 The beam spot spacing is 5mm, 1MU / beam spot, and the proton imaging energy range is 70-244MeV; the detector is placed behind the object to be imaged and has a size of 30×30cm 2 , thickness is 5mm, each detector unit is 1×1mm 2 Resolution. To facilitate comparison with the energy-resolved dose function library, the measured energy-resolved dose function curve at each pixel and the curves in the energy-resolved dose function library were normalized. Curves were interpolated to 1 MeV energy intervals using cubic splines, and between curves were linearly interpolated to 1 mm equivalent thickness intervals. The equivalent thickness value for each pixel was determined by minimizing the mean square error between the measured energy-resolved dose function curve and the curves in the energy-resolved dose function library.

[0090] S4, using several sets of proton imaging results to train a deep learning network, with the input being the proton image before de-scattering and the output being the proton image after de-scattering; the trained deep learning network is used to perform scattering calibration for proton imaging.

[0091] Neural network for lung diseases

[0092] Several CT images of patients with lung tumors were selected, and proton images before and after descattering were obtained to form paired data samples. For a single patient's CT images, a set of paired data was simulated at regular angle intervals to obtain a set of paired data. For four-dimensional CT images, a set of paired data was simulated at regular angle intervals for each phase of the image. A deep learning network was trained using proton images before and after descattering, which can be input into existing network models such as Cycle-GAN for training, resulting in a deep learning network that can be used to correct for proton image scattering. Descattered proton images show significant improvements in signal-to-noise ratio, equivalent thickness accuracy, and image edge clarity. Figure 11 and Figure 12 These are the simulation results of proton imaging of the patient's lungs before and after de-scattering.

[0093] This embodiment proposes a proton imaging scatter correction method for lung tumors combined with deep learning. By modifying the physical processes in Monte Carlo simulation software, a scatter-corrected reference lung proton image is generated. The unscattered proton image and the scatter-corrected proton image form paired data for training a deep neural network. Scatter correction can be applied to lung proton images. Actual measured lung proton images of patients are then fed into the trained network to produce scatter-corrected proton images.

[0094] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.< / t> < / t> < / t>

Claims

1. A proton imaging scattering calibration method based on Monte Carlo and deep learning, characterized in that: Includes the following: Modeling a proton accelerator in Monte Carlo software; Expand and develop Monte Carlo software to enable on / off control of proton angle deflection during proton imaging, thereby controlling the proton transport process with or without scattering in the medium; The energy-resolved dose function library was constructed using the expanded Monte Carlo software, and the energy-resolved dose function library before and after de-scattering was obtained respectively. Proton imaging was performed using the expanded Monte Carlo software and the energy-resolved dose function library before and after descattering, and the proton images before and after descattering were obtained respectively. Collect several sets of proton imaging results to construct a sample set; The deep learning network is trained using a sample set, with the input being the proton image before de-scattering and the output being the proton image after de-scattering. The trained deep learning network is used to perform scattering calibration for proton imaging.

2. The Monte Carlo and deep learning based proton imaging scattering calibration method according to claim 1, characterized in that: The energy-analyzed dose function library is constructed using the expanded Monte Carlo software, as shown below: Setting a set of monoenergetic proton beams of different energies for imaging an object; A first solid water with a set thickness is placed in sequence in the beam direction as a range shifter and a second solid water with a set thickness is placed as a detector, and a third solid water with an increasing thickness is placed at the isocenter position between the range shifter and the detector; In the expanded and developed Monte Carlo software, proton angle deflection is enabled, that is, a simulation is performed for the case without descattering. Using a set of monoenergetic proton beams with different energies, an energy-resolved dose function curve is established for each thickness of the third solid water, thus forming an energy-resolved dose function library before descattering. In the expanded and developed Monte Carlo software, the proton angle deflection is turned off, that is, the descattering situation is simulated, and a set of monoenergetic proton beams with different energies are used to establish an energy-resolved dose function curve for each thickness of the third solid water, thereby forming an energy-resolved dose function library after descattering.

3. The Monte Carlo and deep learning based proton imaging scattering calibration method according to claim 2, characterized in that: Proton imaging was performed using the expanded Monte Carlo software and energy-resolved dose function library, as shown below: Place the object to be imaged at the isocenter position; A set of monoenergetic proton beams with different energies are used to image the object to be imaged. An energy-resolved dose function curve is obtained at each pixel of the object to be imaged. The energy-resolved dose function curve at each pixel is compared with the curve in the energy-resolved dose function library, and the closest curve is found from the energy-resolved dose function library. The thickness corresponding to the closest curve is the equivalent thickness value at the pixel.

4. The Monte Carlo and deep learning based proton imaging scattering calibration method according to claim 3, characterized in that: The patient's CT image is input into the expanded and developed Monte Carlo software, and the human body material, that is, the object to be imaged, is obtained using the CT image and placed at the isocenter position.

5. The Monte Carlo and deep learning based proton imaging scattering calibration method according to claim 4, characterized in that: The CT images of lung cancer patients were input into the developed Monte Carlo software to perform proton imaging, and several sets of proton imaging results of lung cancer patients were obtained. The deep learning network is trained using the results of several sets of proton imaging of lung tumor patients. The trained deep learning network is then used to perform scattering calibration for proton imaging.

6. The Monte Carlo and deep learning based proton imaging scattering calibration method according to claim 1, characterized in that: During proton imaging, protons interact with matter and are deflected by four physical processes: protons interact with the electrons outside the nucleus of atoms, protons interact with the Coulomb force of the nucleus, protons collide with the nucleus elastically, and protons collide with the nucleus inelasticly. In the Monte Carlo software, the codes of the four physical processes are rewritten so that after the protons pass through the four physical processes in the medium, only the energy transfer in the proton transport process is retained and the angle change is cancelled.

7. A computer program product, characterized in that It includes a computer program / instruction, which, when executed by a processor, implements a Monte Carlo and deep learning-based proton imaging scattering calibration method according to any one of claims 1 to 6.

8. A readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, the proton imaging scattering calibration method based on Monte Carlo and deep learning according to any one of claims 1 to 6 is implemented.

9. An electronic device, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the Monte Carlo and deep learning-based proton imaging scattering calibration method according to any one of claims 1 to 6.

Images