Finite angle CBCT reconstruction method and device based on fractional diffusion model

Abstract

The invention discloses a finite angle CBCT reconstruction method and device based on a fractional diffusion model. The method comprises the steps that a finite angle CBCT image is acquired; establishing a diffusion model based on fractions, wherein the diffusion model comprises a condition refining model and a backbone network model; diffusion model training based on scores; and fraction-based diffusion model sampling. According to the method, the complex relation of CBCT image tissue is learned through the diffusion model based on the fractions, a more accurate recovery result can be achieved, and recovery errors are reduced; compared with a traditional FDK method, the fractional diffusion model can quickly carry out high-quality recovery on the reconstructed image, and real-time application in clinical radiotherapy is facilitated; the method has high robustness for noise, deformation, partial shielding and the like in a limited angle CBCT image, and can stably work in a complex clinical environment.

Description

Technical Field

[0001] The invention belongs to the field of medical image processing, and in particular relates to a limited angle CBCT reconstruction method and device based on a fractional diffusion model. Background Art

[0002] Cone beam computed tomography (CBCT) has been widely used in various medical imaging tasks. In radiotherapy, CBCT images can provide the patient's three-dimensional structure as a reference for making subsequent treatment plans. However, in practical applications, full-angle rotation (360° rotation) is sometimes not possible due to the obstruction of additional equipment. Limited angle CBCT (LA-CBCT) is used to solve this problem because it has fewer requirements for the gantry rotation angle. But LA-CBCT imaging is a serious problem, which makes LA-CBCT always suffer from severe wedge artifacts and distortion.

[0003] Recently, deep learning (DL)-based medical imaging techniques have attracted much attention, and their results are impressive compared with traditional methods. The core idea of ​​deep learning-based methods is to map projection data to high-quality images through different strategies such as post-processing and hybrid processing. Despite these exciting successes in LA-CBCT imaging, the following difficulties still exist:

[0004] 1) CNN models have shown strong image restoration capabilities, but they cannot maintain clear contours and accurate structures when the rotation range is insufficient.

[0005] 2) The soft tissue contrast of CBCT images is low, and detailed features are more likely to be lost in the denoising step.

[0006] 3) The current reconstruction methods have poor universality and generalization, which increases the uncertainty of the reconstructed images. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a limited angle CBCT reconstruction method and device based on a fractional diffusion model.

[0008] In order to achieve the above object, the technical solution of the present invention is as follows:

[0009] In a first aspect, the present invention discloses a limited angle CBCT reconstruction method based on a fractional diffusion model, comprising:

[0010] Step S1: Acquire limited-angle CBCT images;

[0011] Step S2: establishing a score-based diffusion model, the diffusion model includes: a conditional refinement model and a backbone network model;

[0012] The conditional refinement model is used to receive three adjacent limited-angle CBCT images each time, compress and refine the extracted features into the latent space as the input of the backbone network model;

[0013] The backbone network model is used to maximize the feature extraction of limited-angle CBCT images;

[0014] Step S3: Based on the score-based diffusion model training, two adjacent limited-angle CBCT images and the image to be restored after adding noise are input into the diffusion model, and a rough prediction is directly performed after passing through the conditional refinement model and the backbone network model in sequence, and a perceptual loss is performed with the original image before adding noise to extract the high-order texture of the limited-angle CBCT image;

[0015] Step S4: Based on the fractional diffusion model sampling, three adjacent limited-angle CBCT images and the noise image are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in turn, the noise image is gradually perturbed slightly and the noise is removed by using the prediction correction sampling method. Finally, a noise-free restored target limited-angle CBCT image is output.

[0016] Based on the above technical solution, the following improvements can be made:

[0017] As a preferred solution, step S1 includes:

[0018] Step S1.1: Acquiring CBCT projection data;

[0019] Step S1.2: reconstructing the CBCT projection data using a filtered back-projection algorithm to obtain a limited-angle CBCT image;

[0020] Step S1.3: performing correction processing on the reconstructed limited-angle CBCT image;

[0021] Step S1.4: normalize the corrected limited-angle CBCT image.

[0022] As a preferred solution, the conditional refinement model includes: a two-dimensional discrete wavelet transform unit, a cross-attention unit and a distillation system, and the cross-attention unit and the distillation system are both composed of several convolutional layers.

[0023] As a preferred solution, the backbone network model includes: an encoder, a bottleneck layer, and a decoder;

[0024] The encoder includes: five encoder modules, each of which consists of two 3×3 convolutional layers, a ReLU activation layer, and a maximum pooling layer;

[0025] The bottleneck layer consists of two 3×3 convolutional layers and a ReLU activation layer, and no pooling is performed in this layer;

[0026] The decoder includes: five decoder modules, each of which contains a deconvolution layer;

[0027] The convolutional layers corresponding to the encoder module concatenate the parameters to the feature maps of the decoder module.

[0028] In a second aspect, the present invention discloses a limited angle CBCT reconstruction device based on a fractional diffusion model, comprising:

[0029] An image acquisition module, used for acquiring limited-angle CBCT images;

[0030] A model building module is used to build a score-based diffusion model, which includes a conditional refinement model and a backbone network model;

[0031] The conditional refinement model is used to receive three adjacent limited-angle CBCT images each time, compress and refine the extracted features into the latent space as the input of the backbone network model;

[0032] The backbone network model is used to maximize the feature extraction of limited-angle CBCT images;

[0033] The model training module is used for score-based diffusion model training. Two adjacent limited-angle CBCT images and the image to be restored after adding noise are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in sequence, a rough prediction is directly performed, and a perceptual loss is performed with the original image before adding noise to extract the high-order texture of the limited-angle CBCT image.

[0034] The model sampling module is used for fractional-based diffusion model sampling. Three adjacent limited-angle CBCT images and a noise image are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in sequence, the noise image is gradually perturbed slightly and the noise is removed by using the prediction correction sampling method. Finally, a noise-free restored target limited-angle CBCT image is output.

[0035] As a preferred solution, the image acquisition module includes:

[0036] A projection data acquisition unit, used for acquiring CBCT projection data;

[0037] A reconstruction unit, used for reconstructing the CBCT projection data using a filtered back-projection algorithm to obtain a limited-angle CBCT image;

[0038] A correction unit, used for performing correction processing on the reconstructed limited-angle CBCT image;

[0039] The normalization unit is used to perform normalization processing on the corrected limited-angle CBCT image.

[0040] As a preferred solution, the conditional refinement model includes: a two-dimensional discrete wavelet transform unit, a cross-attention unit and a distillation system, and the cross-attention unit and the distillation system are both composed of several convolutional layers.

[0041] As a preferred solution, the backbone network model includes: an encoder, a bottleneck layer, and a decoder;

[0042] The encoder includes: five encoder modules, each of which consists of two 3×3 convolutional layers, a ReLU activation layer, and a maximum pooling layer;

[0043] The bottleneck layer consists of two 3×3 convolutional layers and a ReLU activation layer, and no pooling is performed in this layer;

[0044] The decoder includes: five decoder modules, each of which contains a deconvolution layer;

[0045] The convolutional layers corresponding to the encoder module concatenate the parameters to the feature maps of the decoder module.

[0046] In a third aspect, the present invention discloses a computing device, comprising:

[0047] one or more processors;

[0048] Memory;

[0049] And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, and the one or more programs include instructions for any one of the above-mentioned limited angle CBCT reconstruction methods based on the fractional diffusion model.

[0050] In a fourth aspect, the present invention discloses a storage medium storing one or more computer-readable programs, wherein the one or more programs include instructions suitable for being loaded by a memory and executing any of the above-mentioned limited angle CBCT reconstruction methods based on the fractional diffusion model.

[0051] The present invention discloses a limited angle CBCT reconstruction method and device based on a fractional diffusion model, which has the following beneficial effects:

[0052] First, the present invention has high restoration accuracy and can achieve more accurate restoration results and reduce restoration errors by learning the complex relationship of CBCT image tissues through a score-based diffusion model.

[0053] Second, the present invention is highly efficient. Compared with the traditional FDK method, the fractional diffusion model can quickly restore the reconstructed image with high quality, which is helpful for real-time application in clinical radiotherapy.

[0054] Third, the present invention has strong robustness. It has strong robustness to noise, deformation, partial occlusion and other conditions in limited-angle CBCT images, and can work stably in complex clinical environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0056] Figure 1 The overall flow chart of the limited angle CBCT reconstruction method provided by the embodiment of the present invention.

[0057] Figure 2 A flow chart of training, sampling, etc. of a score-based diffusion model provided in an embodiment of the present invention.

[0058] Figure 3 This is a flowchart of the implementation of the limited angle CBCT reconstruction method provided by an embodiment of the present invention.

[0059] Figure 4 A block diagram of a limited-angle CBCT reconstruction device provided in an embodiment of the present invention.

[0060] Figure 5 A block diagram of a computing device provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0063] The expression of “comprising” an element is an “open” expression, which merely means that a corresponding component or step exists, and should not be interpreted as excluding additional components or steps.

[0064] In order to achieve the purpose of the present invention, in some embodiments of a limited angle CBCT reconstruction method and device based on a fractional diffusion model, as Figure 1-3 As shown, the limited angle CBCT reconstruction method includes:

[0065] Step S101: Acquire a limited-angle CBCT image;

[0066] Step S102: establishing a score-based diffusion model, the diffusion model including: a conditional refinement model (CRM) and a backbone network model;

[0067] The conditional refinement model is used to receive three adjacent limited-angle CBCT images each time, compress and refine the extracted features into the latent space as the input of the backbone network model;

[0068] The backbone network model is used to maximize the feature extraction of limited-angle CBCT images;

[0069] Step S103: Based on the score-based diffusion model training, two adjacent limited-angle CBCT images and the image to be restored after adding noise are input into the diffusion model, and a rough prediction is directly performed after passing through the conditional refinement model and the backbone network model in sequence, and a perceptual loss is performed with the original image before adding noise to extract the high-order texture of the limited-angle CBCT image;

[0070] Step S104: Based on the fractional diffusion model sampling, three adjacent limited-angle CBCT images and Gaussian noise images are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in turn, the Gaussian noise image is gradually perturbed and noise is removed using the predictor-corrector sampling method. Finally, a noise-free restored target limited-angle CBCT image is output.

[0071] Furthermore, the above embodiments are further elaborated in detail.

[0072] During radiation therapy, the patient needs to be precisely positioned and prepared before limited-angle CBCT imaging is performed. The entire process may include:

[0073] 1) Patient positioning. The accelerator will have a treatment bed, and the patient needs to be placed in a specific position, usually using a laser guidance system to ensure that the patient's posture matches the previous planning CT.

[0074] 2) Fixation devices. In order to ensure that the patient does not move during the imaging process, tools such as head frames and body fixation devices are usually used to prevent the patient from unconsciously swaying during the scan and causing unnecessary motion artifacts.

[0075] 3) Select a suitable radiotherapy accelerator for CBCT shooting. This radiotherapy accelerator can not only be used for radiotherapy, but also for imaging through the onboard cone beam CT (CBCT) device. The accelerator itself can emit x-rays and can adjust the energy level of the rays. The accelerator is equipped with an x-ray source and a flat-panel detector mounted opposite. In CBCT mode, the x-ray source and detector will rotate around the patient to obtain images at different angles. Subsequent radiotherapy work after the shooting is completed is also carried out on this device.

[0076] 4) Selection of scanning angle: Due to the limited-angle CBCT, the X-ray source and detector will not rotate a full 360 degrees around the patient, but will select a limited angle range, which can be determined according to different application scenarios and requirements.

[0077] After the patient is fixed and the necessary initial parameters are set for the accelerator, limited angle CBCT data is collected. The detector records the attenuation image (i.e., projection image) of the rays after passing through the body. During the acquisition process, the accelerator will only rotate a certain angle according to the previously set parameters and obtain the projection data of that part, rather than performing a complete scan.

[0078] Specifically, step S101 includes:

[0079] Step S101.1: Acquire CBCT projection data;

[0080] Step S101.2: reconstructing the CBCT projection data using a filtered back-projection algorithm to obtain a limited-angle CBCT image;

[0081] Step S101.3: performing correction processing on the reconstructed limited-angle CBCT image;

[0082] Step S101.4: normalize the corrected limited-angle CBCT image.

[0083] It is worth noting that after the scanning is completed, the accelerator transmits the obtained projection data (the data obtained in step S101.1) to the cloud service system for online computing, and subsequent data processing (such as: steps S101.1-S101.4) is performed on the cloud service system.

[0084] The reconstruction process of limited-angle CBCT images is more complicated than that of full-angle CBCT, because the incomplete angle leads to the "missing data" problem in image reconstruction. For this reason, the present invention first performs data preprocessing on the cloud service system, and then uses the filtered back projection (Feldkamp-Davis-Kress, FDK) algorithm to reconstruct the projection data online. In order to obtain higher reconstruction quality, a series of corrections need to be performed on the reconstructed image after the reconstruction is completed, including: air correction, geometric correction, etc. In order to normalize the image to be restored after the correction is completed, the Hu value of the image is transformed to between [-1,1] for subsequent network training and recovery.

[0085] The deep learning model used in this invention is a score-based diffusion model, which has shown excellent ability in image generation and has been successfully applied to medical imaging. It includes a forward process and a reverse process. The forward process gradually adds noise to the image, and the reverse process samples the possible distribution of the image from the noise.

[0086] The conditional refinement model includes: a two-dimensional discrete wavelet transform unit (DWT), a cross-attention unit (CA) and a distillation system (DS), and the cross-attention unit and the distillation system are both composed of several convolutional layers.

[0087] The backbone network model is the Simple-Unet backbone network, including: encoder, bottleneck layer, and decoder;

[0088] The encoder includes: five encoder modules, each of which consists of two 3×3 convolutional layers, a ReLU activation layer, and a maximum pooling layer;

[0089] The bottleneck layer consists of two 3×3 convolutional layers and a ReLU activation layer, and no pooling is performed in this layer;

[0090] The decoder consists of five decoder modules, each of which contains a deconvolution layer to double the spatial size of the feature map;

[0091] The convolutional layer corresponding to the encoder module concatenates the parameters to the feature map of the decoder module, which can combine high-resolution features.

[0092] Furthermore, there are two 3×3 convolutional layers after the deconvolution layer to extract and restore local details. The backbone network model ends with a 1×1 convolutional layer to map the feature map output by the decoder to the final output image.

[0093] The system involved in the present invention achieves the purpose of improving the model recovery ability through a special training method.

[0094] The specific principle is as follows: The score-based diffusion model (SDM) constructs a continuous diffusion process x t , indexed by t∈[0,T], which is a continuous time step, the image set x0 is fully distributed p0.

[0095] This process can be modeled as: dx = f(x, t)dt + g(t)dw as a stochastic differential equation (SDE).

[0096] f(·,t)∈R d →R d is the drift coefficient, w is the standard Wiener process, and g(·)∈R→R represents x t The diffusion coefficient.

[0097] p t (x t ) represents x t The probability indicator, p kt (x t |x k ) indicates that from x k to x t The transmission core of , 0≤k≤t≤T.

[0098] In order to t To obtain samples from the diffusion, it is necessary to define the reverse process of the diffusion. The reverse diffusion formula is as follows:

[0099]

[0100] Where dt is the negative infinitesimal time step, is the score of the marginal distribution, is the reverse standard Wiener process.

[0101] The distribution of scores can be obtained by using a time-independent score-based model s θ The (x,t) prediction is as follows:

[0102]

[0103] Among them, λ(t) is a positive weight function, and Respectively represent the data distribution p(x0) and conditional distribution p(x t |x0) samples x0 and x t When training SDEs, continuous processes are usually discretized.

[0104] The present invention adopts Variance Exploding-Stochastic Differential Equation (VE-SDE).

[0105] The perceptual loss (PL) involved in the present invention is a commonly used method in image restoration, which aims to overcome the blurring phenomenon caused by the L1 or L2 cost function. It is usually expressed as follows:

[0106]

[0107] Among them, gt is the real image, gen represents the generated result, Φ i (·) represents the feature extractor. The perceptual loss maps the image to the feature space and calculates the loss between different subjects in the high-dimensional domain.

[0108] The Condition-Refined Module (CRM) involved in the present invention has the following construction principle: the input of CRM consists of three consecutive LA-CBCT images , where s is the slice position of LA-CBCT. Next, these images are decomposed into low-frequency and high-frequency features by a two-dimensional discrete wavelet transform (DWT) unit. Then they are upsampled to match the initial size. After that, the high-frequency features will enter the criss-cross attention unit and be concatenated with the low-frequency features.

[0109] The criss-cross attention unit will calculate c s-1 , c s and c s+1 , c s Attention between:

[0110]

[0111] Among them, Q,K,V∈R H×W×C Represents the query, keyword and value of three slices, R∈R H×W×1 It is a relative position embedding.

[0112] At the end of the crisscross attention unit, the result is projected and used as the output. The output is then merged with the low-frequency features and passed through 4 convolutional blocks, each with one convolutional layer and one ReLU layer. At the end of each block, the single-channel feature map is split and the remaining feature maps are passed to the next block. When all blocks have passed, the result is merged with the previously split part and added to the remaining part. Perturbed image x t Will be connected and output at the end of CRM. CRM can be summarized as:

[0113] c(x t )=CRM(c s-1 ,c s ,c s+1 ,x t )

[0114] Among them, c(x t ) is a t The refinement condition prior.

[0115] The training method involved in the present invention can improve the recovery ability of the model. The principle is as follows: the ancestral sampling method, a sampling method commonly used in diffusion models, is used to introduce perceptual loss:

[0116]

[0117] t∈[0,1]

[0118] Among them, s θ* Representatives θ The optimal solution of represents the trained network output, c(x t ) is the input of the network, is a standard Gaussian noise with a mean of 0 and a variance of 1. is known during training.

[0119] When t-1 is 0, the above sampling formula will be modified to:

[0120]

[0121] in, is a rough prediction result. When the model is fully trained, In getting When, yes Add x t The same perturbation z t , and get the perturbed image Then pass To improve the final detail retention.

[0122] In the present invention, the pre-trained SDM model is used as Φ i (·), which can adapt to inputs of various noise amplitudes. Specifically, the SDM model uses ResUNet as the backbone network and selects the output of its intermediate block to calculate the perceptual loss.

[0123] Here, the objective function of the model is updated as follows:

[0124]

[0125] After the limited-angle CBCT image restoration is completed, the system involved in the present invention sends the calculation results to the local server through the online computing cloud service system for reference by doctors and patients for subsequent radiotherapy. At the same time, other private information of the patient will also be transmitted to the local server and stored to prevent the information from being remotely stolen or destroyed, causing property losses to the hospital or the patient.

[0126] The invention collects limited-angle CBCT data through a radiological medical accelerator, and provides a data preprocessing method used in the image restoration process, a cloud service system for online calculation, and a unique limited-angle CBCT reconstruction and restoration algorithm based on a fractional diffusion model.

[0127] The present invention stores the final results after image restoration is completed and displays them for subsequent radiotherapy. The present invention can achieve more stable limited-angle CBCT reconstruction image restoration, including KV-level and MV-level images. The present invention establishes a deep learning reconstruction model based on a fractional diffusion model to achieve high-quality limited-angle CBCT image reconstruction. The present invention uses iterative reconstruction and data consistency constraints to ensure the Z-axis continuity of the generated image.

[0128] In some embodiments, the present invention discloses a limited angle CBCT reconstruction device based on a fractional diffusion model, such as Figure 4 As shown, including:

[0129] An image acquisition module 201 is used to acquire a limited-angle CBCT image;

[0130] A model building module 202 is used to build a score-based diffusion model, the diffusion model includes: a conditional refinement model and a backbone network model;

[0131] The conditional refinement model is used to receive three adjacent limited-angle CBCT images each time, compress and refine the extracted features into the latent space as the input of the backbone network model;

[0132] The backbone network model is used to maximize the feature extraction of limited-angle CBCT images;

[0133] The model training module 203 is used for training the diffusion model based on scores. Two adjacent limited-angle CBCT images and the image to be restored after adding noise are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in sequence, a rough prediction is directly performed. The perceptual loss is performed with the original image before adding noise to extract the high-order texture of the limited-angle CBCT image.

[0134] The model sampling module 204 is used for fractional diffusion model sampling. Three adjacent limited-angle CBCT images and a noise image are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in sequence, the noise image is gradually slightly disturbed and the noise is removed by using a prediction correction sampling method. Finally, a restored target limited-angle CBCT image without noise is output.

[0135] Furthermore, the image acquisition module includes:

[0136] A projection data acquisition unit, used for acquiring CBCT projection data;

[0137] A reconstruction unit, used for reconstructing the CBCT projection data using a filtered back-projection algorithm to obtain a limited-angle CBCT image;

[0138] A correction unit, used for performing correction processing on the reconstructed limited-angle CBCT image;

[0139] The normalization unit is used to perform normalization processing on the corrected limited-angle CBCT image.

[0140] Furthermore, the conditional refinement model includes: a two-dimensional discrete wavelet transform unit, a cross-attention unit and a distillation system, and the cross-attention unit and the distillation system are both composed of several convolutional layers.

[0141] Furthermore, the backbone network model includes: encoder, bottleneck layer, decoder;

[0142] The encoder includes: five encoder modules, each of which consists of two 3×3 convolutional layers, a ReLU activation layer, and a maximum pooling layer;

[0143] The bottleneck layer consists of two 3×3 convolutional layers and a ReLU activation layer, and no pooling is performed in this layer;

[0144] The decoder includes: five decoder modules, each of which contains a deconvolution layer;

[0145] The convolutional layers corresponding to the encoder module concatenate the parameters to the feature maps of the decoder module.

[0146] Further, it should be noted that: when the limited angle CBCT reconstruction device based on the fractional diffusion model provided in the above embodiment performs limited angle CBCT image reconstruction, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the limited angle CBCT reconstruction device based on the fractional diffusion model is divided into different functional modules to complete all or part of the functions described above.

[0147] In addition, the limited angle CBCT reconstruction device based on the fractional diffusion model and the limited angle CBCT reconstruction method based on the fractional diffusion model provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0148] In addition, in some other embodiments, Figure 5 As shown, the present invention also discloses a computing device, including:

[0149] One or more processors 301;

[0150] Memory 302;

[0151] And one or more programs, wherein the one or more programs are stored in the memory 302 and configured to be executed by the one or more processors 301, and the one or more programs include instructions for the limited angle CBCT reconstruction method based on the fractional diffusion model disclosed in the above embodiment.

[0152] The processor 301 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 301 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 301 may also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 301 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 301 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0153] The memory 302 may include one or more computer-readable storage media, which may be non-transitory. The memory 302 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 302 is used to store at least one instruction, which is used to be executed by the processor 301 to implement the limited angle CBCT reconstruction method based on the fractional diffusion model provided in the method embodiment of the present invention.

[0154] In addition, the computing device may also optionally include: a peripheral device interface and at least one peripheral device. The processor 301, the memory 302 and the peripheral device interface may be connected via a bus or a signal line. Each peripheral device may be connected to the peripheral device interface via a bus, a signal line or a circuit board. Schematically, the peripheral devices include but are not limited to: a radio frequency circuit, a touch display screen, an audio circuit, and a power supply.

[0155] Of course, the computing device may also include fewer or more components, which is not limited in this embodiment.

[0156] In addition, in some other embodiments, the present invention further discloses a storage medium, which stores one or more computer-readable programs, and the one or more programs include instructions, and the instructions are suitable for being loaded by a memory and executing the limited angle CBCT reconstruction method based on the fractional diffusion model disclosed in the above embodiments.

[0157] The present invention discloses a limited angle CBCT reconstruction method and device based on a fractional diffusion model, which has the following beneficial effects:

[0158] First, the present invention has high restoration accuracy and can achieve more accurate restoration results and reduce restoration errors by learning the complex relationship of CBCT image tissues through a score-based diffusion model.

[0159] Second, the present invention is highly efficient. Compared with the traditional FDK method, the fractional diffusion model can quickly restore the reconstructed image with high quality, which is helpful for real-time application in clinical radiotherapy.

[0160] Third, the present invention has strong robustness. It has strong robustness to noise, deformation, partial occlusion and other conditions in limited-angle CBCT images, and can work stably in complex clinical environments.

[0161] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for illustrating the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which shall fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A limited angle CBCT reconstruction method based on a fractional diffusion model, characterized in that: include: Step S1: Acquire limited-angle CBCT images; Step S2: establishing a score-based diffusion model, wherein the diffusion model includes a conditional refinement model and a backbone network model; The conditional refinement model is used to receive three adjacent limited-angle CBCT images each time, compress and refine the extracted features into a latent space as input of the backbone network model; The backbone network model is used to maximize the extraction of features of limited-angle CBCT images; Step S3: Based on the score-based diffusion model training, two adjacent limited-angle CBCT images and the image to be restored after adding noise are input into the diffusion model, and a rough prediction is directly performed after passing through the conditional refinement model and the backbone network model in sequence, and a perceptual loss is performed with the original image before adding noise to extract the high-order texture of the limited-angle CBCT image; Step S4: Based on the fractional diffusion model sampling, three adjacent limited-angle CBCT images and the noise image are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in turn, the noise image is gradually perturbed slightly and the noise is removed by using the prediction correction sampling method. Finally, a noise-free restored target limited-angle CBCT image is output.

2. The limited angle CBCT reconstruction method according to claim 1, characterized in that: The step S1 comprises: Step S1.1: Acquiring CBCT projection data; Step S1.2: reconstructing the CBCT projection data using a filtered back-projection algorithm to obtain a limited-angle CBCT image; Step S1.3: performing correction processing on the reconstructed limited-angle CBCT image; Step S1.4: normalize the corrected limited-angle CBCT image.

3. The limited angle CBCT reconstruction method according to claim 1, characterized in that: The conditional refinement model includes: a two-dimensional discrete wavelet transform unit, a cross-attention unit and a distillation system, and the cross-attention unit and the distillation system are both composed of several convolutional layers.

4. The limited angle CBCT reconstruction method according to claim 1, characterized in that: The backbone network model includes: an encoder, a bottleneck layer, and a decoder; The encoder includes: five encoder modules, each of which is composed of two 3×3 convolutional layers, a ReLU activation layer, and a maximum pooling layer; The bottleneck layer consists of two 3×3 convolutional layers and a ReLU activation layer, and no pooling is performed on this layer; The decoder comprises: five decoder modules, each of the decoder modules comprising a deconvolution layer; The convolutional layers corresponding to the encoder module concatenate the parameters to the feature maps of the decoder module.

5. A limited angle CBCT reconstruction device based on a fractional diffusion model, characterized in that: include: An image acquisition module, used for acquiring limited-angle CBCT images; A model building module, used to build a score-based diffusion model, wherein the diffusion model includes: a conditional refinement model and a backbone network model; The conditional refinement model is used to receive three adjacent limited-angle CBCT images each time, compress and refine the extracted features into a latent space as input of the backbone network model; The backbone network model is used to maximize the extraction of features of limited-angle CBCT images; The model training module is used for score-based diffusion model training. Two adjacent limited-angle CBCT images and the image to be restored after adding noise are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in sequence, a rough prediction is directly performed, and a perceptual loss is performed with the original image before adding noise to extract the high-order texture of the limited-angle CBCT image. The model sampling module is used for fractional-based diffusion model sampling. Three adjacent limited-angle CBCT images and a noise image are input into the diffusion model. After passing through the conditional refinement model and the backbone network model in sequence, the noise image is gradually perturbed slightly and the noise is removed by using the prediction correction sampling method. Finally, a noise-free restored target limited-angle CBCT image is output.

6. The limited angle CBCT reconstruction device according to claim 5, characterized in that: The image acquisition module comprises: A projection data acquisition unit, used for acquiring CBCT projection data; A reconstruction unit, used for reconstructing the CBCT projection data using a filtered back-projection algorithm to obtain a limited-angle CBCT image; A correction unit, used for performing correction processing on the reconstructed limited-angle CBCT image; The normalization unit is used to perform normalization processing on the corrected limited-angle CBCT image.

7. The limited angle CBCT reconstruction device according to claim 5, characterized in that: The conditional refinement model includes: a two-dimensional discrete wavelet transform unit, a cross-attention unit and a distillation system, and the cross-attention unit and the distillation system are both composed of several convolutional layers.

8. The limited angle CBCT reconstruction device according to claim 5, characterized in that: The backbone network model includes: an encoder, a bottleneck layer, and a decoder; The encoder includes: five encoder modules, each of which is composed of two 3×3 convolutional layers, a ReLU activation layer, and a maximum pooling layer; The bottleneck layer consists of two 3×3 convolutional layers and a ReLU activation layer, and no pooling is performed on this layer; The decoder comprises: five decoder modules, each of the decoder modules comprising a deconvolution layer; The convolutional layers corresponding to the encoder module concatenate the parameters to the feature maps of the decoder module.

9. A computing device, characterized in that include: one or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, and the one or more programs include instructions for the limited angle CBCT reconstruction method based on the fractional diffusion model as described in any one of claims 1 to 4.

10. A storage medium, characterized in that The storage medium stores one or more computer-readable programs, wherein the one or more programs include instructions, and the instructions are suitable for being loaded by the memory and executing the limited angle CBCT reconstruction method based on the fractional diffusion model as described in any one of claims 1 to 4.

Images