Radiotherapy image reconstruction method based on self-supervised learning

Abstract

The invention discloses a radiotherapy image reconstruction method based on self-supervised learning. The method comprises the following steps: S1, obtaining a low-contrast radiotherapy image sequence data set; s2, constructing the low-contrast radiotherapy image sequence data set into a key-query pair according to view angle difference, inputting the key-query pair into a MoCov3 self-supervision feature encoder model, and obtaining an initial feature representation set; s3, obtaining an image enhancement reconstruction result set; s4, generating an updated wolf pack parameter set; s5, outputting the optimal reconstruction parameter configuration and the corresponding high-contrast radiotherapy enhanced image; and S6, inputting the high-contrast radiotherapy enhanced image into a closed-loop quality evaluation module, and performing final quality verification by using an organ sketching consistency index and a dose recalculation error. According to the invention, the response capability of the encoder to the tumor boundary is obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image reconstruction, and in particular to a radiotherapy image reconstruction method based on self-supervised learning. Background Art

[0002] With the development of medical imaging technology, radiotherapy has become one of the important means of precise tumor treatment, especially in the treatment of solid tumors such as lung cancer, brain tumors, and breast cancer. However, due to the frequent use of cone-beam computed tomography systems for dose verification and image guidance during radiotherapy, there are widespread problems such as low contrast, high noise, and blurred structures. These problems seriously affect doctors' ability to accurately delineate tumor boundaries and locate key organs, thereby restricting the accuracy and efficacy of radiotherapy plans.

[0003] In existing technologies, enhancement methods for low-contrast radiotherapy images mainly focus on traditional image reconstruction algorithms and post-processing filtering technologies. Although they have achieved certain results in improving image clarity, they generally lack structural perception capabilities and cannot effectively enhance tumor edge details. In addition, some deep learning methods introduce supervised learning strategies. Although they perform well in image reconstruction accuracy, their reliance on large amounts of high-quality labeled data limits their widespread application in clinical practice.

[0004] Self-supervised learning has become an emerging technology to replace supervised learning in recent years. However, the current mainstream self-supervised models have poor structural generalization in medical imaging applications and difficulty in capturing the microstructural characteristics of tumor boundaries. At the same time, existing image enhancement methods generally adopt fixed parameter design and lack the ability to adaptively adjust for different imaging scenarios, resulting in unstable image enhancement effects and difficulty in meeting the dual requirements of clinical precision radiotherapy for image structure clarity and dose consistency.

[0005] Therefore, there is an urgent need for a new image reconstruction method that can take into account both structure recognition and dose accuracy and has adaptive enhancement capabilities to solve the above problems. Summary of the Invention

[0006] One object of the present invention is to propose a radiotherapy image reconstruction method based on self-supervised learning, which significantly improves the encoder's ability to respond to tumor boundaries.

[0007] A radiotherapy image reconstruction method based on self-supervised learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Acquire low-contrast radiotherapy image sequence data, perform scattering artifact suppression, metal artifact correction, and dose normalization preprocessing on the low-contrast radiotherapy image sequence data, and obtain a low-contrast radiotherapy image sequence data set;

[0009] S2. Construct a MoCov3 self-supervised feature encoder model and initialize the momentum encoder weights. Construct the low-contrast radiotherapy image sequence dataset into key-query pairs based on viewpoint differences and input them into the MoCov3 self-supervised feature encoder model. Perform self-supervised contrastive learning and obtain the MoCov3 self-supervised feature encoder model. Use the MoCov3 self-supervised feature encoder model to perform inference on the low-contrast radiotherapy image sequence dataset to obtain an initial feature representation set.

[0010] S3. Initialize the wolf pack position vectors of the discrete gray wolf optimization algorithm, uniformly discretize the initial feature representation set to form an initial wolf pack parameter set, and for each wolf pack position vector in the initial wolf pack parameter set, couple the corresponding parameter combination with the initial feature representation set, perform image enhancement reconstruction, and obtain an image enhancement reconstruction result set;

[0011] S4. Calculate the tumor edge clarity index and dose consistency index based on the image enhancement reconstruction result set. Fusion of these indexes yields the wolf pack fitness vector. Based on the wolf pack fitness vector, the discrete gray wolf optimization algorithm position update strategy, dynamic step size control, and prey escape mechanism are used to update the wolf pack position vector and generate an updated wolf pack parameter set.

[0012] S5. Drive the MoCov3 self-supervised feature encoder model with the updated wolf pack parameter set, synchronously adjust and refresh the momentum encoder queue based on the learned temperature parameters, obtain an updated feature representation set, re-perform image enhancement reconstruction using the updated feature representation set and the updated wolf pack parameter set, and obtain an iterative image enhancement reconstruction result set. Repeat steps S4 to S5 until the wolf pack fitness vector meets the preset convergence condition, and output the optimal reconstruction parameter configuration and the corresponding high-contrast radiotherapy-enhanced image.

[0013] S6. Input the high-contrast radiotherapy enhanced image into the closed-loop quality evaluation module, and perform final quality verification using the organ outline consistency index and dose recalculation error. If the final quality verification result meets the clinical threshold, the high-contrast radiotherapy image result and the optimal reconstruction parameter configuration after final quality verification are output and the optimal reconstruction parameter configuration is stored for initialization of image enhancement reconstruction on subsequent treatment days. If the final quality verification result does not meet the clinical threshold, the final quality verification result is fed back to the wolf pack fitness vector and returns to step S4 to continue iterating until the clinical threshold is met.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Acquire low-contrast radiotherapy image sequence data by a cone-beam computed tomography system within a preset time window to form a low-contrast radiotherapy image frame set, where each radiotherapy image frame in the low-contrast radiotherapy image frame set is a two-dimensional image having uniform height and width dimensions;

[0016] S12. Perform scattering artifact suppression processing on each radiotherapy image frame in the low-contrast radiotherapy image frame set. A frequency-domain occlusion compensation model is constructed to estimate the low-frequency component of each radiotherapy image frame. The low-frequency component estimation result is subtracted from the original radiotherapy image as a scattering term to generate a scatter-suppressed low-contrast radiotherapy image frame. The scattering term is used to represent the grayscale interference area caused by scattering in the radiotherapy image.

[0017] S13. Performing metal artifact correction on each frame of the radiotherapy image after scatter suppression processing, extracting metal regions to form a binary metal mask matrix, performing an interpolation and reconstruction operation based on texture features of non-metal neighborhoods within the binary metal mask matrix, and outputting a radiotherapy image frame after metal artifact correction. The radiotherapy image frame after metal artifact correction consists of retaining original pixel values ​​in non-metal regions of the original image and restoring pixel values ​​in metal regions through texture interpolation;

[0018] S14. Perform dose normalization on the radiotherapy image frames after metal artifact correction. This process extracts the maximum grayscale value of each radiotherapy image frame as an intensity benchmark, normalizes the grayscale values ​​of all pixels in the radiotherapy image frame according to this benchmark, and uniformly maps them to the set dose reference value, ensuring consistency in intensity across all radiotherapy image frames.

[0019] S15. Reorganize all radiotherapy image frames that have undergone scattering artifact suppression processing, metal artifact correction processing, and dose normalization processing to form a pre-processed low-contrast radiotherapy image sequence data set.

[0020] Optionally, the S2 includes the following steps:

[0021] S21. Construct an improved MoCov3 self-supervised feature encoder model, which includes a multi-scale tumor edge-aware encoder and a momentum-driven soft tissue structure consistency enhancement module, and initialize the network weights of the multi-scale tumor edge-aware encoder and the momentum-driven soft tissue structure consistency enhancement module;

[0022] S22. Construct an improved key-query pair based on the preprocessed low-contrast radiotherapy image sequence data set according to the perspective difference. The improved key-query pair is defined as:

[0023]

[0024] in, represents the query image with the purpose of enhancing the features of the tumor edge area, represents the key image for the purpose of highlighting the internal structural information of soft tissue, and N is the total number of improved key-query pairs;

[0025] S23. Input the improved key-query pair into the multi-scale tumor edge-aware encoder to obtain a multi-scale fused tumor edge-aware feature representation vector Enhanced feature representation vector with soft tissue structure The vectors have the same length and dimension;

[0026] S24. Based on the momentum-driven soft tissue structure consistency enhancement module, a soft structure constraint function is constructed between the tumor edge perception feature representation vector and the soft tissue structure enhancement feature representation vector in the feature space. Adaptive differentiation and enhancement of tumor margins and soft tissue structures in the feature space are achieved. The soft structure constraint function is defined as:

[0027]

[0028] in, represents the dot product similarity between the multi-scale tumor edge perception feature representation vector and the soft tissue structure enhancement feature representation vector, It represents the adaptive feature boundary enhancement item between the tumor edge area and the soft tissue structure area in the feature space, which is used to highlight the difference between the tumor edge area and the soft tissue structure area in the feature space. α is the feature boundary enhancement factor, which is used to adjust the weight of the feature boundary enhancement item. τ is the contrast learning temperature parameter, which is used to adjust the similarity scale. represents the jth negative sample feature vector, M is the total number of negative sample feature vectors;

[0029] S25. According to the soft structure constraint function The calculation results are used to update the network weights of the multi-scale tumor edge perception encoder and the momentum-driven soft tissue structure consistency enhancement module, and steps S23 to S25 are iteratively performed until the improved MoCov3 self-supervised feature encoder model converges;

[0030] S26. Use the trained improved MoCov3 self-supervised feature encoder model to perform feature inference on the preprocessed low-contrast radiotherapy image sequence data set, extract multi-scale fusion features with high discrimination, and form an initial feature representation set.

[0031] Optionally, the S21 includes the following steps:

[0032] S211. Construct a multi-scale tumor edge-aware encoder. The multi-scale tumor edge-aware encoder includes a backbone feature extraction network and a multi-scale edge attention module. The backbone feature extraction network performs convolution operations on each dose-normalized low-contrast radiotherapy image frame at multiple scales to extract a basic texture feature map. The basic texture feature map is used to represent the structural texture information of the image under different receptive fields. The spatial size of the basic texture feature map is consistent with the two-dimensional size of the input image frame.

[0033] S212. Construct an edge activation function based on the basic texture feature map at each scale to extract the tumor edge region. The edge activation function calculates the horizontal and vertical gradient maps of the basic texture feature map, superimposes the two gradient maps, and inputs them into a normalization function to obtain an edge response map. The edge response map is then multiplied point by point with the original basic texture feature map to obtain an edge-perceived feature map. The edge-perceived feature map is used to enhance the response of the edge region in the image, thereby improving the discernibility of the tumor boundary.

[0034] S213. Concatenate the edge-aware feature maps at all scales according to the channel dimension, and input the concatenated feature maps into a feature fusion network for interactive fusion to form a fused multi-scale tumor edge-aware feature representation vector. The fused multi-scale tumor edge-aware feature representation vector serves as a query feature vector in feature comparison calculation to capture multi-scale expression differences of the tumor edge in the feature space.

[0035] S214. Construct a momentum-driven soft tissue structural consistency enhancement module. The momentum-driven soft tissue structural consistency enhancement module consists of a structure-aware main encoder and a sliding momentum update encoder. The structure-aware main encoder and the tumor edge perception pathway share a shallow network structure. In the deep layer, an attention mechanism is embedded to enhance the texture representation capability of low-contrast regions.

[0036] S215. Initialize the parameters of the sliding momentum update encoder. The initial parameter values ​​are set to the current parameter values ​​of the structure-aware main encoder. In each subsequent training iteration, the parameters of the momentum encoder are updated using a sliding average. The sliding average update is obtained by setting a momentum update coefficient and proportionally weighted fusion of the current momentum encoder parameters and the parameters of the structure-aware main encoder in the current training round to ensure consistency in soft tissue structure representation and stability in the feature space.

[0037] S216. Input the key image into the structure-aware main encoder and the sliding momentum update encoder to extract the structure-enhanced feature representation vector and the momentum reference feature representation vector, respectively. The structure-enhanced feature representation vector is used for subsequent feature similarity calculation between the participating and query feature vectors, and the momentum reference feature representation vector is used to maintain encoder stability and serve as a reference for the negative sample queue.

[0038] S217. Complete the construction and parameter initialization of the backbone feature extraction network, the multi-scale edge attention module, and the momentum-driven soft tissue structure consistency enhancement module, and output the query feature representation vector and key feature representation vector required for the first round of training.

[0039] Optionally, the discrete coding is obtained by constructing a discretization reconstruction-coding integrated hyperparameter dictionary, which includes the following ten discrete hyperparameters: super-resolution convolution kernel index Regularization weight level Comparative learning temperature levels Filter kernel index Sampling mask index Feature boundary enhancement factor Momentum coefficient Multi-scale branch depth Negative sample queue length Edge-Structure Loss Weight Set the minimum and maximum values ​​for each discrete hyperparameter to form a corresponding set of discrete value intervals.

[0040] Optionally, S3 includes the following steps:

[0041] S31. Use the vector quantization mapping function to represent each feature vector f in the initial feature representation set i Mapped to the nearest neighbor feature center index c i , get the discrete feature index set C, each index in the discrete feature index set is in a unified integer dimension and corresponds to the same feature dictionary size;

[0042] S32. Set the number of gray wolves G and the total number of discrete hyperparameter dimensions D p = 10, generate a random number for each pair (g, j) and map it to integer coordinates according to the corresponding discrete value interval, and then obtain the wolf pack position matrix X (0) The number of rows in the wolf pack position matrix is ​​equal to the number of gray wolves, and the number of columns is equal to the total number of discrete hyperparameter dimensions. Each integer coordinate in the wolf pack position matrix maintains the same dimension corresponding to the minimum and maximum values ​​of its discrete hyperparameters.

[0043] S33. Using one-to-one decoding function Φ -1 : The wolf pack position matrix X (0) Each row of the integer coordinate vector decodes to a complete discrete hyperparameter combination And gather to get the initial wolf pack parameter set Θ( 0), each discrete hyperparameter combination in the initial wolf pack parameter set contains both the reconstruction hyperparameters and the internal hyperparameters of the improved MoCov3 encoder and maintains consistency in symbols and semantics;

[0044] S34. For the initial wolf pack parameter set Θ (0) Each discrete hyperparameter combination in Dynamically instantiate a set of improved MoCov3 self-supervised feature encoder instances E with the feature boundary enhancement factor, momentum coefficient, multi-scale branch depth, negative sample queue length, edge-structure loss weight, and contrastive learning temperature level g , the discrete hyperparameter combination The super-resolution convolution kernel index, filter kernel index, sampling mask index and regularization weight level in the image enhancement reconstruction operator R are synchronously injected into the image enhancement reconstruction operator R, thereby converting the initial feature representation set F init With encoder instance E g and discrete hyperparameter combinations Coupled image enhancement reconstruction and output of the corresponding reconstructed image

[0045] S35. Collect all reconstructed images Forming a set of image enhancement reconstruction results

[0046] Optionally, the S4 includes the following steps:

[0047] S41. Image enhancement and reconstruction result set Each reconstructed image Calculate the tumor edge clarity index E g Dose consistency index D g , tumor edge clarity index E g The dose consistency index D is obtained by calculating the mean gradient amplitude of the tumor edge area in the reconstructed image. The mean gradient amplitude is used to quantify the clarity of the spatial details of the tumor edge in the reconstructed image. g The normalized mean square error (NMEE) is obtained by calculating the normalized mean square error (NMEE) between the grayscale value distribution of the tumor area in the reconstructed image and the target dose distribution set in the radiotherapy plan. The NMEE is used to measure the accuracy of the reconstructed image in terms of dose.

[0048] S42. Constructing a dual-index fusion fitness function F for tumor edge clarity and dose consistency g , and to meet the clinical needs of low-contrast radiotherapy image enhancement reconstruction, a nonlinear balance factor w is introduced balance The weight relationship between the two indicators is adaptively adjusted so that the fitness value of each individual gray wolf takes into account both structural clarity and dosage accuracy:

[0049] Fg =w balance ·E g +(1-w balance )·(1-D g );

[0050] Nonlinear balance factor w balance The value of is dynamically adjusted with the changing trend of the tumor edge clarity index and the dose consistency index of the gray wolf individual during the iteration process. Specifically, when the tumor edge clarity index E g The dose consistency index D g If there is still a large fluctuation, w will be automatically reduced. balance The value of w is increased, so that the optimization focuses more on dose consistency; on the contrary, when the dose consistency index tends to be stable and the tumor edge clarity index still fluctuates greatly, w is automatically increased. balance The value of makes the optimization focus more on the clarity of tumor edges and improves the clinical adaptability of gray wolf individuals in the process of image enhancement reconstruction;

[0051] S43. Calculate the individual fitness values ​​of gray wolves using a dual-index fusion fitness function. Divide the wolf pack into two independent subgroups based on the composition ratio of the individual fitness values: an image structure optimization wolf pack and a dose accuracy optimization wolf pack. Wolves whose tumor edge clarity index accounts for a dominant proportion of their individual fitness values ​​are assigned to the image structure optimization wolf pack to improve the tumor edge clarity index. Wolves whose dose consistency index accounts for a dominant proportion of their individual fitness values ​​are assigned to the dose accuracy optimization wolf pack to improve the dose consistency index. The optimal individual gray wolf positions in the two subgroups are determined as the structural leader and dose leader of the corresponding subgroups, replacing the guiding structure of the global optimal, suboptimal, and third-optimal individual gray wolf positions in the traditional discrete gray wolf optimization algorithm.

[0052] S44. In the image structure optimization wolf pack, define the structure guidance factor S guide Dynamically control the update strategy of each gray wolf individual moving towards the position of the structural leader wolf, the structural guidance factor S guide The value of is determined by the difference in tumor edge clarity between the gray wolf individual and the structural leader wolf. When the difference is greater than the threshold, the structural guidance factor S is increased. guide , strengthen the approach to the position of the structural leader, and reduce the structural guidance factor S when the gap is less than the threshold guide , maintain the diversity of search exploration;

[0053] S45. Define the dose guidance factor D in the dose accuracy optimization wolf group guide Dynamically control the update strategy of each gray wolf individual moving towards the position of the dose leader wolf, the dose guidance factor D guideThe value of is determined by the difference between the individual gray wolf and the dose leader wolf in the dose consistency index. When the difference is greater than the threshold, the dose guidance factor D is increased. guide , strengthen the approach to the dose leader position, and reduce the dose guidance factor D when the gap is less than the threshold guide , to maintain the diversity of search exploration;

[0054] S46. Introduce a structure-dose cross-group synergy mechanism. During each iteration, the overall fitness of the two subgroups of the wolf pack is improved based on the image structure optimization and the dose accuracy optimization. Some gray wolf individuals are swapped across subgroups. Specifically, if the overall fitness of a subgroup does not improve after multiple consecutive iterations, several gray wolf individuals in that subgroup are selected and swapped to adjust their positions to the search space of another subgroup. This optimizes the cross-group information sharing capability and overall convergence efficiency of the discrete gray wolf optimization algorithm in low-contrast radiotherapy image enhancement reconstruction.

[0055] S47. After each round of iteration, the updated results of the positions of all individual gray wolves in the two subgroups are reassembled and formed into an updated wolf pack position matrix, and a one-to-one decoding function consistent with S3 is used to decode the wolf pack position matrix into an updated wolf pack parameter set. The updated wolf pack parameter set is used for image enhancement reconstruction and the dynamic optimization process of the MoCov3 self-supervised feature encoder model.

[0056] Optionally, the S5 includes the following steps:

[0057] S51. Update each discrete hyperparameter combination in the wolf pack parameter set Dynamically configure the improved MoCov3 self-supervised feature encoder model by setting the operating parameters of the boundary separation module, momentum update module, multi-scale structure perception module, and queue management strategy in the encoder according to the feature boundary enhancement factor, momentum coefficient, multi-scale branch depth, negative sample queue length, edge-structure loss weight, and contrastive learning temperature parameters, and construct a self-supervised feature encoder instance in the current round

[0058] S52. Refresh the momentum encoder queue of the current round of self-supervised feature encoder. The momentum encoder queue is refreshed by the positive sample feature set participating in the training from the previous round. The positive sample feature set output by the current round encoder Extract feature representations and fuse them using an exponential sliding average method based on the momentum update coefficient to generate a new momentum encoder queue to maintain the stability of the negative sample distribution during model training.

[0059] S53. Use the self-supervised feature encoder instance configured in the current round Perform feature extraction on the pre-processed low-contrast radiotherapy image sequence data to obtain the updated feature representation set F (t) , the updated feature representation set maintains the same dimension and semantic distribution as the initial feature set, and has the characteristics of the current discrete hyperparameter combination The coding features such as structure perception, boundary enhancement and temperature control are determined by it;

[0060] S54. Update the feature representation set F extracted in the current round (t) Combined with the current discrete hyperparameters The images are input into the image enhancement reconstruction operator together to complete the image enhancement reconstruction operation. During the reconstruction process, the specific reconstruction path is configured according to the super-resolution convolution kernel index, filter kernel index, sampling mask index and regularization weight level in the discrete hyperparameter combination, and an enhanced reconstructed image corresponding to the discrete hyperparameter combination is output. The enhanced reconstructed image is named the enhanced reconstructed image of the current gray wolf individual in the current round.

[0061] S55. Enhanced reconstructed images output by all gray wolf individuals in the current round Arrange and summarize in order to construct a complete set of image enhancement and reconstruction results for this round The image enhancement reconstruction result set contains the reconstructed image output corresponding to all current discrete hyperparameter combinations;

[0062] S56. Using image enhancement to reconstruct the result set Tumor edge clarity index Dose consistency index And construct a dual-index fusion fitness function The dual-index fusion fitness function dynamically weights the structural clarity and dose accuracy to form the fitness vector of the current round. Based on this fitness vector, the image structure optimization wolf group and the dose accuracy optimization wolf group are re-divided into two subgroups. The structure guidance, dose guidance, adaptive factor adjustment and structure-dose cross-group coordination mechanism are respectively implemented in the two subgroups to update the position vector of each gray wolf individual, and finally generate a new round of wolf group position matrix X (t+1) and its corresponding parameter combination set Θ (t+1) ;

[0063] S57. Determine the current round fitness vector F (t) Whether the preset convergence threshold ∈ conv , if satisfied, then select the gray wolf individual with the largest value from the fitness vector and output its corresponding optimal hyperparameter combination θ (*) and the corresponding final high contrast radiotherapy enhanced image I (*)If not, the iteration count is updated to t+1, and S51 to S56 are repeated until the convergence condition is met.

[0064] Optionally, the S6 includes the following steps:

[0065] S61. Enhanced imaging in high-contrast radiotherapy I (*) Import the closed-loop quality evaluation module, which will load the optimal reconstruction parameter configuration θ synchronously (*) Initialize the final quality verification process with the patient's original DICOM-RT structure set and radiotherapy plan dose distribution;

[0066] S62. In the closed-loop quality evaluation module, the organ delineation consistency index based on contour overlap is used to align the tumor target area with the key organs at risk, and the high-contrast radiotherapy enhanced image I (*) The structure contours obtained by semi-automatic segmentation were compared with the corresponding contours of the clinical DICOM-RT structure set at the pixel level to obtain the organ outline consistency index Dice ROI ,Organ delineation consistency index is used to quantify the degree of spatial consistency between the segmented contours of high-contrast radiotherapy-enhanced images and the reference contours;

[0067] S63. In the closed-loop quality assessment module, the dose recalculation error index based on the recalculated dose volume error is used to verify the accuracy of the dose measurement, and the original beam parameters of the radiotherapy plan are mapped to the high-contrast radiotherapy enhanced image I. (*) The three-dimensional dose distribution is then recalculated to obtain the recalculated dose volume, and the dose recalculation error index Dose is obtained by differentiating the recalculated dose volume from the original planned dose volume at the voxel level. RMSE ,The dose recalculation error index is used to quantify the degree of deviation in the dosimetry of high-contrast radiotherapy-enhanced images;

[0068] S64. Based on the clinical threshold set {τ Dice ,τ Dose Dice index for organ delineation consistency ROI and dose recalculation error index Dose RMSE Perform joint judgment:

[0069] When organ delineation consistency index Dice ROI ≥τ Dice And the dose recalculation error index Dose RMSE ≤τ Dose When the final quality verification is passed;

[0070] When any of the above conditions is not met, the final quality verification is deemed to have failed;

[0071] S65. If the final quality verification is passed, the high-contrast radiotherapy image result I that has passed the final quality verification is output. (*) and the corresponding optimal reconstruction parameter configuration θ (*) ;

[0072] S66. If the final quality verification fails, the organ will be outlined with the consistency index Dice ROI and dose recalculation error index Dose RMSE It is converted into a negative feedback gain value in proportion and injected into the wolf pack fitness vector of the previous round, correspondingly reducing the fitness contribution in the indicator direction that does not meet the threshold. Then it returns to S41 and re-executes the iterative process of image enhancement reconstruction and wolf pack position update until the joint threshold condition is met.

[0073] The beneficial effects of the present invention are:

[0074] (1) The present invention introduces an improved MoCov3 self-supervised feature encoder model, and constructs a separation perception mechanism for tumor edges and soft tissue textures through feature boundary enhancement factors, multi-scale branch structures and momentum consistency queues. It can accurately capture weak structural features in low-contrast areas in the absence of pixel-level labels. By introducing a structure-boundary separation module and a soft structure consistency enhancement mechanism, a structural contrast learning path for radiotherapy sequences is designed, which significantly improves the encoder's responsiveness to tumor boundaries.

[0075] (2) This paper designs a structure-dose dual-guided discrete gray wolf optimization framework. By dividing the gray wolf group into an imaging structure optimization subgroup and a dose consistency optimization subgroup, the structure guidance factor and the dose guidance factor are introduced for differential iteration, and supplemented by a structure-dose cross-group synergy mechanism, the convergence speed and global search capability of the optimization strategy in multi-objective reconstruction scenarios are effectively improved. Starting from the clinical indicators of radiotherapy, different subgroups are guided to optimize around specific goals, and a dynamic exchange mechanism is introduced between subgroups to avoid falling into local optimality.

[0076] (3) This invention establishes a cyclically consistent closed-loop feedback mechanism for image enhancement and quality verification. By introducing a joint quality assessment standard based on the Dice coefficient and dose-volume error, the final verification result is injected into the fitness function as a negative gain, thereby enhancing the sensitivity to clinical usability indicators and target consistency during the parameter search process. The introduction of closed-loop evaluation and dynamic feedback in each iteration allows the algorithm to continuously adjust the reconstruction parameters to converge to dose consistency while meeting structural clarity, ensuring the practicality and reliability of images in treatment path planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0078] Figure 1 This is a flowchart of a radiotherapy image reconstruction method based on self-supervised learning proposed by the present invention;

[0079] Figure 2 This is a schematic diagram of the structure of the improved MoCov3 self-supervised feature encoder model in the radiotherapy image reconstruction method based on self-supervised learning proposed in the present invention. DETAILED DESCRIPTION

[0080] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0081] refer to Figure 1-Figure 2 , a radiotherapy image reconstruction method based on self-supervised learning, comprising the following steps:

[0082] S1. Acquire low-contrast radiotherapy image sequence data, perform scattering artifact suppression, metal artifact correction, and dose normalization preprocessing on the low-contrast radiotherapy image sequence data, and obtain a low-contrast radiotherapy image sequence data set;

[0083] S2. Construct a MoCov3 self-supervised feature encoder model and initialize the momentum encoder weights. Construct the low-contrast radiotherapy image sequence dataset into key-query pairs based on viewpoint differences and input them into the MoCov3 self-supervised feature encoder model. Perform self-supervised contrastive learning and obtain the MoCov3 self-supervised feature encoder model. Use the MoCov3 self-supervised feature encoder model to perform inference on the low-contrast radiotherapy image sequence dataset to obtain an initial feature representation set.

[0084] S3. Initialize the wolf pack position vectors of the discrete gray wolf optimization algorithm, uniformly discretize the initial feature representation set to form an initial wolf pack parameter set, and for each wolf pack position vector in the initial wolf pack parameter set, couple the corresponding parameter combination with the initial feature representation set, perform image enhancement reconstruction, and obtain an image enhancement reconstruction result set;

[0085] S4. Calculate the tumor edge clarity index and dose consistency index based on the image enhancement reconstruction result set. Fusion of these indexes yields the wolf pack fitness vector. Based on the wolf pack fitness vector, the discrete gray wolf optimization algorithm position update strategy, dynamic step size control, and prey escape mechanism are used to update the wolf pack position vector and generate an updated wolf pack parameter set.

[0086] S5. driving the MoCov3 self-supervised feature encoder model with the updated wolf pack parameter set, synchronously adjusting the contrast learning temperature parameter and refreshing the momentum encoder queue, obtaining an updated feature representation set, and re-executing the image enhancement reconstruction using the updated feature representation set and the updated wolf pack parameter set to obtain an iterative image enhancement reconstruction result set, repeating steps S4 to S5 until the wolf pack fitness vector meets the preset convergence condition, and outputting the optimal reconstruction parameter configuration and the corresponding high-contrast radiotherapy enhancement image;

[0087] S6. inputting the high-contrast radiotherapy enhancement image into the closed-loop quality evaluation module, performing final quality verification using the organ contour consistency index and dose recalculation error, if the final quality verification result meets the clinical threshold, outputting the high-contrast radiotherapy image result after final quality verification and the optimal reconstruction parameter configuration, and storing the optimal reconstruction parameter configuration for subsequent treatment day image enhancement reconstruction initialization, if the final quality verification result does not meet the clinical threshold, feeding back the final quality verification result to the wolf pack fitness vector and returning to step S4 for iteration until the clinical threshold is met.

[0088] In this embodiment, S1 includes the following steps:

[0089] S11. acquiring a low-contrast radiotherapy image sequence data in a preset time window by a cone beam computed tomography system to form a low-contrast radiotherapy image frame set, each frame of the low-contrast radiotherapy image frame set being a two-dimensional image with uniform height and width dimensions;

[0090] S12. performing scatter artifact suppression processing on each frame of the low-contrast radiotherapy image frame set, estimating a low-frequency component of each frame of the radiotherapy image by constructing a frequency domain occlusion compensation model, and subtracting the low-frequency component estimation result as a scatter term from the original radiotherapy image to generate a low-contrast radiotherapy image frame after scatter suppression, the scatter term being used to represent a gray scale interference area caused by scatter in the radiotherapy image;

[0091] S13. performing metal artifact correction processing on each frame of the radiotherapy image after scatter suppression processing, extracting a metal region to form a binary metal mask matrix, and performing an interpolation reconstruction operation based on the texture features of the non-metal neighborhood in the binary metal mask matrix to output a radiotherapy image frame after metal artifact correction processing, the radiotherapy image frame after metal artifact correction processing being composed of original pixel values in non-metal regions and restored pixel values by texture interpolation in metal regions;

[0092] S14. Perform dose normalization on the radiotherapy image frames after metal artifact correction. This process extracts the maximum grayscale value of each radiotherapy image frame as an intensity benchmark, normalizes the grayscale values ​​of all pixels in the radiotherapy image frame according to this benchmark, and uniformly maps them to the set dose reference value, ensuring consistency in intensity across all radiotherapy image frames.

[0093] S15. Reorganize all radiotherapy image frames that have undergone scattering artifact suppression processing, metal artifact correction processing, and dose normalization processing to form a pre-processed low-contrast radiotherapy image sequence data set.

[0094] In this embodiment, S2 includes the following steps:

[0095] S21. Construct an improved MoCov3 self-supervised feature encoder model, which includes a multi-scale tumor edge-aware encoder and a momentum-driven soft tissue structure consistency enhancement module, and initialize the network weights of the multi-scale tumor edge-aware encoder and the momentum-driven soft tissue structure consistency enhancement module;

[0096] S22. Construct an improved key-query pair based on the preprocessed low-contrast radiotherapy image sequence data set according to the perspective difference. The improved key-query pair is defined as:

[0097]

[0098] in, represents the query image with the purpose of enhancing the features of the tumor edge area, represents the key image for the purpose of highlighting the internal structural information of soft tissue, and N is the total number of improved key-query pairs;

[0099] S23. Input the improved key-query pair into the multi-scale tumor edge-aware encoder to obtain a multi-scale fused tumor edge-aware feature representation vector Enhanced feature representation vector with soft tissue structure The vectors have the same length and dimension;

[0100] S24. Based on the momentum-driven soft tissue structure consistency enhancement module, a soft structure constraint function is constructed between the tumor edge perception feature representation vector and the soft tissue structure enhancement feature representation vector in the feature space. To achieve adaptive differentiation and enhancement of tumor margins and soft tissue structures in the feature space, the soft structure constraint function is defined as:

[0101]

[0102] in, represents the dot product similarity between the multi-scale tumor edge perception feature representation vector and the soft tissue structure enhancement feature representation vector, It represents the adaptive feature boundary enhancement item between the tumor edge area and the soft tissue structure area in the feature space, which is used to highlight the difference between the tumor edge area and the soft tissue structure area in the feature space. α is the feature boundary enhancement factor, which is used to adjust the weight of the feature boundary enhancement item. τ is the contrast learning temperature parameter, which is used to adjust the similarity scale. represents the jth negative sample feature vector, M is the total number of negative sample feature vectors;

[0103] By introducing the adaptive feature boundary enhancement term The encoder prioritizes the representation of tumor margins, improving its ability to discern soft tissue differences without labels. This prevents the modified MoCov3 self-supervised feature encoder model from experiencing modal collapse (feature vectors converge to the mean) under low-contrast inputs, improving stability in radiotherapy images. By incorporating structural saliency as a constraint, the modified MoCov3 self-supervised feature encoder model can adapt to different patient anatomy and tumor types, improving generalization of heterogeneous structures.

[0104] S25. According to the soft structure constraint function The calculation results are used to update the network weights of the multi-scale tumor edge perception encoder and the momentum-driven soft tissue structure consistency enhancement module, and steps S23 to S25 are iteratively performed until the improved MoCov3 self-supervised feature encoder model converges;

[0105] S26. Use the trained improved MoCov3 self-supervised feature encoder model to perform feature inference on the preprocessed low-contrast radiotherapy image sequence data set, extract multi-scale fusion features with high discrimination, and form an initial feature representation set.

[0106] In this embodiment, S21 includes the following steps:

[0107] S211. Construct a multi-scale tumor edge-aware encoder. The multi-scale tumor edge-aware encoder includes a backbone feature extraction network and a multi-scale edge attention module. The backbone feature extraction network performs convolution operations on each dose-normalized low-contrast radiotherapy image frame at multiple scales to extract a basic texture feature map. The basic texture feature map is used to represent the structural texture information of the image under different receptive fields. The spatial size of the basic texture feature map is consistent with the two-dimensional size of the input image frame.

[0108] S212. Construct an edge activation function based on the basic texture feature map at each scale to extract the tumor edge region. The edge activation function calculates the horizontal and vertical gradient maps of the basic texture feature map, superimposes the two gradient maps, and inputs them into a normalization function to obtain an edge response map. The edge response map is then multiplied point by point with the original basic texture feature map to obtain an edge-perceived feature map. The edge-perceived feature map is used to enhance the response of the edge region in the image, thereby improving the discernibility of the tumor boundary.

[0109] S213. Concatenate the edge-aware feature maps at all scales according to the channel dimension, and input the concatenated feature maps into a feature fusion network for interactive fusion to form a fused multi-scale tumor edge-aware feature representation vector. The fused multi-scale tumor edge-aware feature representation vector serves as a query feature vector in feature comparison calculation to capture multi-scale expression differences of the tumor edge in the feature space.

[0110] S214. Construct a momentum-driven soft tissue structural consistency enhancement module. The momentum-driven soft tissue structural consistency enhancement module consists of a structure-aware main encoder and a sliding momentum update encoder. The structure-aware main encoder and the tumor edge perception pathway share a shallow network structure. In the deep layer, an attention mechanism is embedded to enhance the texture representation capability of low-contrast regions.

[0111] S215. Initialize the parameters of the sliding momentum update encoder. The initial parameter values ​​are set to the current parameter values ​​of the structure-aware main encoder. In each subsequent training iteration, the parameters of the momentum encoder are updated using a sliding average. The sliding average update is obtained by setting a momentum update coefficient and proportionally weighted fusion of the current momentum encoder parameters and the parameters of the structure-aware main encoder in the current training round to ensure consistency in soft tissue structure representation and stability in the feature space.

[0112] S216. Input the key image into the structure-aware main encoder and the sliding momentum update encoder to extract the structure-enhanced feature representation vector and the momentum reference feature representation vector, respectively. The structure-enhanced feature representation vector is used for subsequent feature similarity calculation between the participating and query feature vectors, and the momentum reference feature representation vector is used to maintain encoder stability and serve as a reference for the negative sample queue.

[0113] S217. Complete the construction and parameter initialization of the backbone feature extraction network, the multi-scale edge attention module, and the momentum-driven soft tissue structure consistency enhancement module, and output the query feature representation vector and key feature representation vector required for the first round of training.

[0114] In this embodiment, the discrete coding is obtained by constructing a discretization reconstruction-coding integrated hyperparameter dictionary, which includes the following ten discrete hyperparameters: super-resolution convolution kernel index Regularization weight level Comparative learning temperature levels Filter kernel index Sampling mask index Feature boundary enhancement factor Momentum coefficient Multi-scale branch depth Negative sample queue length Edge-Structure Loss Weight Set the minimum and maximum values ​​for each discrete hyperparameter to form a corresponding set of discrete value intervals.

[0115] In this embodiment, S3 includes the following steps:

[0116] S31. Use the vector quantization mapping function to represent each feature vector f in the initial feature representation set i Mapped to the nearest neighbor feature center index c i , get the discrete feature index set C, each index in the discrete feature index set is in a unified integer dimension and corresponds to the same feature dictionary size;

[0117] The vector quantization mapping function is used to map each initial feature representation vector to a discrete feature index, which represents its nearest center in the predefined feature dictionary. The specific construction method is as follows:

[0118] Construct a feature center dictionary: Sampling or clustering is used to extract a set of representative feature center vectors from the initial feature representation set. All feature center vectors form a feature center dictionary. Each feature center in the feature center dictionary has the same dimension as the initial feature vector.

[0119] Define the distance metric: Euclidean distance is used as the distance metric to evaluate the similarity between any feature representation vector and all feature centers in the dictionary. Euclidean distance is used to ensure that the mapping is performed in a unified numerical dimension.

[0120] Perform nearest neighbor matching: For each feature vector in the initial feature representation set, calculate its Euclidean distance to all feature centers, and select the feature center with the smallest distance as the target index mapped by the vector. The obtained index value is the discretized code of the feature representation.

[0121] Construct a discrete feature index set: All initial feature representation vectors are converted into corresponding feature index values ​​through the above-mentioned nearest neighbor matching process, and finally a discrete feature index set in a unified format is formed for the subsequent coupling of wolf pack parameter positions and feature space.

[0122] S32. Set the number of gray wolves G and the total number of discrete hyperparameter dimensions D p = 10, generate a random number for each pair (g, j) and map it to integer coordinates according to the corresponding discrete value interval, and then obtain the wolf pack position matrix X (0) The number of rows in the wolf pack position matrix is ​​equal to the number of gray wolves, and the number of columns is equal to the total number of discrete hyperparameter dimensions. Each integer coordinate in the wolf pack position matrix maintains the same dimension corresponding to the minimum and maximum values ​​of its discrete hyperparameters.

[0123] S33. Using one-to-one decoding function Φ -1 : The wolf pack position matrix X (0) Each row of the integer coordinate vector decodes to a complete discrete hyperparameter combination And gather to get the initial wolf pack parameter set Θ (0) , each discrete hyperparameter combination in the initial wolf pack parameter set contains both the reconstruction hyperparameters and the internal hyperparameters of the improved MoCov3 encoder and maintains consistency in symbols and semantics;

[0124] S34. For the initial wolf pack parameter set Θ( 0) Each discrete hyperparameter combination in Dynamically instantiate a set of improved MoCov3 self-supervised feature encoder instances E with the feature boundary enhancement factor, momentum coefficient, multi-scale branch depth, negative sample queue length, edge-structure loss weight, and contrastive learning temperature level g , the discrete hyperparameter combination The super-resolution convolution kernel index, filter kernel index, sampling mask index and regularization weight level in the image enhancement reconstruction operator R are synchronously injected into the image enhancement reconstruction operator R, thereby converting the initial feature representation set F init With encoder instance E g and discrete hyperparameter combinations Coupled image enhancement reconstruction and output of the corresponding reconstructed image

[0125] S35. Collect all reconstructed images Forming a set of image enhancement reconstruction results

[0126] In this embodiment, S4 includes the following steps:

[0127] S41. Image enhancement and reconstruction result set Each reconstructed image Calculate the tumor edge clarity index E g Dose consistency index D g , tumor edge clarity index E g The dose consistency index D is obtained by calculating the mean gradient amplitude of the tumor edge area in the reconstructed image. The mean gradient amplitude is used to quantify the clarity of the spatial details of the tumor edge in the reconstructed image. g The normalized mean square error (NMEE) is obtained by calculating the normalized mean square error (NMEE) between the grayscale value distribution of the tumor area in the reconstructed image and the target dose distribution set in the radiotherapy plan. The NMEE is used to measure the accuracy of the reconstructed image in terms of dose.

[0128] S42. Constructing a dual-index fusion fitness function F for tumor edge clarity and dose consistency g , and to meet the clinical needs of low-contrast radiotherapy image enhancement reconstruction, a nonlinear balance factor w is introduced balance The weight relationship between the two indicators is adaptively adjusted so that the fitness value of each individual gray wolf takes into account both structural clarity and dosage accuracy:

[0129] F g =w balance ·E g +(1-w balance )·(1-D g );

[0130] Nonlinear balance factor w balance The value of is dynamically adjusted with the changing trend of the tumor edge clarity index and the dose consistency index of the gray wolf individual during the iteration process. Specifically, when the tumor edge clarity index E g The dose consistency index D g If there is still a large fluctuation, w will be automatically reduced. balance The value of w is increased, so that the optimization focuses more on dose consistency; on the contrary, when the dose consistency index tends to be stable and the tumor edge clarity index still fluctuates greatly, w is automatically increased. balance The value of makes the optimization focus more on the clarity of tumor edges and improves the clinical adaptability of gray wolf individuals in the process of image enhancement reconstruction;

[0131] The dual-index fusion fitness function breaks through the limitations of traditional image reconstruction optimization objectives, which are mostly structural indicators or single dose matching items. It integrates the two clinical goals of structural quality and dose availability of radiotherapy image reconstruction into the evolutionary optimization model with an adjustable dynamic mechanism.

[0132] Simultaneously optimize both structure and dose: This solves the problem of structure optimization sacrificing dosimetric accuracy in previous greedy parameter adjustment or single-target search, improving clinical adaptability.

[0133] Adaptive adjustment of search direction: through nonlinear balance factor w balance Dynamically adjust the current optimization bias so that the discrete grey wolf search can be adaptively guided under the changes in the dominance of different indicators, avoiding falling into a single solution domain.

[0134] S43. Calculate the individual fitness values ​​of gray wolves using a dual-index fusion fitness function. Divide the wolf pack into two independent subgroups based on the composition ratio of the individual fitness values: an image structure optimization wolf pack and a dose accuracy optimization wolf pack. Wolves whose tumor edge clarity index accounts for a dominant proportion of their individual fitness values ​​are assigned to the image structure optimization wolf pack to improve the tumor edge clarity index. Wolves whose dose consistency index accounts for a dominant proportion of their individual fitness values ​​are assigned to the dose accuracy optimization wolf pack to improve the dose consistency index. The optimal individual gray wolf positions in the two subgroups are determined as the structural leader and dose leader of the corresponding subgroups, replacing the guiding structure of the global optimal, suboptimal, and third-optimal individual gray wolf positions in the traditional discrete gray wolf optimization algorithm.

[0135] S44. In the image structure optimization wolf pack, define the structure guidance factor S guide Dynamically control the update strategy of each gray wolf individual moving towards the position of the structural leader wolf, the structural guidance factor S guide The value of is determined by the difference in tumor edge clarity between the gray wolf individual and the structural leader wolf. When the difference is greater than the threshold, the structural guidance factor S is increased. guide , strengthen the approach to the position of the structural leader, and reduce the structural guidance factor S when the gap is less than the threshold guide , maintain the diversity of search exploration;

[0136] S45. Define the dose guidance factor D in the dose accuracy optimization wolf group guide Dynamically control the update strategy of each gray wolf individual moving towards the position of the dose leader wolf, the dose guidance factor D guide The value of is determined by the difference between the individual gray wolf and the dose leader wolf in the dose consistency index. When the difference is greater than the threshold, the dose guidance factor D is increased. guide , strengthen the approach to the dose leader position, and reduce the dose guidance factor D when the gap is less than the threshold guide , to maintain the diversity of search exploration;

[0137] S46. Introduce a structure-dose cross-group synergy mechanism. During each iteration, the overall fitness of the two subgroups of the wolf pack is improved based on the image structure optimization and the dose accuracy optimization. Some gray wolf individuals are swapped across subgroups. Specifically, if the overall fitness of a subgroup does not improve after multiple consecutive iterations, several gray wolf individuals in that subgroup are selected and swapped to adjust their positions to the search space of another subgroup. This optimizes the cross-group information sharing capability and overall convergence efficiency of the discrete gray wolf optimization algorithm in low-contrast radiotherapy image enhancement reconstruction.

[0138] S47. After each round of iteration, the updated results of the positions of all individual gray wolves in the two subgroups are reassembled and formed into an updated wolf pack position matrix, and a one-to-one decoding function consistent with S3 is used to decode the wolf pack position matrix into an updated wolf pack parameter set. The updated wolf pack parameter set is used for image enhancement reconstruction and the dynamic optimization process of the MoCov3 self-supervised feature encoder model.

[0139] In this embodiment, S5 includes the following steps:

[0140] S51. Update each discrete hyperparameter combination in the wolf pack parameter set Dynamically configure the improved MoCov3 self-supervised feature encoder model by setting the operating parameters of the boundary separation module, momentum update module, multi-scale structure perception module, and queue management strategy in the encoder according to the feature boundary enhancement factor, momentum coefficient, multi-scale branch depth, negative sample queue length, edge-structure loss weight, and contrastive learning temperature parameters, and construct a self-supervised feature encoder instance in the current round

[0141] S52. Refresh the momentum encoder queue of the current round of self-supervised feature encoder. The momentum encoder queue is refreshed by the positive sample feature set participating in the training from the previous round. The positive sample feature set output by the current round encoder Extract feature representations and fuse them using an exponential sliding average method based on the momentum update coefficient to generate a new momentum encoder queue to maintain the stability of the negative sample distribution during model training.

[0142] S53. Use the self-supervised feature encoder instance configured in the current round Perform feature extraction on the pre-processed low-contrast radiotherapy image sequence data to obtain the updated feature representation set F (t) , the updated feature representation set maintains the same dimension and semantic distribution as the initial feature set, and has the characteristics of the current discrete hyperparameter combination The coding features such as structure perception, boundary enhancement and temperature control are determined by it;

[0143] S54. The set of updated feature representations F extracted in the current round (t) combined with the current discrete hyperparameter combination are input into the image enhancement reconstruction operator to perform the image enhancement reconstruction operation. During the reconstruction process, a specific reconstruction path is configured according to the super-resolution convolution kernel index, the filter kernel index, the sampling mask index, and the regularization weight level in the discrete hyperparameter combination, and an enhanced reconstruction image corresponding to the discrete hyperparameter combination is output. The enhanced reconstruction image is named the enhanced reconstruction image of the current wolf individual in the current round

[0144] S55. The enhanced reconstruction images output by all wolf individuals in the current round are collected to form a complete image enhancement reconstruction result set for the current round The image enhancement reconstruction result set includes the reconstruction images output by the current discrete hyperparameter combination;

[0145] S56. The image enhancement reconstruction result set is used to calculate the tumor edge sharpness index and the dose consistency index and construct a double-index fusion fitness function The double-index fusion fitness function dynamically weights between structural clarity and dose accuracy to form a fitness vector for the current round. According to the fitness vector, the image structure optimization wolf group and the dose accuracy optimization wolf group are re-divided, and the structure-guided, dose-guided, adaptive factor adjustment, and structure-dose cross-group coordination mechanisms are executed in the two subgroups, respectively, to update the position vector of each wolf individual, and finally generate a new round of wolf group position matrix X (t+1) and the corresponding parameter combination set Θ (t+1)

[0146] S57. Determine whether the fitness vector F of the current round (t) satisfies the preset convergence threshold ∈ conv If it is satisfied, the wolf individual with the largest value is selected from the fitness vector, and the corresponding optimal hyperparameter combination θ (*) and the corresponding final high-contrast radiotherapy enhancement image I (*) are output; if it is not satisfied, the iteration count is updated to t+1, and steps S51 to S56 are repeated until the convergence condition is satisfied.

[0147] In this embodiment, S6 includes the following steps:

[0148] S61. The high-contrast radiotherapy enhancement image I (*) is imported into the closed-loop quality evaluation module, and the closed-loop quality evaluation module synchronously loads the optimal reconstruction parameter configuration θ​(*) Initialize the final quality verification process with the patient's original DICOM-RT structure set and radiotherapy plan dose distribution;

[0149] The closed-loop quality evaluation module consists of three sub-modules: the structural consistency assessment sub-module, the dose recalculation verification sub-module, and the result feedback control sub-module. Together, the three form a quality assessment and optimization feedback closed loop.

[0150] The structural consistency assessment submodule is used to evaluate the tumor area and organ contours in the enhanced radiotherapy images. It is constructed as follows: a multi-structure delineation algorithm based on deep morphological segmentation and boundary fusion is called to extract the contours of the tumor target area, spinal cord, lung, and liver key tissue areas in the enhanced images; then, automatic alignment and contour comparison methods are used to overlap the extracted contours with the reference contours in the original DICOM-RT structure set, and the organ delineation consistency index used to measure the structural spatial consistency is output.

[0151] The dose recalculation and verification submodule is used to recalculate the dose on the enhanced image and compare and verify it with the dose target in the radiotherapy plan. It is constructed as follows: the beam arrangement, energy level and irradiation parameters in the original plan are imported into the Monte Carlo or convolution / super-resolution dose calculation engine, and the enhanced image is used as the solution basis to complete the new dose field calculation; the new dose volume and the planned dose volume are then calculated at the voxel level, and the dose recalculation error index is output.

[0152] The result feedback control submodule is used to use the combined results of the structural consistency index and the dose error index as the optimization feedback input. It is constructed by jointly judging the above two indicators against the preset clinical threshold and generating a corresponding negative feedback factor based on the type and degree of deviation of the failed indicator. This feedback factor will be used to dynamically adjust the fitness evaluation function weight or optimization direction in the next round of discrete grey wolf optimization, thereby achieving a dynamic balance and iterative improvement between structural accuracy and dosimetric reliability.

[0153] The three submodules operate in series in a structure assessment-dose verification-feedback driven manner, forming an information closed loop between the enhanced reconstruction results and the optimization engine, and realizing closed-loop quality control of low-contrast radiotherapy images in both structure and dose dimensions.

[0154] S62. In the closed-loop quality evaluation module, the organ delineation consistency index based on contour overlap is used to align the tumor target area with the key organs at risk, and the high-contrast radiotherapy enhanced image I (*) The structure contours obtained by semi-automatic segmentation were compared with the corresponding contours of the clinical DICOM-RT structure set at the pixel level to obtain the organ outline consistency index Dice ROI,Organ delineation consistency index is used to quantify the degree of spatial consistency between the segmented contours of high-contrast radiotherapy-enhanced images and the reference contours;

[0155] S63. In the closed-loop quality assessment module, the dose recalculation error index based on the recalculated dose volume error is used to verify the accuracy of the dose measurement, and the original beam parameters of the radiotherapy plan are mapped to the high-contrast radiotherapy enhanced image I. (*) The three-dimensional dose distribution is then recalculated to obtain the recalculated dose volume, and the dose recalculation error index Dose is obtained by differentiating the recalculated dose volume from the original planned dose volume at the voxel level. RMSE ,The dose recalculation error index is used to quantify the degree of deviation in the dosimetry of high-contrast radiotherapy-enhanced images;

[0156] S64. Based on the clinical threshold set {τ Dice ,τ Dose Dice index for organ delineation consistency ROI and dose recalculation error index Dose RMSE Perform joint judgment:

[0157] When organ delineation consistency index Dice ROI ≥τ Dice And the dose recalculation error index Dose RMSE ≤τ Dose When the final quality verification is passed;

[0158] When any of the above conditions is not met, the final quality verification is deemed to have failed;

[0159] S65. If the final quality verification is passed, the high-contrast radiotherapy image result I that has passed the final quality verification is output. (*) and the corresponding optimal reconstruction parameter configuration θ (*) ;

[0160] S66. If the final quality verification fails, the organ will be outlined with the consistency index Dice ROI and dose recalculation error index Dose RMSE It is converted into a negative feedback gain value in proportion and injected into the wolf pack fitness vector of the previous round, correspondingly reducing the fitness contribution in the indicator direction that does not meet the threshold. Then it returns to S41 and re-executes the iterative process of image enhancement reconstruction and wolf pack position update until the joint threshold condition is met.

[0161] Example 1:

[0162] The Radiology Department of the Cancer Hospital in City A welcomed a 57-year-old male patient, surnamed Zhang, who was diagnosed with early-stage non-small cell lung cancer in the left lower lobe. To develop a precise stereotactic radiotherapy plan, the doctor arranged a routine CBCT scan for image positioning. However, the acquired images were limited by the scattering of the metal stent and the overlap of the soft tissue structure of the lungs, resulting in blurred tumor margins and severely low contrast in the dose setting area. As a result, the doctor made multiple adjustments during the image delineation process, but was still unable to clearly determine the boundary between the tumor and the adjacent bronchus.

[0163] In the traditional process, radiology technicians use standard bilateral filtering and gamma correction methods for image enhancement. Although the overall brightness of the image is improved, the tumor boundary still appears severely blurred in grayscale. Radiation oncologists are unable to determine the safe boundary based on this image, resulting in the need to arrange additional enhanced CT examinations, which prolongs the time it takes to formulate treatment plans and increases the patient's radiation exposure.

[0164] To solve this problem, the research team introduced the present invention and attempted to perform a new type of enhanced reconstruction processing on Zhang's CBCT image data to evaluate its practicality and stability in clinical radiotherapy positioning.

[0165] The research team first called up the original CBCT sequence obtained by Zhang, totaling 96 frames, with an image resolution of 512×512 pixels and a thickness of 1.25mm per frame. The standard deviation of the average CT value distribution in the original image was only 22.4, which was significantly lower than the tissue identifiable threshold of 30.

[0166] First, the original sequence was processed for scattering artifact suppression. The frequency-domain occlusion compensation model was used for low-frequency estimation, and the estimated grayscale was removed from the image as a scattering term. Subsequently, a binary metal mask was constructed for the metal artifact area, and the missing area was restored using interpolation based on texture similarity. Finally, in the dose normalization stage, the maximum grayscale value of each frame was used as the intensity benchmark and uniformly mapped to the standard dose reference value of 1.0.

[0167] The preprocessed data was input into the improved MoCov3 self-supervised feature encoder. In the process of constructing training sample pairs, six pairs were constructed with the 12th, 20th, 28th, 36th, 44th, and 52nd frames as the perspective difference reference frames. The query image emphasized the tumor boundary through the edge attention mechanism, while the key image retained the structural consistency characteristics of the junction area between the lung lobe and the blood vessels. The network used ResNet-50 as the backbone structure, the momentum coefficient was initially set to 0.996, and the edge-structure loss weighting factor was set to 0.65.

[0168] After initial training, the extracted multi-scale fusion feature map showed a significantly enhanced response in the lower left edge of the tumor, but structural adhesions still existed in the central bronchus region. To this end, the MoCov3 feature vector was fused with the discrete gray wolf optimization algorithm to initialize 100 gray wolf individuals. Each individual contained a 10-dimensional discrete hyperparameter combination, and the full stack participated in the search, from convolution kernel selection and sampling mask to contrast learning temperature and regularization level.

[0169] In the first round of reconstructed output images, the average tumor edge clarity index was 0.328, and the dose consistency index was 0.743. The gray wolf group was divided into a structure optimization subgroup and a dose optimization subgroup, which iteratively executed structure guidance and dose guidance strategies respectively. In the fifth round of iteration, the fitness was significantly improved, with the Edge index reaching 0.417 and the Dose index reaching 0.801. After the ninth round, the group converged, and the edge clarity in the final output reconstructed image was improved to 0.459, and the Dose index was 0.834, reaching the preset clinical standards of DiceROI ≥ 0.85 and DoseRMSE ≤ 0.17.

[0170] Table 1 Data comparison of the method of the present invention, the traditional filtering enhancement method and the original MoCov3 method

[0171]

[0172] To verify the comparative advantages of the proposed method, the research team conducted a real-world control experiment involving 20 low-contrast radiotherapy cases. They processed CBCT data using the traditional filter enhancement method, the original MoCov3 method, and the proposed method. The performance of each method in terms of edge structure recognition accuracy, dose consistency index, and model stability was statistically analyzed.

[0173] For example, after the final enhanced and reconstructed image entered the closed-loop quality assessment process, the automatically segmented tumor region achieved a DiceROI value of 0.869 and a dose recalculation error (DoseRMSE) of 0.153, both meeting and exceeding clinical reference thresholds. During re-evaluation and verification within the image-guided radiotherapy system, the reconstructed image was assessed by doctors as "featuring feasibility and boundary accuracy for tumor target replanning," significantly shortening treatment preparation cycles.

[0174] According to clinical feedback, it not only improves the clarity and reliability of tumor boundary identification, but also completes high-quality reconstruction without the need for additional enhanced CT acquisition, reducing patients' additional radiation risk and medical costs, and significantly improving the practical application value of low-contrast CBCT images in tumor radiotherapy.

[0175] This paper introduces an improved MoCov3 self-supervised feature encoder model, and constructs a separation perception mechanism for tumor edges and soft tissue textures through feature boundary enhancement factors, multi-scale branch structures and momentum consistency queues. It can accurately capture weak structural features in low-contrast areas in the absence of pixel-level labels. By introducing a structure-boundary separation module and a soft structure consistency enhancement mechanism, a structural contrast learning path for radiotherapy sequences is designed, which significantly improves the encoder's responsiveness to tumor boundaries.

[0176] This paper designs a structure-dose dual-guided discrete gray wolf optimization framework. By dividing the gray wolf population into an imaging structure optimization subgroup and a dose consistency optimization subgroup, the paper introduces structure-guiding factors and dose-guiding factors for differentiated iteration. This is supplemented by a structure-dose cross-group synergy mechanism, effectively improving the convergence speed and global search capabilities of the optimization strategy in multi-objective reconstruction scenarios. Starting from radiotherapy clinical indicators, the paper guides different subgroups to optimize around specific goals, while introducing a dynamic exchange mechanism between subgroups to avoid falling into local optimality.

[0177] This invention establishes a cyclically consistent closed-loop feedback mechanism for image enhancement and quality verification. By introducing a joint quality assessment criterion based on the Dice coefficient and dose-volume error, and injecting the final verification results as a negative gain into the fitness function, the sensitivity to clinical usability indicators and target consistency during the parameter search process are enhanced. The introduction of closed-loop evaluation and dynamic feedback in each iteration allows the algorithm to continuously adjust reconstruction parameters to converge toward dose consistency while maintaining structural clarity, ensuring the practicality and reliability of images in treatment path planning.

[0178] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A radiotherapy image reconstruction method based on self-supervised learning, characterized in that: The steps include: S1. Obtain low-contrast radiotherapy image sequence data, perform preprocessing, and obtain a low-contrast radiotherapy image sequence data set; S2. Constructing the low-contrast radiotherapy image sequence data set into key-query pairs according to the perspective difference and inputting them into the MoCov3 self-supervised feature encoder model, and performing inference on the low-contrast radiotherapy image sequence data set to obtain an initial feature representation set; S3. The initial feature representation set is uniformly discretely encoded to form an initial wolf pack parameter set. For each wolf pack position vector in the initial wolf pack parameter set, the corresponding parameter combination is coupled with the initial feature representation set to perform image enhancement reconstruction to obtain an image enhancement reconstruction result set. S4. Calculate the tumor edge clarity index and dose consistency index based on the image enhancement reconstruction result set, and generate an updated wolf pack parameter set; S5. Drive the MoCov3 self-supervised feature encoder model with the updated wolf pack parameter set to obtain an updated feature representation set. Re-perform image enhancement reconstruction using the updated feature representation set and the updated wolf pack parameter set to obtain an iterative image enhancement reconstruction result set. Repeat steps S4 to S5 until the preset convergence condition is met, and output a high-contrast radiotherapy-enhanced image. S6. Input the high-contrast radiotherapy enhanced image into the closed-loop quality evaluation module to calculate the error for final quality verification. If the final quality verification result meets the clinical threshold, the result is output. If the final quality verification result does not meet the clinical threshold, the final quality verification result is fed back to the wolf pack fitness vector and returns to step S4 to continue iterating until the clinical threshold is met.

2. The radiotherapy image reconstruction method based on self-supervised learning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Acquiring low-contrast radiotherapy image sequence data within a preset time window by a cone-beam computed tomography system to form a low-contrast radiotherapy image frame set; S12. Perform scattering artifact suppression processing on each radiotherapy image frame in the low-contrast radiotherapy image frame set. A frequency-domain occlusion compensation model is constructed to estimate the low-frequency component of each radiotherapy image frame. The low-frequency component estimation result is subtracted from the original radiotherapy image as a scattering term to generate a scatter-suppressed low-contrast radiotherapy image frame. The scattering term is used to represent the grayscale interference area caused by scattering in the radiotherapy image. S13. Performing metal artifact correction on each frame of the radiotherapy image after scatter suppression processing, extracting the metal area to form a binary metal mask matrix, performing an interpolation reconstruction operation based on the texture features of the non-metal neighborhood within the binary metal mask matrix, and outputting the radiotherapy image frame after metal artifact correction; S14. Performing dose normalization on the radiotherapy image frames after metal artifact correction by extracting the maximum grayscale value of each radiotherapy image frame as an intensity benchmark and normalizing the grayscale values ​​of all pixels in the radiotherapy image frame according to this benchmark; S15. Reorganize all radiotherapy image frames that have undergone scattering artifact suppression processing, metal artifact correction processing, and dose normalization processing to form a pre-processed low-contrast radiotherapy image sequence data set.

3. The radiotherapy image reconstruction method based on self-supervised learning according to claim 2, characterized in that: The S2 comprises the following steps: S21. Construct an improved MoCov3 self-supervised feature encoder model, which includes a multi-scale tumor edge-aware encoder and a momentum-driven soft tissue structure consistency enhancement module, and initialize the network weights of the multi-scale tumor edge-aware encoder and the momentum-driven soft tissue structure consistency enhancement module; S22. Construct an improved key-query pair based on the preprocessed low-contrast radiotherapy image sequence data set according to the perspective difference. The improved key-query pair is defined as: in, represents the query image with the purpose of enhancing the features of the tumor edge area, represents the key image for the purpose of highlighting the internal structural information of soft tissue, and N is the total number of improved key-query pairs; S23. Input the improved key-query pair into the multi-scale tumor edge-aware encoder to obtain a multi-scale fused tumor edge-aware feature representation vector Enhanced feature representation vector with soft tissue structure The vectors have the same length and dimension; S24. Based on the momentum-driven soft tissue structure consistency enhancement module, a soft structure constraint function is constructed between the tumor edge perception feature representation vector and the soft tissue structure enhancement feature representation vector in the feature space. in, represents the dot product similarity between the multi-scale tumor edge perception feature representation vector and the soft tissue structure enhancement feature representation vector, It represents the adaptive feature boundary enhancement item between the tumor edge area and the soft tissue structure area in the feature space, which is used to highlight the difference between the tumor edge area and the soft tissue structure area in the feature space. α is the feature boundary enhancement factor, and τ is the contrast learning temperature parameter, which is used to adjust the similarity scale. represents the jth negative sample feature vector, M is the total number of negative sample feature vectors; S25. According to the soft structure constraint function The calculation results are used to update the network weights of the multi-scale tumor edge perception encoder and the momentum-driven soft tissue structure consistency enhancement module, and steps S23 to S25 are iteratively performed until the improved MoCov3 self-supervised feature encoder model converges; S26. Use the trained improved MoCov3 self-supervised feature encoder model to perform feature inference on the preprocessed low-contrast radiotherapy image sequence data set, extract multi-scale fusion features with high discrimination, and form an initial feature representation set.

4. The radiotherapy image reconstruction method based on self-supervised learning according to claim 3, characterized in that: The S21 includes the following steps: S211. Construct a multi-scale tumor edge-aware encoder, comprising a backbone feature extraction network and a multi-scale edge attention module. The backbone feature extraction network performs convolution operations on each dose-normalized low-contrast radiotherapy image frame at multiple scales to extract a basic texture feature map. S212. Construct an edge activation function based on the basic texture feature map at each scale to extract the tumor edge region. The edge activation function calculates the horizontal and vertical gradient maps of the basic texture feature map, superimposes the two gradient maps, and inputs them into a normalization function to obtain an edge response map. The edge response map is then multiplied point by point with the original basic texture feature map to obtain an edge perception feature map. S213. Concatenate the edge-aware feature maps at all scales according to the channel dimension, and input the concatenated feature maps into a feature fusion network for interactive fusion to form a fused multi-scale tumor edge-aware feature representation vector. The fused multi-scale tumor edge-aware feature representation vector serves as a query feature vector in feature comparison calculation. S214. Construct a momentum-driven soft tissue structural consistency enhancement module. The momentum-driven soft tissue structural consistency enhancement module consists of a structure-aware main encoder and a sliding momentum update encoder. The structure-aware main encoder and the tumor edge perception pathway share a shallow network structure. In the deep layer, an attention mechanism is embedded to enhance the texture representation capability of low-contrast regions. S215. Initialize the parameters of the sliding momentum update encoder, set the initial parameter values ​​to the current parameter values ​​of the structure-aware main encoder, and obtain the current momentum encoder parameters by weighted fusion with the parameters of the structure-aware main encoder in the current round of training; S216. Input the key image into the structure-aware main encoder and the sliding momentum update encoder to extract the structure enhancement feature representation vector and the momentum reference feature representation vector respectively; S217. Complete the construction and parameter initialization of the backbone feature extraction network, the multi-scale edge attention module, and the momentum-driven soft tissue structure consistency enhancement module, and output the query feature representation vector and key feature representation vector required for the first round of training.

5. The radiotherapy image reconstruction method based on self-supervised learning according to claim 1, characterized in that: The discrete coding is achieved by constructing a discretization reconstruction-coding integrated hyperparameter dictionary P ＊ The discretized reconstruction-encoding integrated hyperparameter dictionary contains the following ten discrete hyperparameters: super-resolution convolution kernel index Regularization weight level Comparative learning temperature levels Filter kernel index Sampling mask index Feature boundary enhancement factor Momentum coefficient Multi-scale branch depth Negative sample queue length Edge-Structure Loss Weight Set the minimum and maximum values ​​for each discrete hyperparameter to form a corresponding set of discrete value intervals.

6. The radiotherapy image reconstruction method based on self-supervised learning according to claim 5, characterized in that: The S3 includes the following steps: S31. Use the vector quantization mapping function to represent each feature vector f in the initial feature representation set i Mapped to the nearest neighbor feature center index c i , get the discrete feature index set C; S32. Set the number of gray wolves G and the total number of discrete hyperparameter dimensions D p = 10, generate a random number for each pair (g, j) and map it to integer coordinates according to the corresponding discrete value interval, and obtain the wolf pack position matrix X (0) ; S33. Using one-to-one decoding function The wolf pack position matrix X (0) Each row of the integer coordinate vector decodes to a complete discrete hyperparameter combination And gather to get the initial wolf pack parameter set Θ (0) , each discrete hyperparameter combination in the initial wolf pack parameter set contains both the reconstruction hyperparameters and the internal hyperparameters of the improved MoCov3 encoder and maintains consistency in symbols and semantics; S34. For the initial wolf pack parameter set Θ (0) Each discrete hyperparameter combination in Dynamically instantiate a set of improved MoCov3 self-supervised feature encoder instances E with the feature boundary enhancement factor, momentum coefficient, multi-scale branch depth, negative sample queue length, edge-structure loss weight, and contrastive learning temperature level g , the discrete hyperparameter combination The super-resolution convolution kernel index, filter kernel index, sampling mask index and regularization weight level in the image enhancement reconstruction operator R are synchronously injected into the image enhancement reconstruction operator R, thereby converting the initial feature representation set F init With encoder instance E g and discrete hyperparameter combinations Coupled image enhancement reconstruction and output of the corresponding reconstructed image S35. Collect all reconstructed images Forming a set of image enhancement reconstruction results 7. The radiotherapy image reconstruction method based on self-supervised learning according to claim 6, characterized in that: The S4 comprises the following steps: S41. Image enhancement and reconstruction result set Each reconstructed image Calculate the tumor edge clarity index E g Dose consistency index D g , tumor edge clarity index E g The dose consistency index D is obtained by calculating the mean gradient amplitude of the tumor edge area in the reconstructed image. The mean gradient amplitude is used to quantify the clarity of the spatial details of the tumor edge in the reconstructed image. g It is obtained by calculating the normalized mean square error between the gray value distribution of the tumor area in the reconstructed image and the target dose distribution set in the radiotherapy plan; S42. Constructing a dual-index fusion fitness function F for tumor edge clarity and dose consistency g , and to meet the clinical needs of low-contrast radiotherapy image enhancement reconstruction, a nonlinear balance factor w is introduced balance Adaptively adjust the weight relationship between the two indicators: F g =w balance ·E g +(1-w balance )·(1-D g ); S43. Calculate the individual fitness values ​​of gray wolves using a dual-index fusion fitness function. Divide the wolf pack into two independent subgroups based on the proportion of their individual fitness values: an imaging structure optimization wolf pack and a dose accuracy optimization wolf pack. Wolves whose tumor edge clarity index accounts for a dominant proportion of their individual fitness values ​​are assigned to the imaging structure optimization wolf pack; Wolves whose dose consistency index accounts for a dominant proportion of their individual fitness values ​​are assigned to the dose accuracy optimization wolf pack. The optimal individual gray wolf positions in each subgroup are determined as the structure leader and dose leader, respectively. S44. In the image structure optimization wolf pack, define the structure guidance factor S guide Dynamically control the update strategy of each gray wolf individual moving towards the position of the structural leader wolf, the structural guidance factor S guide The value of is determined by the difference in tumor edge clarity between the gray wolf individual and the structural leader wolf. When the difference is greater than the threshold, the structural guidance factor S is increased. guide , strengthen the approach to the position of the structural leader, and reduce the structural guidance factor S when the gap is less than the threshold guide ; S45. Define the dose guidance factor D in the dose accuracy optimization wolf group guide Dynamically control the update strategy of each gray wolf individual moving towards the position of the dose leader wolf, the dose guidance factor D guide The value of is determined by the difference between the individual gray wolf and the dose leader wolf in the dose consistency index. When the difference is greater than the threshold, the dose guidance factor D is increased. guide , strengthen the approach to the dose leader position, and reduce the dose guidance factor D when the gap is less than the threshold guide ; S46. Introduce a structure-dose cross-group synergy mechanism. During each iteration, the overall fitness of the two subgroups of the wolf pack is improved based on the image structure optimization and the dose accuracy optimization. Some gray wolf individuals are swapped across subgroups. Specifically, if the overall fitness of a subgroup does not improve after multiple consecutive iterations, several gray wolf individuals in that subgroup are selected and swapped to adjust their positions to the search space of another subgroup. This optimizes the cross-group information sharing capability and overall convergence efficiency of the discrete gray wolf optimization algorithm in low-contrast radiotherapy image enhancement reconstruction. S47. After each round of iteration, the updated positions of all individual gray wolves in the two subgroups are reassembled and formed into an updated wolf pack position matrix, and the wolf pack position matrix is ​​decoded into an updated wolf pack parameter set using a one-to-one decoding function consistent with S3.

8. The radiotherapy image reconstruction method based on self-supervised learning according to claim 7, characterized in that: The S5 comprises the following steps: S51. Update each discrete hyperparameter combination in the wolf pack parameter set Dynamically configure the improved MoCov3 self-supervised feature encoder model by setting the operating parameters of the boundary separation module, momentum update module, multi-scale structure perception module, and queue management strategy in the encoder according to the feature boundary enhancement factor, momentum coefficient, multi-scale branch depth, negative sample queue length, edge-structure loss weight, and contrastive learning temperature parameters, and construct a self-supervised feature encoder instance in the current round S52. Refresh the momentum encoder queue of the current round of self-supervised feature encoder. The momentum encoder queue is refreshed by the positive sample feature set participating in the training from the previous round. The positive sample feature set output by the current round encoder Extract feature representation, and fuse the features using exponential sliding average according to the momentum update coefficient to generate a new momentum encoder queue; S53. Use the self-supervised feature encoder instance configured in the current round Perform feature extraction on the pre-processed low-contrast radiotherapy image sequence data to obtain the updated feature representation set F (t) ; S54. Update the feature representation set F extracted in the current round (t) Combined with the current discrete hyperparameters The images are input into the image enhancement reconstruction operator together to complete the image enhancement reconstruction operation. During the reconstruction process, the specific reconstruction path is configured according to the super-resolution convolution kernel index, filter kernel index, sampling mask index and regularization weight level in the discrete hyperparameter combination, and an enhanced reconstructed image corresponding to the discrete hyperparameter combination is output. The enhanced reconstructed image is named the enhanced reconstructed image of the current gray wolf individual in the current round. S55. Enhanced reconstructed images output by all gray wolf individuals in the current round Arrange and summarize in order to construct a complete set of image enhancement and reconstruction results for this round S56. Using image enhancement to reconstruct the result set Tumor edge clarity index Dose consistency index And construct a dual-index fusion fitness function Finally, a new round of wolf position matrix X is generated (t+1) and its corresponding parameter combination set Θ (t+1) ; S57. Determine the current round fitness vector F (t) Whether the preset convergence threshold ∈ conv , if satisfied, then select the gray wolf individual with the largest value from the fitness vector and output its corresponding optimal hyperparameter combination θ (*) and the corresponding final high-contrast radiotherapy-enhanced image I (*) If not, the iteration count is updated to t+1, and S51 to S56 are repeated until the convergence condition is met.

9. The radiotherapy image reconstruction method based on self-supervised learning according to claim 1, characterized in that: The S6 comprises the following steps: S61. Enhanced imaging in high-contrast radiotherapy I (*) Import the closed-loop quality evaluation module, which will load the optimal reconstruction parameter configuration θ synchronously (*) Initialize the final quality verification process with the patient's original DICOM-RT structure set and radiotherapy plan dose distribution; S62. In the closed-loop quality evaluation module, the organ delineation consistency index based on contour overlap is used to align the tumor target area with the key organs at risk, and the high-contrast radiotherapy enhanced image I (*) The structure contours obtained by semi-automatic segmentation were compared with the corresponding contours of the clinical DICOM-RT structure set at the pixel level to obtain the organ outline consistency index Dice ROI ; S63. In the closed-loop quality assessment module, the dose recalculation error index based on the recalculated dose volume error is used to verify the accuracy of the dose measurement, and the original beam parameters of the radiotherapy plan are mapped to the high-contrast radiotherapy enhanced image I. (*) The three-dimensional dose distribution is then recalculated to obtain the recalculated dose volume, and the dose recalculation error index Dose is obtained by differentiating the recalculated dose volume from the original planned dose volume at the voxel level. RMSE ; S64. Based on the clinical threshold set {τ Dice ,τ Dose Dice index for organ delineation consistency ROI and dose recalculation error index Dose RMSE Perform joint judgment: When organ delineation consistency index Dice ROI ≥τ Dice And the dose recalculation error index Dose RMSE ≤τ Dose When the final quality verification is passed; When any of the above conditions is not met, the final quality verification is deemed to have failed; S65. If the final quality verification is passed, the high-contrast radiotherapy image result I that has passed the final quality verification is output. (*) and the corresponding optimal reconstruction parameter configuration θ (*) ; S66. If the final quality verification fails, the organ will be outlined with the consistency index Dice ROI and dose recalculation error index Dose RMSE It is converted into a negative feedback gain value in proportion and injected into the wolf pack fitness vector of the previous round, correspondingly reducing the fitness contribution in the indicator direction that does not meet the threshold. Then it returns to S41 and re-executes the iterative process of image enhancement reconstruction and wolf pack position update until the joint threshold condition is met.

Images