Optimization design method for afterloading 3D printing source applicator based on deep learning

Abstract

The invention relates to the technical field of source applicator dose optimization, in particular to an after-loading 3D printing source applicator optimization design method based on deep learning, which comprises the following steps: S1, acquiring medical image data and clinical dose constraint conditions of a target object, and calculating the clinical dose of the target object based on the medical image data; constructing a three-dimensional anatomical topology network reflecting the spatial relationship and attribute of the anatomical structure; and S2, through the three-dimensional anatomical topology network, through an adaptive structure generation process, determining an initial source applicator biological form configuration and a structure parameter set thereof. An initial source applicator parameter is generated from a medical image through a deep learning model, and optimization is performed in combination with clinical dose constraints, so that the optimal dose is achieved while the final model is ensured to fit an anatomical structure of a patient.

Description

Technical Field

[0001] This invention relates to the field of applicator dosage optimization technology, specifically to a deep learning-based method for optimizing the design of post-loading 3D printed applicators. Background Technology

[0002] In brachytherapy for tumors, the applicator is a key device for delivering radiation to the vicinity of the target area. Currently, the standard applicator based on a limited number of geometric templates is widely used in clinical practice. When implanting the applicator, doctors select the model based on experience and manually adjust its position. This design has inherent limitations. First, its fixed shape is difficult to fully fit the complex and highly individualized anatomical cavities of the patient, resulting in unstable applicator position and inaccurate relative spatial relationship between the radiation source channel and the target area and organs at risk, affecting the accuracy of dose calculation. Second, the channel layout of the standard applicator is preset and cannot be optimized for the tumor shape and surrounding sensitive organs of a specific patient, often leading to uneven dose distribution in the target area or unnecessary irradiation of organs at risk.

[0003] In recent years, with the development of medical imaging and 3D printing technology, existing methods usually reconstruct the patient's anatomical model based on medical images. Then, doctors or engineers manually or semi-manually design the applicator shape and plan the channels in computer-aided design software. This method is highly dependent on the operator's experience and skills. The design process is time-consuming and lacks systematic optimization. Designers often find it difficult to simultaneously consider multiple conflicting goals. They need to ensure that the applicator shape perfectly fits the irregular anatomical structure to ensure stability, while also optimizing the dose distribution generated at the radiation source residence point through the layout of the internal channels. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a deep learning-based optimized design method for post-loading 3D printing applicators.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a deep learning-based post-loading 3D printing applicator optimization design method, comprising the following steps: S1. Obtain medical imaging data and clinical dose constraints of the target object, and construct a three-dimensional anatomical topology network reflecting the spatial relationships and attributes of anatomical structures based on the medical imaging data; S2. Through a three-dimensional anatomical topology network and an adaptive structure generation process, the initial donor biomorphic configuration and its set of structural parameters are determined. S3. Based on clinical dose constraints and applicator morphology, perform multi-objective coordination to determine the optimized applicator adaptation configuration and channel layout. Based on the optimized applicator adaptation configuration and channel layout, generate manufacturing instructions for 3D printing.

[0006] In a preferred embodiment, in step S1, medical image data of the target object is acquired, and based on the medical image data, the acquired medical image data is preprocessed. The preprocessing includes adjusting the window width and window level of the image to optimize the display contrast and performing noise filtering to enhance the signal-to-noise ratio. Iterate through all voxels in the entire medical image data. For each voxel that has not yet been assigned to any region, check if its intensity value is higher than a preset global intensity threshold. If it is higher than the preset global intensity threshold, mark the voxel as an initial seed point. At the same time, create an empty list to record the set of candidate regions that are currently growing, and create a labeling matrix with the same size as the image data to initially mark all voxels as unvisited. Starting from the unvisited initial seed point, perform the following iterative operations: S11. Using the initial seed point as the core, create a new current growth region and add it to the candidate region set. Mark the initial seed point as belonging to the current region. S12. Check all spatial neighbor voxels that have been assigned to the voxels of the current growth region. For each neighbor voxel, if it is unvisited, proceed to S13 for judgment; otherwise, ignore the neighbor voxel. S13. Calculate the difference between the intensity value of the neighboring voxel and the average intensity of all voxels in the current growth region. If the absolute value of the difference is less than or equal to the preset intensity similarity tolerance, then determine that the neighboring voxel has intensity similarity with the current region. S14. Add the neighboring voxels determined by intensity similarity to the current growth region and mark them as belonging to the current region; S15. Repeat S12-S14, continuously exploring the neighborhood of the latest voxel added to the current growth region, making similarity judgments and merging them, until no new voxel is added to the current growth region in a complete neighborhood exploration cycle. At this time, the growth of the current growth region terminates. Record the current growth region where growth terminates as a candidate continuous region. Then find the next seed point that is still marked as unvisited and use the unvisited seed point as the new seed. Repeat the entire process of iteration to generate the next candidate continuous region. After traversing all voxels, output an initial segmentation result consisting of multiple candidate continuous regions. Based on a pre-defined anatomical atlas, each candidate contiguous region is classified and labeled. The pre-defined anatomical atlas defines the typical spatial location, morphological features, and relative relationships of different anatomical tissues in the target image coordinate system, and the following steps are performed: The geometric center and spatial range of each candidate continuous region are compared with the expected location range of each standard anatomical structure defined in the preset anatomical atlas. For each candidate region, its shape features and its proximity relationship with the preliminary judgment region are extracted. If the spatial location of a candidate region highly coincides with the expected location of a natural cavity in a pre-defined anatomical atlas, and its shape is tubular or potential channel-like, then the candidate region is marked as a natural cavity region. If the spatial location of a candidate region matches the expected location of the target region in the preset anatomical atlas, and its shape is usually clump-like, then the candidate region is marked as the target region. If the spatial location of a candidate region matches the expected location of an organ at risk in a preset anatomical atlas, then it is marked as the corresponding organ at risk region according to the specific definition of the preset anatomical atlas. The classified and labeled regions are formed into a structured dataset, which includes the identified and categorized natural cavity regions, target regions, and one or more organ-at-risk regions. Based on the identified and labeled anatomical structures, a three-dimensional anatomical topology network is constructed. Several points are extracted on the central path of the natural cavity region as first-class nodes, and several representative points are extracted on the surface or inside the target area and the organ at risk region as second-class nodes. Nodes with spatial proximity are connected by connecting edges, and each connecting edge is assigned a parameter as its first mechanical property. Among all the first type of nodes in the three-dimensional anatomical topology network, a specific node is selected according to the preset path starting point rule and defined as the initial growth point. The goal of the first morphology generation is to maximize the number of nodes in the contact network. Based on the connecting edges and their first mechanical properties, the generated structure's orientation conforms to the first mechanical property as the second morphology generation target. An adaptive structure generation process is then initiated, taking the initial growth point, the first morphology generation target, and the second morphology generation target as inputs. Specifically, this process includes the following steps: An initial structural unit is generated at the spatial location of the initial growth point. Starting from the initial structural unit, the branch structure is iteratively expanded within the constraint space defined by the three-dimensional anatomical topology network. In each expansion iteration, each candidate expansion direction is evaluated based on the current structural state, the comprehensive contribution of each direction to achieving the first form generation goal and the second form generation goal is calculated, and the direction with the highest comprehensive contribution is selected to perform the actual expansion. When the adaptive structure generation process meets the preset termination condition, the iteration stops and the final generated three-dimensional branch structure is output. The three-dimensional branch structure includes the trunk and branches defined by the spatial coordinate sequence, which is defined as the initial donor morphological configuration. All geometric parameters of the initial donor morphological configuration constitute the initial structure parameter set.

[0007] In a preferred embodiment, in step S2, at the spatial coordinate position of the initial growth point, an initial structural unit of a preset basic size is generated, a dynamic structural record list is created, the geometric parameters of the initial structural unit are stored in the dynamic structural record list as the first record, and the leading geometric feature points of the initial structural unit are marked as the current growth front set. Based on the dynamic structure record list and the current growth front set, an adaptive structure generation process is performed, which includes the following steps: S21. For each growth front point in the current growth front set, find all connecting edges in the three-dimensional anatomical topology network that start from the node where the growth front point is located and have an angle less than a preset angle with the tangent direction triggered from the growth front point. Take the spatial direction indicated by the connecting edge as a set of candidate expansion directions. S22. For each candidate extension direction, perform virtual extension along the candidate extension direction. After moving a preset standard step distance, the point reached in three-dimensional space is taken as the endpoint of the virtual extension line. Sequentially obtain the three-dimensional spatial coordinates of each node in the three-dimensional anatomical topology network. For each node in the network, calculate the difference between the X-axis coordinate of the node and the X-axis coordinate of the endpoint of the virtual extension line, the difference between the Y-axis coordinate of the node and the Y-axis coordinate of the endpoint, and the difference between the Z-axis coordinate of the node and the Z-axis coordinate of the endpoint. Square the three differences respectively and add them together. Then, perform a square root operation on the sum. The final value is the spatial distance between the node and the endpoint of the virtual extension line. The number of nodes whose spatial distance is less than the preset contact tolerance is counted, and the total number of nodes that the current growth structure has already contacted is subtracted from the number of nodes. The difference is used as the contribution value of the candidate expansion direction to the new contact node of the first morphology generation target. S23. For the same candidate expansion direction, obtain the first mechanical property value of the connecting edge corresponding to the candidate expansion direction. According to a preset mapping relationship, the first mechanical attribute value is converted into a compliance score representing the degree of compliance. Specifically, a scoring reference range is set, which has a preset upper limit and a lower limit. The first mechanical attribute value is compared with the upper limit and the lower limit of the scoring reference range. If the first mechanical attribute value is greater than or equal to the upper limit, the compliance score is determined as a first preset high score. If the first mechanical attribute value is less than or equal to the lower limit, the compliance score is determined as a first preset low score. If the first mechanical attribute value is between the upper limit and the lower limit, the compliance score is determined by linear difference calculation based on its relative position in the scoring reference range. The compliance score is the contribution value of the candidate expansion direction to the second morphology generation target. S24. For the same candidate expansion direction, the contribution value of the new contact node and the compliance score are weighted and summed. The result is the comprehensive contribution of the candidate expansion direction. The comprehensive contribution of all candidate expansion directions is compared and the direction with the highest comprehensive contribution is selected as the actual expansion direction for this iteration. S25. Along the actual expansion direction, with preset compensation and preset cross-sectional diameter, generate a new three-dimensional geometric structure segment. The starting coordinates of the new three-dimensional geometric structure segment are the three-dimensional coordinates of the current growth front point that triggered this iteration. The calculation of the ending coordinates needs to start from the starting coordinates, along the actual expansion direction determined in this iteration, with a preset step size as the moving distance, and add the components of the starting coordinates in the X, Y, and Z directions to the product of the components of the actual expansion direction vector in the corresponding directions and the preset step size, so as to obtain a new three-dimensional coordinate point. The new three-dimensional coordinate point is the ending coordinate. Add the spatial starting and ending coordinates of the new three-dimensional geometric structure segment as a new record to the dynamic structure record list. S26. Update the endpoint of the new three-dimensional geometric structure segment to the new growth front point in the current growth front set, and remove the growth front state at the starting point of this iteration. The adaptive structure generation step is repeated. When the total length of the geometric structure segments in the dynamic structure record list reaches the preset upper limit, the adaptive structure generation step is terminated. Based on the spatial start and end coordinates of all geometric structure segments recorded in the dynamic structure record list, a complete three-dimensional connected skeleton is reconstructed and output as the initial donor morphology configuration. At the same time, all geometric parameters in the dynamic structure record list are summarized to form the structural parameter set of the initial donor morphology configuration.

[0008] In a preferred embodiment, in step S3, clinical dose constraints, initial applicator morphology configuration, set of structural parameters, and three-dimensional anatomical topology network are obtained. On the three-dimensional anatomical topology network, a dynamic dose is initialized using the channel position of the initial applicator morphology configuration as the dose release source, and its initial propagation parameters are set. A coordinated iterative process is executed, in which the following operations are performed synchronously in each iteration: Dose assessment difference: Based on the current propagation parameters of dynamic dose, the spatial location distribution of dose release sources defined in the dynamic dose on the three-dimensional anatomical topology network and the current intensity parameters of each dose release source are obtained. Then, the process of dose energy spreading from each dose release source point to surrounding nodes along the connecting edge network is simulated. By simulating this process, a cumulative received dose energy value is calculated for each node in the three-dimensional anatomical topology network. The set of dose energy values ​​of all nodes constitutes the current dose distribution. Obtain the current dose energy value of each node in the three-dimensional anatomical topology network, read the expected dose value corresponding to the same node from the dose constraint, and for each node, subtract its expected dose value from the current dose energy value of the node to obtain the dose difference of the node. After calculating the dose difference for all network nodes, the absolute values ​​of the dose differences for all nodes are first taken and summed to obtain the total absolute difference. Then, a preset first threshold is called. If the total absolute difference is less than or equal to the first threshold, the result of the dose difference meeting the requirements is output. If the total absolute difference is greater than the first threshold, the result of the dose difference not meeting the requirements is output. Mechanical impact assessment: For each connecting edge in the three-dimensional anatomical topology network, it is determined whether the line segment of the connecting edge in three-dimensional space intersects with the geometric surface model of the applicator configuration in space. If the above spatial condition is met, it is considered that the connecting edge may be affected by the mechanical impact of the applicator configuration. The first mechanical attribute value of all affected connecting edges is compared with a preset mechanical impact risk threshold one by one. If the first biomechanical attribute value of any affected connecting edge is lower than the preset mechanical impact risk threshold, it means that the tissue mechanical tolerance is poor and is initially judged to have an unacceptable mechanical risk. If the attribute values ​​of all affected connecting edges are higher than the preset mechanical impact risk threshold, the mechanical impact is judged to be acceptable. Based on the difference between the dose difference assessment results and the preset first requirement, adjustment instructions for the morphological parameters of the applicator configuration and the propagation parameters of the dynamic dose are generated. At the same time, based on the difference between the mechanical influence assessment results and the preset second requirement, adjustment instructions for the morphological parameters of the applicator configuration are generated. The adjustment instructions are executed in combination to update the morphological parameters of the applicator configuration and the propagation parameters of the dynamic dose. Repeat the coordination and iteration process until the difference between the current dose distribution and the clinical dose constraint meets the first requirement, and the mechanical influence of the applicator configuration meets the first requirement. Then, take the updated applicator configuration as the optimized applicator adaptation configuration, lock the dose release source distribution and propagation path of the current dynamic dose, and convert them into three-dimensional channel centerline coordinates as the final channel layout. Based on the optimized bio-adaptive configuration of the applicator and the final channel layout, a unified three-dimensional digital model is integrated, and the unified three-dimensional data model is 3D printed.

[0009] The beneficial effects of this invention are: by constructing a three-dimensional anatomical structure network of the patient and using it to drive adaptive structure generation and coordinated optimization, this invention effectively overcomes the problems of mismatch between traditional standard applicators and individual anatomical structures and the inability of manual customization to simultaneously optimize biomechanical adaptability. Initial applicator parameters are generated from medical images using a deep learning model and optimized by combining them with clinical dose constraints. This ensures that the final model achieves the optimal dose while conforming to the patient's anatomy, improving treatment preparation efficiency. Furthermore, by directly outputting 3D printing instructions, real-time conversion between the design and the physical device is guaranteed. Attached Figure Description

[0010] Figure 1 This is a flowchart of the present invention; Figure 2 This is an iterative flowchart of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0013] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0014] like Figure 1 This embodiment provides a deep learning-based method for optimizing the design of post-loading 3D printing applicators, including the following steps: S1. Obtain medical imaging data and clinical dose constraints of the target object, and construct a three-dimensional anatomical topology network reflecting the spatial relationships and attributes of anatomical structures based on the medical imaging data; Furthermore, in S1, medical image data of the target object is acquired, and based on the medical image data, the acquired medical image data is preprocessed. The preprocessing includes adjusting the window width and window level of the image to optimize the display contrast and performing noise filtering to enhance the signal-to-noise ratio. Iterate through all voxels in the entire medical image data. For each voxel that has not yet been assigned to any region, check if its intensity value is higher than a preset global intensity threshold. If it is higher than the preset global intensity threshold, mark the voxel as an initial seed point. At the same time, create an empty list to record the set of candidate regions that are currently growing, and create a labeling matrix with the same size as the image data to initially mark all voxels as unvisited. Starting from the unvisited initial seed point, perform the following iterative operations: S11. Using the initial seed point as the core, create a new current growth region and add it to the candidate region set. Mark the initial seed point as belonging to the current region. S12. Check all spatial neighbor voxels (e.g., six-neighbor or twenty-six-neighbor) that have been assigned to the voxel in the current growth region. For each neighbor voxel, if it is unvisited, proceed to S13 for judgment; otherwise, ignore the neighbor voxel. S13. Calculate the difference between the intensity value of the neighboring voxel and the average intensity of all voxels in the current growth region. If the absolute value of the difference is less than or equal to the preset intensity similarity tolerance, then determine that the neighboring voxel has intensity similarity with the current region. The preset intensity similarity tolerance is the maximum allowable intensity deviation for determining whether a neighboring voxel can be included in the current growth region during the region growth process. In medical imaging, the maximum expected range of intensity fluctuation for voxels belonging to uniform homogeneous anatomical tissues within a local area is determined by the preset intensity similarity tolerance. This tolerance controls the consistency of region growth and the precision of segmentation, ensuring that the intensity values ​​of all voxels merged into a unified region are close to the average level of that region. This ensures that the generated candidate continuous regions maintain internal uniformity in image intensity. When the difference between the neighboring voxel and the average intensity of the region exceeds the preset intensity similarity tolerance, growth stops. Therefore, the preset intensity similarity tolerance directly determines the boundary position of region growth and is key to distinguishing the boundaries of different tissues (e.g., tumors and normal tissues, different organs). The preset intensity similarity tolerance is determined by selecting a known, typical sample region in the image that represents a single tissue (such as muscle or soft tissue tumor), calculating the standard deviation of the intensity of all voxels in that region, and multiplying the standard deviation by a coefficient, typically 2-3 times, as an empirical value for the preset intensity similarity tolerance.

[0015] S14. Add the neighboring voxels determined by intensity similarity to the current growth region and mark them as belonging to the current region; S15. Repeat S12-S14, continuously exploring the neighborhood of the latest voxel added to the current growth region, making similarity judgments and merging them, until no new voxel is added to the current growth region in a complete neighborhood exploration cycle. At this time, the growth of the current growth region terminates. Record the current growth region where growth terminates as a candidate continuous region. Then find the next seed point that is still marked as unvisited and use the unvisited seed point as the new seed. Repeat the entire process of iteration to generate the next candidate continuous region. After traversing all voxels, output an initial segmentation result consisting of multiple candidate continuous regions. Based on a pre-defined anatomical atlas, each candidate contiguous region is classified and labeled. The pre-defined anatomical atlas defines the typical spatial location, morphological features, and relative relationships of different anatomical tissues in the target image coordinate system, and the following steps are performed: The geometric center and spatial range of each candidate continuous region are compared with the expected location range of each standard anatomical structure defined in the preset anatomical atlas. For each candidate region, its shape features and its proximity relationship with the preliminary judgment region are extracted. If the spatial location of a candidate region highly coincides with the expected location of a natural cavity (such as the uterine canal or vaginal fornix) in a pre-defined anatomical atlas, and its shape is tubular or potential channel-like, then the candidate region is marked as a natural cavity region. If the spatial location of a candidate region matches the expected location of a target region (such as a tumor) in a pre-defined anatomical atlas, and its shape is usually clumpy, then the candidate region is marked as the target region. If the spatial location of a candidate region matches the expected location of an organ at risk (such as the bladder, rectum, or sigmoid colon) in a preset anatomical atlas, then it is marked as the corresponding organ at risk region according to the specific definition of the preset anatomical atlas. The classified and labeled regions are formed into a structured dataset, which includes the identified and categorized natural cavity regions, target regions, and one or more organ-at-risk regions. Based on the identified and labeled anatomical structures, a three-dimensional anatomical topology network is constructed. Several points are extracted on the central path of the natural cavity region as first-class nodes, and several representative points are extracted on the surface or inside the target area and the organ at risk region as second-class nodes. Nodes with spatial proximity are connected by connecting edges, and each connecting edge is assigned a parameter as its first mechanical property. Among all the first type of nodes in the three-dimensional anatomical topology network, a specific node is selected according to the preset path starting point rule and defined as the initial growth point. The goal of the first morphology generation is to maximize the number of nodes in the contact network. Based on the connecting edges and their first mechanical properties, the generated structure's orientation conforms to the first mechanical property as the second morphology generation target. An adaptive structure generation process is then initiated, taking the initial growth point, the first morphology generation target, and the second morphology generation target as inputs. Specifically, this process includes the following steps: An initial structural unit is generated at the spatial location of the initial growth point. Starting from the initial structural unit, the branch structure is iteratively expanded within the constraint space defined by the three-dimensional anatomical topology network. In each expansion iteration, each candidate expansion direction is evaluated based on the current structural state, the comprehensive contribution of each direction to achieving the first form generation goal and the second form generation goal is calculated, and the direction with the highest comprehensive contribution is selected to perform the actual expansion. When the adaptive structure generation process meets the preset termination condition, the iteration stops and the final generated three-dimensional branch structure is output. The three-dimensional branch structure includes the trunk and branches defined by the spatial coordinate sequence, which is defined as the initial donor morphological configuration. All geometric parameters of the initial donor morphological configuration constitute the initial structure parameter set.

[0016] The pre-defined anatomical atlas is a predefined set of standard anatomical knowledge about human body parts. The pre-defined anatomical atlas contains all the key structures that need to be identified in the target treatment area (such as endovascular treatment), such as the uterine canal, natural cavities, tumor target area, bladder, rectum, and sigmoid colon. It also describes the approximate location, spatial distribution range, and normal morphological variation range of each standard anatomical structure in three-dimensional space, as well as the intensity and texture features of each anatomical structure in specific modal medical images. Furthermore, it defines the inherent spatial positional relationships between the various structures, such as the rectum being located behind the cervix. By comparing the spatial location and morphological features of each candidate region with the standard structures defined in the preset anatomical atlas, it is possible to identify and classify them into natural cavity regions, target regions, and regions of organs at risk. The preset anatomical atlas can quantify the average spatial location, morphological variation data, and probability distribution of the relative positions of each anatomical structure in the population by using machine learning training or statistical analysis on a large amount of historical patient medical image data annotated by clinical experts.

[0017] S2. Through a three-dimensional anatomical topology network and an adaptive structure generation process, the initial donor biomorphic configuration and its set of structural parameters are determined. Furthermore, in S2, at the spatial coordinate position of the initial growth point, an initial structural unit of a preset basic size is generated, a dynamic structural record list is created, the geometric parameters of the initial structural unit are stored in the dynamic structural record list as the first record, and the leading geometric feature points of the initial structural unit are marked as the current growth front set. Based on the dynamic structure record list and the current growth front set, an adaptive structure generation process is performed, which includes the following steps: S21. For each growth front point in the current growth front set, find all connecting edges in the three-dimensional anatomical topology network that start from the node where the growth front point is located and have an angle less than a preset angle with the tangent direction triggered from the growth front point. Take the spatial direction indicated by the connecting edge as a set of candidate expansion directions. S22. For each candidate extension direction, perform virtual extension along the candidate extension direction. After moving a preset standard step distance, the point reached in three-dimensional space is taken as the endpoint of the virtual extension line. Sequentially obtain the three-dimensional spatial coordinates of each node in the three-dimensional anatomical topology network. For each node in the network, calculate the difference between the X-axis coordinate of the node and the X-axis coordinate of the endpoint of the virtual extension line, the difference between the Y-axis coordinate of the node and the Y-axis coordinate of the endpoint, and the difference between the Z-axis coordinate of the node and the Z-axis coordinate of the endpoint. Square the three differences respectively and add them together. Then, perform a square root operation on the sum. The final value is the spatial distance between the node and the endpoint of the virtual extension line. The number of nodes whose spatial distance is less than the preset contact tolerance is counted, and the total number of nodes that the current growth structure has already contacted is subtracted from the number of nodes. The difference is used as the contribution value of the candidate expansion direction to the new contact node of the first morphology generation target. The preset contact tolerance is a positive distance threshold used to determine whether the spatial distance is close enough. If the spatial distance between a point on the applicator structure and a node in the anatomical topology network is less than or equal to this threshold, it is considered that the applicator structure has contacted the node in the anatomical topology network and the spatial position of the node in the anatomical topology network has been effectively covered. It is preset based on the physical size of the applicator and the clinical fitting accuracy requirements.

[0018] S23. For the same candidate expansion direction, obtain the first mechanical property value of the connecting edge corresponding to the candidate expansion direction. According to a preset mapping relationship, the first mechanical attribute value is converted into a compliance score representing the degree of compliance. Specifically, a scoring reference range is set, which has a preset upper limit and a lower limit. The first mechanical attribute value is compared with the upper limit and the lower limit of the scoring reference range. If the first mechanical attribute value is greater than or equal to the upper limit, the compliance score is determined as a first preset high score. If the first mechanical attribute value is less than or equal to the lower limit, the compliance score is determined as a first preset low score. If the first mechanical attribute value is between the upper limit and the lower limit, the compliance score is determined by linear difference calculation based on its relative position in the scoring reference range. The compliance score is the contribution value of the candidate expansion direction to the second morphology generation target. S24. For the same candidate expansion direction, the contribution value of the new contact node and the compliance score are weighted and summed. The result is the comprehensive contribution of the candidate expansion direction. The comprehensive contribution of all candidate expansion directions is compared and the direction with the highest comprehensive contribution is selected as the actual expansion direction for this iteration. S25. Along the actual expansion direction, with preset compensation and preset cross-sectional diameter, generate a new three-dimensional geometric structure segment. The starting coordinates of the new three-dimensional geometric structure segment are the three-dimensional coordinates of the current growth front point that triggered this iteration. The calculation of the ending coordinates needs to start from the starting coordinates, along the actual expansion direction determined in this iteration, with a preset step size as the moving distance, and add the components of the starting coordinates in the X, Y, and Z directions to the product of the components of the actual expansion direction vector in the corresponding directions and the preset step size, so as to obtain a new three-dimensional coordinate point. The new three-dimensional coordinate point is the ending coordinate. Add the spatial starting and ending coordinates of the new three-dimensional geometric structure segment as a new record to the dynamic structure record list. S26. Update the endpoint of the new three-dimensional geometric structure segment to the new growth front point in the current growth front set, and remove the growth front state at the starting point of this iteration. The adaptive structure generation step is repeated. When the total length of the geometric structure segments in the dynamic structure record list reaches the preset upper limit, the adaptive structure generation step is terminated. Based on the spatial start and end coordinates of all geometric structure segments recorded in the dynamic structure record list, a complete three-dimensional connected skeleton is reconstructed and output as the initial donor morphology configuration. At the same time, all geometric parameters in the dynamic structure record list are summarized to form the structural parameter set of the initial donor morphology configuration.

[0019] The preset upper limit is a maximum allowable value for limiting the total length of the applicator geometry. An excessively long applicator may increase the difficulty of implantation, the risk of patient discomfort, or the risk of unexpected bending or displacement in the body. The preset upper limit is a clinically recognized safe and operable length range. In brachytherapy, the dose distribution has the characteristics of high dose at close range and rapid drop at long range. Setting the preset upper limit helps to focus the design optimization on the most effective distance range around the core target area, avoiding excessive consumption of structural complexity in the distal region where the dose contribution is weak. The preset upper limit is based on the standard treatment procedure and typical anatomical dimensions of the target treatment site. By conducting preliminary dosimetric simulations on typical cases, the impact of different applicator lengths on target area dose coverage and organ-at-risk dose is evaluated, and a risk-controllable length value is selected as the preset upper limit.

[0020] S3. Based on clinical dose constraints and applicator morphology, perform multi-objective coordination to determine the optimized applicator adaptation configuration and channel layout. Based on the optimized applicator adaptation configuration and channel layout, generate manufacturing instructions for 3D printing. Furthermore, in S3, the clinical dose constraint conditions, the initial applicator morphology configuration, the set of structural parameters, and the three-dimensional anatomical topology network are obtained. On the three-dimensional anatomical topology network, the channel position of the initial applicator morphology configuration is used as the dose release source to initialize a dynamic dose and set its initial propagation parameters. A coordinated iterative process is executed, in which the following operations are performed synchronously in each iteration: Dose assessment difference: Based on the current propagation parameters of dynamic dose, the spatial location distribution of dose release sources defined in the dynamic dose on the three-dimensional anatomical topology network and the current intensity parameters of each dose release source are obtained. Then, the process of dose energy spreading from each dose release source point to surrounding nodes along the connecting edge network is simulated. By simulating this process, a cumulative received dose energy value is calculated for each node in the three-dimensional anatomical topology network. The set of dose energy values ​​of all nodes constitutes the current dose distribution. Obtain the current dose energy value of each node in the three-dimensional anatomical topology network, read the expected dose value corresponding to the same node from the dose constraint, and for each node, subtract its expected dose value from the current dose energy value of the node to obtain the dose difference of the node. After calculating the dose difference for all network nodes, the absolute values ​​of the dose differences for all nodes are first taken and summed to obtain the total absolute difference. Then, a preset first threshold is called. If the total absolute difference is less than or equal to the first threshold, the result of the dose difference meeting the requirements is output. If the total absolute difference is greater than the first threshold, the result of the dose difference not meeting the requirements is output. The preset first threshold is the maximum upper limit that the total absolute difference calculated in dose difference assessment is not allowed to exceed. It serves as the core quantitative indicator for judging whether the difference between the current dose distribution and the dose constraint is acceptable. When the total absolute difference calculated iteratively is lower than the preset first threshold, the dose target is considered to have been achieved. The preset first threshold is mainly based on the dosimetric standards and clinical practice experience of clinical radiotherapy. By analyzing the statistical deviation between the dose distribution and the ideal target of a large number of successful treatment plans, a typical deviation tolerance value is determined as the preset first threshold.

[0021] Mechanical impact assessment: For each connecting edge in the three-dimensional anatomical topology network, it is determined whether the line segment of the connecting edge in three-dimensional space intersects with the geometric surface model of the applicator configuration in space. If the above spatial condition is met, it is considered that the connecting edge may be affected by the mechanical impact of the applicator configuration. The first mechanical attribute value of all affected connecting edges is compared with a preset mechanical impact risk threshold one by one. If the first biomechanical attribute value of any affected connecting edge is lower than the preset mechanical impact risk threshold, it means that the tissue mechanical tolerance is poor and is initially judged to have an unacceptable mechanical risk. If the attribute values ​​of all affected connecting edges are higher than the preset mechanical impact risk threshold, the mechanical impact is judged to be acceptable. The preset mechanical impact risk threshold is the safe upper limit that the estimated mechanical impact of the applicator configuration on the connection edge is not allowed to exceed in the mechanical impact assessment. This threshold is set based on clinical safety guidelines.

[0022] Based on the difference between the dose difference assessment results and the preset first requirement, adjustment instructions for the morphological parameters of the applicator configuration and the propagation parameters of the dynamic dose are generated. At the same time, based on the difference between the mechanical influence assessment results and the preset second requirement, adjustment instructions for the morphological parameters of the applicator configuration are generated. The adjustment instructions are executed in combination to update the morphological parameters of the applicator configuration and the propagation parameters of the dynamic dose. Repeat the coordination and iteration process until the difference between the current dose distribution and the clinical dose constraint meets the first requirement, and the mechanical influence of the applicator configuration meets the first requirement. Then, take the updated applicator configuration as the optimized applicator adaptation configuration, lock the dose release source distribution and propagation path of the current dynamic dose, and convert them into three-dimensional channel centerline coordinates as the final channel layout. Based on the optimized bio-adaptive configuration of the applicator and the final channel layout, a unified three-dimensional digital model is integrated, and the unified three-dimensional data model is 3D printed.

[0023] The preset first requirement defines the overall termination condition in the dose dimension of the target coordination process. The dose target is considered to be met only when all aspects of the dose assessment difference meet the first requirement. The preset first requirement is set in advance by medical experts based on treatment guidelines and clinical goals. The preset second requirement is a criterion for determining whether the mechanical influence of the applicator configuration is sufficiently safe. The preset second requirement is defined based on medical device safety standards and clinical experience.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make further changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A deep learning-based method for optimizing the design of post-loading 3D printing applicators, characterized in that, Includes the following steps: S1. Obtain medical imaging data and clinical dose constraints of the target object, and construct a three-dimensional anatomical topology network reflecting the spatial relationships and attributes of anatomical structures based on the medical imaging data; S2. Through a three-dimensional anatomical topology network and an adaptive structure generation process, the initial donor biomorphic configuration and its set of structural parameters are determined. S3. Based on clinical dose constraints and applicator morphology, perform multi-objective coordination to determine the optimized applicator adaptation configuration and channel layout. Based on the optimized applicator adaptation configuration and channel layout, generate manufacturing instructions for 3D printing.

2. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 1, characterized in that, In step S1, medical image data of the target object is acquired. Based on the medical image data, the acquired medical image data is preprocessed. The preprocessing includes adjusting the window width and window level of the image to optimize the display contrast and performing noise filtering to enhance the signal-to-noise ratio. Iterate through all voxels in the entire medical image data. For each voxel that has not yet been assigned to any region, check if its intensity value is higher than a preset global intensity threshold. If it is higher than the preset global intensity threshold, mark the voxel as an initial seed point. At the same time, create an empty list to record the set of candidate regions that are currently growing, and create a labeling matrix with the same size as the image data to initially mark all voxels as unvisited.

3. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 2, characterized in that, Starting from the unvisited initial seed point, perform the following iterative operations: S11. Using the initial seed point as the core, create a new current growth region and add it to the candidate region set. Mark the initial seed point as belonging to the current region. S12. Check all spatial neighbor voxels that have been assigned to the voxels of the current growth region. For each neighbor voxel, if it is unvisited, proceed to S13 for judgment; otherwise, ignore the neighbor voxel. S13. Calculate the difference between the intensity value of the neighboring voxel and the average intensity of all voxels in the current growth region. If the absolute value of the difference is less than or equal to the preset intensity similarity tolerance, then determine that the neighboring voxel has intensity similarity with the current region. S14. Add the neighboring voxels determined by intensity similarity to the current growth region and mark them as belonging to the current region. S15. Repeat S12-S14, continuously exploring the neighborhood of the latest voxel added to the current growth region, making similarity judgments and merging them, until no new voxels are added to the current growth region in a complete neighborhood exploration cycle. At this point, the growth of the current growth region terminates.

4. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 3, characterized in that, Record the current growth region where growth terminates as a candidate continuous region. Then find the next seed point that is still marked as unvisited and use the unvisited seed point as the new seed. Repeat the entire process of iteration to generate the next candidate continuous region. After traversing all voxels, output an initial segmentation result consisting of multiple candidate continuous regions. Based on a pre-defined anatomical atlas, each candidate contiguous region is classified and labeled. The pre-defined anatomical atlas defines the typical spatial location, morphological features, and relative relationships of different anatomical tissues in the target image coordinate system, and the following steps are performed: The geometric center and spatial range of each candidate continuous region are compared with the expected location range of each standard anatomical structure defined in the preset anatomical atlas. For each candidate region, its shape features and its proximity relationship with the preliminary judgment region are extracted. If the spatial location of a candidate region highly coincides with the expected location of a natural cavity in a pre-defined anatomical atlas, and its shape is tubular or potential channel-like, then the candidate region is marked as a natural cavity region. If the spatial location of a candidate region matches the expected location of the target region in the preset anatomical atlas, and its shape is usually clump-like, then the candidate region is marked as the target region. If the spatial location of a candidate region matches the expected location of an organ at risk in a preset anatomical atlas, then it is marked as the corresponding organ at risk region according to the specific definition of the preset anatomical atlas. The classified and labeled regions are formed into a structured dataset, which includes the identified and categorized natural cavity regions, target regions, and one or more organ-at-risk regions. Based on the identified and labeled anatomical structures, a three-dimensional anatomical topology network is constructed. Several points are extracted on the central path of the natural cavity region as first-class nodes, and several representative points are extracted on the surface or inside the target area and the organ at risk region as second-class nodes. Nodes with spatial proximity are connected by connecting edges, and each connecting edge is assigned a parameter as its first mechanical property. Among all the first type of nodes in the three-dimensional anatomical topology network, a specific node is selected according to the preset path starting point rule and defined as the initial growth point. The goal of the first morphology generation is to maximize the number of nodes in the contact network.

5. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 4, characterized in that, Based on the connecting edges and their first mechanical properties, the generated structure's orientation conforms to the first mechanical property as the second morphology generation target. An adaptive structure generation process is then initiated, taking the initial growth point, the first morphology generation target, and the second morphology generation target as inputs. Specifically, this process includes the following steps: An initial structural unit is generated at the spatial location of the initial growth point. Starting from the initial structural unit, the branch structure is iteratively expanded within the constraint space defined by the three-dimensional anatomical topology network. In each expansion iteration, each candidate expansion direction is evaluated based on the current structural state, the comprehensive contribution of each direction to achieving the first form generation goal and the second form generation goal is calculated, and the direction with the highest comprehensive contribution is selected to perform the actual expansion. When the adaptive structure generation process meets the preset termination condition, the iteration stops and the final generated three-dimensional branch structure is output. The three-dimensional branch structure includes the trunk and branches defined by the spatial coordinate sequence, which is defined as the initial donor morphological configuration. All geometric parameters of the initial donor morphological configuration constitute the initial structure parameter set.

6. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 1, characterized in that, In step S2, at the spatial coordinates of the initial growth point, an initial structural unit of a preset basic size is generated, a dynamic structural record list is created, the geometric parameters of the initial structural unit are stored in the dynamic structural record list as the first record, and the leading geometric feature points of the initial structural unit are marked as the current growth front set.

7. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 6, characterized in that, Based on the dynamic structure record list and the current growth front set, an adaptive structure generation process is performed, which includes the following steps: S21. For each growth front point in the current growth front set, find all connecting edges in the three-dimensional anatomical topology network that start from the node where the growth front point is located and have an angle less than a preset angle with the tangent direction triggered from the growth front point. The spatial direction indicated by the connecting edge is taken as a set of candidate expansion directions. S22. For each candidate extension direction, perform virtual extension along the candidate extension direction. After moving a preset standard step distance, the point reached in three-dimensional space is taken as the endpoint of the virtual extension line. Sequentially obtain the three-dimensional spatial coordinates of each node in the three-dimensional anatomical topology network. For each node in the network, calculate the difference between the X-axis coordinate of the node and the X-axis coordinate of the endpoint of the virtual extension line, the difference between the Y-axis coordinate of the node and the Y-axis coordinate of the endpoint, and the difference between the Z-axis coordinate of the node and the Z-axis coordinate of the endpoint. Square the three differences respectively and add them together. Then, perform a square root operation on the sum. The final value is the spatial distance between the node and the endpoint of the virtual extension line. The number of nodes whose spatial distance is less than the preset contact tolerance is counted, and the total number of nodes that the current growth structure has already contacted is subtracted from the number of nodes. The difference is used as the contribution value of the candidate expansion direction to the new contact node of the first morphology generation target. S23. For the same candidate expansion direction, obtain the first mechanical property value of the connecting edge corresponding to the candidate expansion direction. According to a preset mapping relationship, the first mechanical attribute value is converted into a compliance score representing the degree of compliance. Specifically, a scoring reference range is set, which has a preset upper limit and a lower limit. The first mechanical attribute value is compared with the upper limit and the lower limit of the scoring reference range. If the first mechanical attribute value is greater than or equal to the upper limit, the compliance score is determined as a first preset high score. If the first mechanical attribute value is less than or equal to the lower limit, the compliance score is determined as a first preset low score. If the first mechanical attribute value is between the upper limit and the lower limit, the compliance score is determined by linear difference calculation based on its relative position in the scoring reference range. The compliance score is the contribution value of the candidate expansion direction to the second morphology generation target. S24. For the same candidate expansion direction, the contribution value of the new contact node and the compliance score are weighted and summed. The result is the comprehensive contribution of the candidate expansion direction. The comprehensive contribution of all candidate expansion directions is compared and the direction with the highest comprehensive contribution is selected as the actual expansion direction for this iteration. S25. Along the actual expansion direction, with preset compensation and preset cross-sectional diameter, generate a new three-dimensional geometric structure segment. The starting coordinates of the new three-dimensional geometric structure segment are the three-dimensional coordinates of the current growth front point that triggered this iteration. The calculation of the ending coordinates needs to start from the starting coordinates, along the actual expansion direction determined in this iteration, with a preset step size as the moving distance, and add the components of the starting coordinates in the X, Y, and Z directions to the product of the components of the actual expansion direction vector in the corresponding directions and the preset step size, so as to obtain a new three-dimensional coordinate point. The new three-dimensional coordinate point is the ending coordinate. Add the spatial starting and ending coordinates of the new three-dimensional geometric structure segment as a new record to the dynamic structure record list. S26. Update the endpoint of the new three-dimensional geometric structure segment to the new growth front point in the current growth front set, and remove the growth front state at the starting point of this iteration.

8. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 7, characterized in that, Repeat the adaptive structure generation step. When the total length of the geometric structure segments in the dynamic structure record list reaches the preset upper limit, the adaptive structure generation step is terminated. Based on the spatial start and end coordinates of all geometric structure segments recorded in the dynamic structure record list, a complete three-dimensional connected skeleton is reconstructed and output as the initial donor morphology configuration. At the same time, all geometric parameters in the dynamic structure record list are summarized to form the structural parameter set of the initial donor morphology configuration.

9. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 1, characterized in that, In S3, clinical dose constraints, initial applicator morphology configuration, set of structural parameters, and three-dimensional anatomical topology network are obtained. On the three-dimensional anatomical topology network, the channel position of the initial applicator morphology configuration is used as the dose release source to initialize a dynamic dose and set its initial propagation parameters. A coordinated iterative process is executed, in which the following operations are performed synchronously in each iteration: Dose assessment difference: Based on the current propagation parameters of dynamic dose, the spatial location distribution of dose release sources defined in the dynamic dose on the three-dimensional anatomical topology network and the current intensity parameters of each dose release source are obtained. Then, the process of dose energy spreading from each dose release source point to surrounding nodes along the connecting edge network is simulated. By simulating this process, a cumulative received dose energy value is calculated for each node in the three-dimensional anatomical topology network. The set of dose energy values ​​of all nodes constitutes the current dose distribution. Obtain the current dose energy value of each node in the three-dimensional anatomical topology network, read the expected dose value corresponding to the same node from the dose constraint, and for each node, subtract its expected dose value from the current dose energy value of the node to obtain the dose difference of the node. After calculating the dose difference for all network nodes, the absolute values ​​of the dose differences for all nodes are first taken and summed to obtain the total absolute difference. Then, a preset first threshold is called. If the total absolute difference is less than or equal to the first threshold, the result of the dose difference meeting the requirements is output. If the total absolute difference is greater than the first threshold, the result of the dose difference not meeting the requirements is output. Mechanical impact assessment: For each connecting edge in the three-dimensional anatomical topology network, it is determined whether the line segment of the connecting edge in three-dimensional space intersects with the geometric surface model of the applicator configuration in space. If the above spatial condition is met, it is considered that the connecting edge may be affected by the mechanical impact of the applicator configuration. The first mechanical attribute value of all affected connecting edges is compared with a preset mechanical impact risk threshold one by one. If the first biomechanical attribute value of any affected connecting edge is lower than the preset mechanical impact risk threshold, it means that the tissue mechanical tolerance is poor and is initially judged to have an unacceptable mechanical risk. If the attribute values ​​of all affected connecting edges are higher than the preset mechanical impact risk threshold, the mechanical impact is judged to be acceptable. Based on the difference between the dose difference assessment results and the preset first requirement, adjustment instructions for the morphological parameters of the applicator configuration and the propagation parameters of the dynamic dose are generated. At the same time, based on the difference between the mechanical influence assessment results and the preset second requirement, adjustment instructions for the morphological parameters of the applicator configuration are generated. The adjustment instructions are executed in combination to update the morphological parameters of the applicator configuration and the propagation parameters of the dynamic dose.

10. The deep learning-based post-loading 3D printing applicator optimization design method according to claim 9, characterized in that, Repeat the coordination and iteration process until the difference between the current dose distribution and the clinical dose constraint meets the first requirement, and the mechanical influence of the applicator configuration meets the first requirement. Then, take the updated applicator configuration as the optimized applicator adaptation configuration, lock the dose release source distribution and propagation path of the current dynamic dose, and convert them into three-dimensional channel centerline coordinates as the final channel layout. Based on the optimized bio-adaptive configuration of the applicator and the final channel layout, an integrated three-dimensional digital model is formed, and the integrated three-dimensional data model is 3D printed.

Images