Lung tumor tracking method based on four-dimensional ct registration prior and projection time series observation

Abstract

The application discloses a lung tumor tracking method based on four-dimensional CT registration prior and projection time sequence observation. The method firstly acquires a lung four-dimensional CT image sequence collected last time, labels a tumor region at a reference respiratory phase, and transmits the tumor region to the rest of the respiratory phases through multi-phase deformable registration, calculates three-dimensional displacement states of each phase relative to the reference respiratory phase, and constructs an individualized tumor displacement prior sequence. Subsequently, an X-ray projection image sequence is acquired or a digital reconstructed radiogram image sequence is generated according to a lung four-dimensional CT image sequence collected last time as time sequence projection observation, and a differential correlation ratio constraint observation fusion module is used to reduce domain difference. Finally, the displacement prior feature and the enhanced projection observation feature are fused, input into a time sequence tracking network, and continuous estimation and dynamic tracking of the three-dimensional displacement state of the lung tumor are realized.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, computer vision, and artificial intelligence, specifically to a method for dynamic tracking of lung tumors based on four-dimensional CT temporal registration prior and projection temporal observation. This method can be applied to scenarios such as thoracic tumor radiotherapy, image-guided radiotherapy, and respiratory motion management, for continuous dynamic tracking of lung tumors under the influence of respiratory motion, and estimation of their three-dimensional motion state or trajectory. Background Technology

[0002] Lung tumors undergo significant changes in position, shape, and local tissue deformation with a patient's breathing, especially during chest radiotherapy, where respiratory motion causes periodic or non-periodic spatial displacement of the tumor target area. If the real-time position and motion of the tumor during treatment cannot be accurately determined, it can easily lead to target area deviation, affecting the accuracy of radiotherapy dose coverage and increasing the risk of radiation exposure to surrounding normal tissues. Therefore, achieving precise dynamic tracking of lung tumors under respiratory motion conditions has always been a key issue in image-guided radiotherapy and motion management.

[0003] Among existing methods for managing lung tumor motion, four-dimensional computed tomography (4DCT) provides three-dimensional anatomical information at different phases within a respiratory cycle, making it a crucial tool for establishing tumor respiratory motion models. By registering 4D CT images at different respiratory phases, the temporal deformation relationship between the tumor and surrounding tissues can be estimated, thus providing a basis for modeling patient respiratory motion. However, 4D CT scans are typically acquired before treatment or at a specific moment, reflecting motion patterns under offline acquisition conditions, making them difficult to directly use for continuous online tracking during subsequent treatment. Especially at different fractions or time points, the patient's respiratory amplitude, rhythm, and the motion relationship between the tumor and surrounding tissues may change. Static motion models established solely based on a single 4D CT scan often fail to adequately represent the true motion state in subsequent stages.

[0004] In clinical practice, X-ray fluoroscopy or cone-beam projection images have advantages such as fast imaging speed and ease of acquisition, and are often used for online monitoring during treatment. Tumor tracking based on projection images can reflect the patient's current respiratory status and anatomical changes to some extent. However, because projection images are two-dimensional, their depth information is significantly lacking, and they are easily affected by the overlap of ribs, cardiac silhouettes, mediastinum, and other high-density structures, resulting in low contrast and unclear boundaries in the tumor area. This increases the difficulty of directly reconstructing the three-dimensional tumor location from two-dimensional projection. Especially under conditions where there are no implanted markers or weakly visible targets, relying on only a single frame or a small number of projection images for three-dimensional state estimation often suffers from instability, susceptibility to noise interference, and insufficient robustness.

[0005] To address the aforementioned issues, some studies have attempted to combine prior motion models with online projection observations to improve the stability of tumor tracking. Existing techniques include methods that construct patient-specific respiratory motion models to estimate subsequent online states using motion information obtained before treatment; and methods that use deep learning models to directly regress tumor location or displacement parameters from projection sequences. While these methods improve tracking efficiency to some extent, they still have the following shortcomings: First, existing prior construction methods typically rely on precise multi-phase tumor annotation, which is costly and complex to implement clinically, making it difficult to efficiently establish individualized temporal motion priors. Second, some methods do not fully utilize the deformation propagation relationships between multiple temporal phases of four-dimensional CT, making it difficult to form dynamic prior expressions with continuous temporal constraints. Third, the use of two-dimensional projection information in the online observation phase is mainly based on feature regression, lacking explicit constraints consistent with the imaging observation process. This leads to problems such as decreased generalization performance and accumulated state estimation biases under different imaging conditions, different data fractions, or cross-domain projection conditions.

[0006] Therefore, a novel dynamic tracking method for lung tumors is urgently needed. This method should fully utilize patient-specific respiratory motion information from previously acquired 4D CT scans, construct individualized motion priors through temporal registration propagation, and continuously estimate tumor motion status by combining subsequent short-term projection observations. Simultaneously, observational constraints consistent with the projection imaging mechanism should be introduced during the tracking process to improve the model's robustness to cross-time and cross-conditional projection observations and the accuracy of 3D trajectory estimation, thus meeting the clinical application needs of dynamic tracking of lung tumors. Summary of the Invention

[0007] Objective: This invention aims to provide a method for dynamic tracking of lung tumors based on 4D CT temporal registration priors and projection temporal observation, addressing the problem in existing technologies where stable and accurate 3D dynamic tracking of lung tumors under the influence of respiratory motion is difficult. This method fully utilizes patient-specific respiratory motion information contained in previously acquired 4D CT scans. By establishing multi-temporal dynamic deformation relationships through temporal registration between different respiratory phases, and relying solely on tumor annotation in a reference respiratory phase, it propagates the tumor region from the reference phase to other respiratory phases, constructing a tumor displacement prior sequence that characterizes the individualized respiratory motion patterns of the patient. This reduces the manual cost and clinical implementation complexity associated with sequential annotation across multiple phases, improving the efficiency and feasibility of individualized motion prior modeling. Building upon this foundation, this invention further integrates projection time-series observation information from subsequent moments to address the cross-modal domain differences between simulated projection observations and real X-ray observations. It constructs an observation enhancement mechanism and introduces a local statistical structure consistency constraint driven by a differentiable correlation ratio during the enhancement process. This improves the model's adaptability to projection observations across time periods, fractions, and imaging conditions, enhancing the stability and robustness of observation information utilization during subsequent tracking. Simultaneously, this invention constructs a priori-guided time-series tracking model. It fuses the displacement prior obtained from four-dimensional CT registration propagation with the enhanced projection observation features and combines this with a time-series recursive estimation mechanism to dynamically predict the three-dimensional displacement state of the lung tumor relative to a reference respiratory phase during continuous respiratory phases. The model outputs the dynamic displacement results of the tumor in the left-right, front-back, and head-to-foot directions, achieving continuous estimation and dynamic tracking of the three-dimensional motion state of the lung tumor. The method proposed in this invention can be effectively applied to clinical scenarios such as stereotactic radiotherapy for lung cancer, image-guided radiotherapy, and respiratory motion management. It has important application value in tumor motion monitoring, dynamic evaluation of the treatment process, and precise tracking of the radiotherapy target area. It helps to improve the accuracy of tumor motion trajectory estimation, reduce irradiation deviation caused by respiratory motion, reduce the risk of false irradiation of normal tissues, and thus improve the safety and therapeutic effect of radiotherapy.

[0008] 1. A method for tracking lung tumors based on four-dimensional CT registration prior and projection temporal observation, characterized by comprising the following steps:

[0009] Step 1.1: Obtain the previous four-dimensional CT image sequence of the lungs of the same patient and mark the tumor region on the reference respiratory phase; use multi-phase deformable registration to propagate the tumor region to other respiratory phases, calculate the three-dimensional displacement state of the tumor relative to the reference respiratory phase in each phase, and construct an individualized tumor displacement prior sequence.

[0010] Step 1.2: Obtain the X-ray projection image sequence acquired before or during radiotherapy, or generate a digitally reconstructed X-ray image sequence based on the subsequently acquired four-dimensional CT image sequence of the lungs, as the temporal projection observation sequence; enhance the temporal projection observation sequence by using a projection observation enhancement module that integrates differentiable correlation ratio constraints to reduce the domain difference between simulated projection observation and real X-ray observation, and obtain the enhanced projection observation sequence;

[0011] Step 1.3: Input the tumor displacement prior sequence into the prior encoder for feature extraction to obtain prior features characterizing the movement and changes of the tumor during continuous respiration; input the enhanced post-projection observation image at the current moment into the observation encoder for feature extraction to obtain observation features characterizing the anatomical structure information and lesion projection response information at the current moment; perform alignment and fusion processing on the prior features and observation features to construct a joint input feature representation containing temporal motion prior information and current observation information;

[0012] Step 1.4: Input the joint input feature representation into the prior-guided temporal tracking network, and recursively update the hidden representation at the current time by combining the hidden state at the previous time step; based on the updated hidden representation at the current time step, output the predicted result of the three-dimensional displacement state of the lung tumor at the current time step relative to the reference respiratory phase through the state prediction module, wherein the three-dimensional displacement state includes left-right displacement components, front-back displacement components, and head-to-toe displacement components; pass the hidden representation at the current time step to the next time step to achieve dynamic tracking of the three-dimensional displacement state of the lung tumor under continuous respiratory phases.

[0013] 2. The method for tracking lung tumors based on four-dimensional CT registration prior and projection temporal observation according to claim 1, characterized in that the method for constructing the individualized tumor displacement prior sequence in step 1.1 is as follows:

[0014] Step 2.1: Represent the sequence of previous four-dimensional CT multi-temporal images of the lungs of the same patient as follows: ,in These represent CT images at different respiratory phases within a complete respiratory cycle; [Selection] As a reference respiratory phase image, the tumor region was identified on the reference respiratory phase image. Outlining;

[0015] Step 2.2: Perform deformable registration on CT images between adjacent respiratory phases, estimate the deformation field between adjacent phases, and obtain the set of deformation fields between adjacent phases within a complete respiratory cycle. :

[0016]

[0017] in, This represents the deformation field from the 00th time phase to the 10th time phase; the meanings of the other symbols are similar.

[0018] Step 2.3: Using the reference breathing phase as a unified benchmark, obtain the deformation relationship from the reference breathing phase to any target breathing phase by cascading adjacent phase deformation fields. And using deformation field-driven spatial transformation operations to transform the tumor region under the reference respiratory phase. Propagated to the remaining time phases, the first... Tumor region at different time phases The propagation formula is as follows:

[0019]

[0020] in, Indicates the spread to the th Tumor regions after each time phase This represents a spatial transformation operation driven by a deformation field. This indicates the deformation relationship between the reference breathing phase and the target breathing phase;

[0021] Step 2.4: Calculate the center location of the tumor region at each time phase, using the reference respiratory phase as a unified benchmark, and calculate the [missing information - likely a specific location or step]. The three-dimensional displacement of the tumor center relative to the reference respiratory phase tumor center at each time phase is calculated using the following formula:

[0022]

[0023] in, Indicates the first Tumor center coordinates at each time phase This indicates the coordinates of the tumor center at the reference respiratory phase. Indicates the first The three-dimensional displacement state of each phase. , , The corresponding displacement states at each time point were organized according to the respiratory time sequence to construct the tumor displacement prior sequence S. .

[0024] 3. The method for tracking lung tumors based on four-dimensional CT registration prior and projection temporal observation according to claim 1, characterized in that the method for cross-domain observation enhancement processing in step 1.2 is as follows:

[0025] Step 3.1: Obtain multi-temporal image sequences of the lungs from subsequent four-dimensional CT scans of the same patient, generate corresponding digitally reconstructed X-ray images using consistent projection geometry parameters, and obtain the original projection observation sequence. ;

[0026] Step 3.2: The projection observation enhancement module includes a generator network for cross-domain mapping, a feature encoder for feature extraction, and a discriminator for distribution discrimination, which are used to realize cross-domain conversion from the digitally reconstructed X-ray image domain to the real X-ray style domain.

[0027] Step 3.3: The projection observation enhancement module is optimized using a joint loss function that includes adversarial loss, block-level contrast loss, and local image block-level differentiable correlation ratio constraint. Its expression is as follows:

[0028]

[0029]

[0030]

[0031] in, This represents the joint loss due to the enhancement of projected observations. This represents the adversarial loss used to push the enhanced projected image to approximate the true X-ray domain in its overall distribution. This represents the block-level contrast loss used to maintain content consistency between the input projected image and the output enhanced image in corresponding local regions. This represents a local block-level differentiable correlation ratio constraint used to enhance the consistency of local statistical structure between the enhanced projected image and the target domain image. and represents the weights of the block-level contrast loss and the local block-level differentiable correlation ratio constraint, respectively, and their values ​​were determined through experimental parameter tuning. Indicates a generator; Indicates the discriminator; Represents the true X-ray image of the target domain; Represents the original projected image; This represents the enhanced projected image; Represents local block features; Represents the set of local block features of negative samples; This represents the contrastive learning loss function;

[0032] Step 3.4: The local image block-level differentiable correlation ratio constraint is obtained by calculating the symmetric differentiable correlation ratio loss at the local image block level; let the enhanced projection image be... Compared with the target domain real X-ray image The For local blocks respectively and Its symmetric differentiable correlation ratio loss is defined as:

[0033]

[0034] in, Represents the differentiable correlation ratio; for the division The local image block-level differentiable correlation ratio constraint is calculated by averaging the symmetric differentiable correlation ratio loss of the local blocks:

[0035]

[0036] Step 3.5: Using the projection observation enhancement module trained with the joint loss function, the input original projection observation sequence is mapped to output an enhanced projection observation sequence that more closely approximates the true X-ray domain representation. .

[0037] Beneficial results of the present invention:

[0038] This invention addresses the challenge of achieving stable and accurate three-dimensional dynamic tracking of lung tumors under the influence of respiratory motion. It proposes a dynamic tracking solution for lung tumors by combining four-dimensional CT temporal registration priors, enhanced projection temporal observations, and a prior-guided temporal recursive estimation mechanism. This method can construct individualized patient displacement priors relying solely on tumor annotations from reference respiratory phases. Furthermore, it enhances the model's adaptability to cross-temporal and cross-conditional projection observations through local block-level differentiable correlation ratio constraints. This improves the accuracy, stability, and robustness of three-dimensional displacement state estimation and motion trajectory recovery for lung tumors, enhancing the application value of this technology in practical clinical scenarios such as lung cancer radiotherapy, image-guided radiotherapy, and respiratory motion management. Attached Figure Description

[0039] Figure 1 This is a flowchart of a method for dynamic tracking of lung tumors based on four-dimensional CT temporal registration prior and projection temporal observation provided by an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of the overall framework of the priori-guided dynamic tracking model for lung tumors provided in this embodiment of the invention;

[0041] Figure 3 This is a schematic diagram of the projection observation enhancement module structure that integrates local block-level differentiable correlation ratio constraints provided in an embodiment of the present invention. Detailed Implementation

[0042] The invention will now be further described with reference to the accompanying drawings.

[0043] like Figures 1 to 3As shown, the lung tumor tracking method based on four-dimensional CT registration prior and projection temporal observation described in this invention mainly includes a prior data input and preprocessing module, a multi-temporal dynamic deformation modeling module, a displacement prior construction and projection observation enhancement module, and a prior-guided temporal tracking module. Its specific implementation process is as follows:

[0044] Step 1: Obtain the previously acquired 4D CT image sequence of the lungs from the same patient. Preprocess the lung 4D CT image sequence, select a reference respiratory phase image, and annotate the lung tumor region on the reference respiratory phase image. The prior data input and preprocessing process is as follows: Figure 1 As shown.

[0045] Step 1.1: Obtain the previous four-dimensional CT image sequence of the lungs for the same patient. The four-dimensional CT image sequence of the lungs consists of three-dimensional CT images corresponding to multiple breaths within a complete respiratory cycle, used to characterize the dynamic changes in the lung anatomy and tumor region under different respiratory states;

[0046] Step 1.2: Preprocess the previously acquired 4D CT lung image sequence. The preprocessing includes operations such as respiratory phase sorting, image size standardization, voxel resolution resampling, and grayscale intensity normalization to reduce the impact of different scanning conditions and spatial resolution differences on the subsequent dynamic deformation modeling process.

[0047] Step 1.3: After preprocessing, the previously acquired four-dimensional CT image sequence of the lungs is represented as follows:

[0048]

[0049] in, This represents a sequence of multi-temporal images of the lungs acquired in a previous 4D CT scan. These represent CT images at different respiratory phases within a complete respiratory cycle;

[0050] Step 1.4: Select a reference respiratory phase image from the previous four-dimensional CT image sequence of the lungs. Preferably, the end-expiratory phase or end-inspiratory phase, where the anatomical structure is relatively stable during the respiratory cycle, is selected as the reference respiratory phase. In this embodiment, the following is selected: As a reference respiratory phase image;

[0051] Step 1.5: Mark the lung tumor region on the reference respiratory phase image to obtain the tumor region at the reference respiratory phase, denoted as... ;

[0052] Step 1.6: After completing the above processing, the previous four-dimensional CT image sequence of the lungs is obtained for subsequent multi-temporal dynamic deformation modeling. and tumor regions at reference respiratory phases Among them, the previous four-dimensional CT image sequence of the lungs Used for deformable registration and dynamic deformation model construction between subsequent adjacent respiratory phases, referencing the tumor region at the respiratory phase. Used for subsequent tumor region propagation and displacement prior sequence construction.

[0053] Step 2: Perform deformable registration on adjacent respiratory phase images in the previously acquired 4D CT lung image sequence to obtain the deformation field between adjacent phases within the respiratory cycle. Based on the deformation field between adjacent phases, construct a multi-phase dynamic deformation model, wherein the multi-phase dynamic deformation modeling process is as follows: Figure 1 , Figure 2 As shown.

[0054] Step 2.1: Represent the previously acquired 4D CT image sequence of the lungs as follows:

[0055]

[0056] in, This represents a sequence of multi-temporal images of the lungs acquired in a previous 4D CT scan. These represent CT images at different respiratory phases within a complete respiratory cycle;

[0057] Step 2.2: Construct adjacent respiratory phase image pairs according to the respiratory time sequence, and perform deformable registration on each pair of adjacent respiratory phase images to obtain the images from the first respiratory phase. The first breath phase to the second The deformation field of each breathing phase is represented as follows:

[0058]

[0059] in, This represents a deformable registration model. Indicates from the first The first breath phase to the second Deformation field of adjacent phases of a breathing phase;

[0060] Step 2.3: Repeat step 2.2 for all adjacent respiratory phase image pairs within the respiratory cycle to obtain the set of deformation fields for adjacent phases:

[0061]

[0062] in, This represents the set of adjacent temporal deformation fields within a complete respiratory cycle;

[0063] Step 2.4: Using reference respiratory phase To unify the spatial reference, the deformation fields of adjacent time phases are cascaded to construct the deformation relationship from the reference breathing phase to any target breathing phase, expressed as:

[0064]

[0065] in, This represents the cascade combination operation of deformation fields. Indicates the transition from the reference breathing phase to the target breathing phase. The cumulative deformation relationship;

[0066] Step 2.5: Set the adjacent temporal deformation fields And the cumulative deformation relationship from the reference respiratory phase to each target respiratory phase. The common organization is a multi-temporal dynamic deformation model within the respiratory cycle, used to characterize the temporal motion relationship and non-rigid deformation relationship of lung tumors and surrounding tissues during continuous respiration.

[0067] Step 3: Using the multi-temporal dynamic deformation model, the lung tumor region in the reference respiratory phase image is propagated to the remaining respiratory phases to obtain the tumor region corresponding to each respiratory phase. Furthermore, the three-dimensional displacement vector of the tumor region relative to the reference respiratory phase is calculated for each respiratory phase, constructing a tumor displacement prior sequence characterizing the patient's individualized respiratory motion patterns. The process of tumor region propagation and displacement prior construction is as follows: Figure 1 , Figure 2 As shown.

[0068] Step 3.1: Record the reference respiratory phase image The upper lung tumor area is ;

[0069] Step 3.2: Based on the multi-temporal dynamic deformation model constructed in Step 2, the lung tumor region in the reference respiratory phase image is propagated to each target respiratory phase to obtain the tumor region corresponding to the t-th respiratory phase, which is represented as:

[0070]

[0071] in, This represents a spatial transformation operation driven by a deformation field. Indicates the reference breathing phase up to the 1st The cumulative deformation relationship of each breathing phase Indicates the spread to the th Lung tumor area after one respiratory phase;

[0072] Step 3.3: Calculate the center location of the lung tumor region at each respiratory phase, and let the first... The tumor center coordinates at each respiratory phase are: The reference tumor center coordinates during the respiratory phase are as follows: ;

[0073] Step 3.4: Using the tumor center coordinates at the reference respiratory phase To standardize the calculation, the first... The three-dimensional displacement vector of the lung tumor relative to a reference respiratory phase at each respiratory phase is expressed as follows:

[0074]

[0075] in, Indicates the first The three-dimensional displacement state vector of the lung tumor relative to the reference respiratory phase at each respiratory phase; Indicates the first Tumor center coordinates at each respiratory phase; This indicates the coordinates of the tumor center at the reference respiratory phase; Indicates the first The left-right displacement component of the tumor during each respiratory phase. This represents the displacement component of the tumor in the anterior-posterior direction during the t-th respiratory phase. Indicates the first The displacement component of the tumor in the head-to-foot direction during each respiratory phase;

[0076] Step 3.5: Organize the three-dimensional displacement vectors corresponding to each respiration according to the respiratory time sequence to construct a priori sequence of lung tumor displacement relative to a reference respiratory phase within a complete respiratory cycle:

[0077]

[0078] Wherein, S represents the lung tumor displacement prior sequence, and T represents the total number of respiratory phases within a complete respiratory cycle; the displacement prior sequence is used to characterize the individualized respiratory movement pattern of the patient and serves as the prior input for the subsequent time-series tracking network.

[0079] Step 4: Acquire the X-ray projection image sequence acquired before or during radiotherapy, or generate a digitally reconstructed X-ray image sequence based on the subsequently acquired four-dimensional CT lung image sequence, as the temporal projection observation sequence; perform observation enhancement processing on the temporal projection observation sequence, introducing a local block-level differentiable correlation ratio constraint during the enhancement process to reduce the cross-modal domain difference between simulated projection observation and real X-ray observation, obtaining the enhanced projection observation sequence, wherein the projection observation enhancement process is as follows: Figure 3 As shown.

[0080] Step 4.1: Perform observation enhancement mapping on the projection images of each phase in the time-series projection observation sequence to obtain the enhanced projection images of each phase, which are represented as follows:

[0081]

[0082] in, Indicates the first Enhanced post-projection images corresponding to each breath Indicates the first The original projection image corresponding to each breath. Represents the observation enhancement mapping function;

[0083] Step 4.2: During the observation enhancement mapping process, the input projected image is... Compared with the enhanced projection image Multi-layer local features are extracted, and content consistency constraints are established between corresponding local regions to maintain the local correspondence between the lung anatomy and tumor-related regions before and after enhancement.

[0084] Step 4.3: Project the enhanced image of the t-th respiratory phase. X-ray images of the target domain A sliding window method is used to divide a local region, and the size of the local window is set to . The sliding step size is Extracting from the enhanced projection image and the target domain X-ray image The local block corresponding to the location is represented as:

[0085]

[0086] in, Indicates the first The set of local block pairs constructed under each breathing phase Indicates the enhanced first The first respiratory phase projection image A local block, Represents the corresponding first element in the target domain X-ray image. A local block;

[0087] Step 4.4: For each pair of local blocks Calculate the differentiable correlation ratio of the local X-ray patch in the target domain with respect to the enhanced projected local patch. And the differentiable correlation ratio of the enhanced projected local block with respect to the target domain X-ray local block. And construct a symmetric differentiable correlation ratio loss:

[0088]

[0089] in, Indicates the first The first breathing phase Symmetrically differentiable correlation ratio loss for local blocks. Represents a differentiable correlation ratio function;

[0090] Step 4.5: For the first The symmetric differentiable correlation ratio losses of all local blocks under each respiratory phase are accumulated and normalized to obtain the first... The corresponding local block-level differentiable correlation ratio constraint loss for each breath:

[0091]

[0092] in, Indicates the first Local block-level differentiable correlation ratio constraint loss under each breathing phase This represents the total number of local block pairs obtained from the partitioning;

[0093] Step 4.6: The local block-level differentiable correlation ratio constraint is used as a local statistical structure consistency constraint in the observation enhancement process. It works together with the content consistency constraint between the input projection image and the enhanced projection image to improve the consistency between the enhanced projection image and the real X-ray image in terms of local gray-level distribution, texture statistics and structural expression.

[0094] Step 4.7: Organize the enhanced temporal projection images according to the original respiratory time sequence to obtain the enhanced temporal projection observation sequence. The enhanced temporal projection observation sequence is then used as the observation input for the subsequent prior-guided temporal tracking network.

[0095] Step 5: Encode the tumor displacement prior sequence and the enhanced projection observation sequence separately to obtain prior features and observation features, and fuse the prior features and observation features; input the fused features into a prior-guided temporal tracking network to recursively estimate the three-dimensional displacement state of the lung tumor relative to a reference respiratory phase during continuous breathing phases, and output the dynamic displacement results of the lung tumor in the left-right, front-back, and head-to-foot directions, wherein the prior-guided temporal tracking process is as follows: Figure 2 As shown.

[0096] Step 5.1: Place the first The enhanced post-projection images at each respiratory phase are input into the observation feature encoder to extract the corresponding observation feature representations:

[0097]

[0098] in, This represents the observational features extracted from the enhanced projection image at the t-th respiration phase. Represents the observation feature encoder. Indicates the first Enhanced post-projection images at each respiratory phase;

[0099] Step 5.2: Input the prior sequence of lung tumor displacement into the prior encoder for feature extraction to obtain prior features used to characterize the individualized respiratory movement patterns of the patient:

[0100]

[0101] in, This represents the prior features obtained by encoding the prior sequence of lung tumor displacement. Indicates the prior encoder, This represents the prior sequence of displacement of a lung tumor relative to a reference respiratory phase within a complete respiratory cycle.

[0102] Step 5.3: Determine the observed features at the current moment. with prior features By integrating the data, a time-series tracing network is constructed in the first stage. Input feature representation under each breathing phase:

[0103]

[0104] in, Indicates the first Fusion characteristics of input time-series tracking networks under each breathing phase. This represents the fusion function of observed features and prior features;

[0105] Step 5.4: Merge the features The hidden state of the previous moment A common input temporal tracking network is used to recursively estimate the motion state of lung tumors during continuous respiratory phases, obtaining the first... Hidden state updated after each breath phase:

[0106]

[0107] in, This indicates that the time-tracing network is in the first... The hidden state after updating in each breathing phase Indicates the first The hidden state under each breathing phase This represents the recursive update function of the time-series tracing network;

[0108] Step 5.5: Update the hidden state Input displacement state regression module to obtain lung tumor in the [number]th [year]. Three-dimensional displacement state prediction results relative to the reference respiratory phase at each respiratory phase:

[0109]

[0110] in, Indicates the first The predicted three-dimensional displacement state vector of the lung tumor relative to a reference respiratory phase at each respiratory phase. This represents the displacement state regression function;

[0111] Step 5.6: Organize the predicted three-dimensional displacement state vectors for all respiratory phases in chronological order to obtain the predicted displacement state sequence for dynamic tracking of lung tumors:

[0112]

[0113] in, This represents the predicted displacement sequence of a lung tumor relative to a reference respiratory phase within a complete respiratory cycle. Indicates the first Predicted three-dimensional displacement state vector at each breathing phase Indicates the total number of respiratory phases;

[0114] Step 5.7: The predicted three-dimensional displacement state vector is further expressed as:

[0115]

[0116] in, Indicates the first Predicted displacement components of lung tumors in the left-right direction during each respiratory phase. Indicates the first Predicted displacement components of lung tumors in the anterior-posterior direction during each respiratory phase. Indicates the first The predicted displacement components of the lung tumor in the head-to-foot direction during each respiratory phase; through the predicted displacement state sequence The displacement components of each phase were analyzed to obtain dynamic displacement curves of the lung tumor in the left-right, front-back, and head-to-toe directions.

[0117] This invention utilizes the temporal registration relationship between different respiratory phases in a previous four-dimensional CT scan to construct a dynamic deformation model from a reference respiratory phase to the remaining respiratory phases. Under the condition of tumor annotation relying solely on the reference respiratory phase, the tumor region is propagated throughout the entire respiratory cycle, thus forming a displacement prior sequence capable of characterizing the individualized respiratory motion patterns of the patient. By introducing an observation enhancement mechanism and local block-level differentiable correlation ratio constraints in the subsequent projection observation modeling process, the consistency of local statistical structure between simulated projection observations and real X-ray observations is improved, enhancing the model's adaptability to projection data across time periods, fractions, and imaging conditions. By constructing a priori-guided temporal tracking network, displacement prior features are fused with enhanced projection observation features, and combined with a temporal recursive estimation mechanism, continuous dynamic prediction of the three-dimensional displacement state of lung tumors is achieved, thereby improving the accuracy, stability, and robustness of tumor motion trajectory estimation under complex respiratory conditions.

[0118] The detailed descriptions listed above are merely specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Any equivalent transformations, substitutions, improvements, or combinations made by those skilled in the art to the above embodiments without departing from the technical concept and essence of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for tracking lung tumors based on registration prior and projection temporal observation of four-dimensional CT computed tomography, characterized in that, Includes the following steps: Step 1.1: Obtain the previous four-dimensional CT image sequence of the lungs of the same patient and mark the tumor region on the reference respiratory phase; use multi-phase deformable registration to propagate the tumor region to other respiratory phases, calculate the three-dimensional displacement state of the tumor relative to the reference respiratory phase in each phase, and construct an individualized tumor displacement prior sequence. Step 1.2: Obtain the X-ray projection image sequence acquired before or during radiotherapy, or generate a digital reconstructed X-ray image sequence based on the subsequently acquired four-dimensional CT lung image sequence, as a time-series projection observation sequence; The time-series projection observation sequence is enhanced by a projection observation enhancement module that integrates differentiable correlation ratio constraints, thereby reducing the domain difference between simulated projection observations and real X-ray observations and obtaining an enhanced projection observation sequence. Step 1.3: Input the tumor displacement prior sequence into the prior encoder for feature extraction to obtain prior features characterizing the movement and change of the tumor during continuous respiration. The enhanced post-projection observation image at the current moment is input into the observation encoder for feature extraction to obtain observation features that characterize the anatomical structure information and lesion projection response information at the current moment; the prior features and observation features are aligned and fused to construct a joint input feature representation that includes temporal motion prior information and current observation information; Step 1.4: Input the joint input feature representation into the prior-guided temporal tracking network, and recursively update the hidden representation at the current time by combining the hidden state at the previous time step; based on the updated hidden representation at the current time step, output the predicted result of the three-dimensional displacement state of the lung tumor at the current time step relative to the reference respiratory phase through the state prediction module, wherein the three-dimensional displacement state includes left-right displacement components, front-back displacement components, and head-to-toe displacement components; pass the hidden representation at the current time step to the next time step to achieve dynamic tracking of the three-dimensional displacement state of the lung tumor under continuous respiratory phases.

2. The method for tracking lung tumors based on four-dimensional CT registration prior and projection temporal observation according to claim 1, characterized in that, The method for constructing the individualized tumor displacement prior sequence in step 1.1 is as follows: Step 2.1: Represent the sequence of previous four-dimensional CT multi-temporal images of the lungs of the same patient as follows: ,in These represent CT images at different respiratory phases within a complete respiratory cycle; [Selection] As a reference respiratory phase image, the tumor region was identified on the reference respiratory phase image. Outlining; Step 2.2: Perform deformable registration on CT images between adjacent respiratory phases, estimate the deformation field between adjacent phases, and obtain the set of deformation fields between adjacent phases within a complete respiratory cycle. : in, Indicates from the first The time phase to the first The deformation field of each phase, and the meanings of the other symbols are similar; Step 2.3: Using the reference breathing phase as a unified benchmark, obtain the deformation relationship from the reference breathing phase to any target breathing phase by cascading adjacent phase deformation fields. And using deformation field-driven spatial transformation operations to transform the tumor region under the reference respiratory phase. Propagated to the remaining time phases, the first... Tumor region at different time phases The propagation formula is as follows: in, Indicates the spread to the th Tumor regions after each time phase This represents a spatial transformation operation driven by a deformation field. This indicates the deformation relationship between the reference breathing phase and the target breathing phase; Step 2.4: Calculate the center location of the tumor region at each time phase, using the reference respiratory phase as a unified benchmark, and calculate the [missing information - likely a specific location or step]. The three-dimensional displacement of the tumor center relative to the reference respiratory phase tumor center at each time phase is calculated using the following formula: in, Indicates the first Tumor center coordinates at each time phase This indicates the coordinates of the tumor center at the reference respiratory phase. Indicates the first The three-dimensional displacement state of each phase. , , The corresponding displacement states at each time point were organized according to the respiratory time sequence to construct the tumor displacement prior sequence S. .

3. The method for tracking lung tumors based on 4DCT registration prior and projection temporal observation according to claim 1, characterized in that, The method for cross-domain observation enhancement processing in step 1.2 is as follows: Step 3.1: Obtain multi-temporal image sequences of the lungs from subsequent four-dimensional CT scans of the same patient, generate corresponding digitally reconstructed X-ray images using consistent projection geometry parameters, and obtain the original projection observation sequence. ; Step 3.2: The projection observation enhancement module includes a generator network for cross-domain mapping, a feature encoder for feature extraction, and a discriminator for distribution discrimination, which are used to realize cross-domain conversion from the digitally reconstructed X-ray image domain to the real X-ray style domain. Step 3.3: The projection observation enhancement module is optimized using a joint loss function that includes adversarial loss, block-level contrast loss, and local block-level differentiable correlation ratio constraint, the expression of which is as follows: in, This represents the joint loss due to the enhancement of projected observations. This represents the adversarial loss used to push the enhanced projected image to approximate the true X-ray domain in its overall distribution. This represents the block-level contrast loss used to maintain content consistency between the input projected image and the output enhanced image in corresponding local regions. This represents a local block-level differentiable correlation ratio constraint used to enhance the consistency of local statistical structure between the enhanced projected image and the target domain image. and represents the weights of the block-level contrast loss and the local block-level differentiable correlation ratio constraint, respectively, and their values ​​were determined through experimental parameter tuning. Indicates a generator; Indicates the discriminator; Represents the true X-ray image of the target domain; Represents the original projected image; This represents the enhanced projected image; Represents local block features; Represents the set of local block features of negative samples; This represents the contrastive learning loss function; Step 3.4: The local block-level differentiable correlation ratio constraint is obtained by calculating the symmetric differentiable correlation ratio loss at the local block level; assuming the enhanced projection image... Compared with the target domain real X-ray image The For local blocks respectively and Its symmetric differentiable correlation ratio loss is defined as: in, Represents the differentiable correlation ratio; for the division The local block-level differentiable correlation ratio constraint is calculated by averaging the symmetric differentiable correlation ratio loss over the local blocks: in, This represents the total number of pairs of local blocks in the partition. Step 3.5: Using the projection observation enhancement module trained with the joint loss function, the input original projection observation sequence is mapped to output an enhanced projection observation sequence that more closely approximates the true X-ray domain representation. .

Images