A method for reconstructing the donor

Abstract

This invention belongs to the field of feature extraction technology for metal applicators under CT images, and discloses an applicator reconstruction method. It obtains CT data of each channel of the applicator through thin-slice CT scanning and fits the curve equations of each channel to establish an applicator template library based on thin-slice CT. An improved sparse cascade network model is established, which can quickly and accurately classify applicators, providing a basis for further matching the corresponding applicator channels in the applicator template library. Based on thin-slice CT, the CT-CT image registration between the CT data in the applicator template library and the CT data of clinical applicators is more efficient and faster. Using the applicator fitting curve established from the thin-slice CT data as a skeleton, the fitting curve of the center point set extracted from the CT data of the clinical applicator is optimized by using piecewise fitting, physical constraints, and other methods, thereby completing the accurate reconstruction of the applicator.

Description

Technical Field

[0001] This invention relates to the field of metal applicator feature extraction technology in CT images, and particularly to an applicator reconstruction method. Background Technology

[0002] Brachytherapy is a crucial treatment modality in radiotherapy, particularly for gynecological tumors. Different applicators are used in brachytherapy depending on the patient's condition and treatment objectives. In the modern radiotherapy planning stage, dosimeters typically need to manually mark the center point of each applicator on CT images; this process is called applicator reconstruction. However, applicator reconstruction is not only time-consuming and labor-intensive, but also prone to various deviations due to human subjectivity, such as center point deviation caused by metal artifacts or applicator tip deviation due to slice thickness. Furthermore, the Fletcher applicator, frequently used in gynecological tumor treatment, is made of metal, exhibiting a bright characteristic on CT images and significant metal artifact noise. Its clinical assembly requires the attachment of an oocyte applicator and a uterine cavity limiting clip, which also contain metal rings and bolts. These components often overlap with the bright characteristic of the applicator on CT images, requiring considerable experience and meticulousness from the dosimeter to accurately mark the center point of the applicator cannula, posing a potential risk to treatment safety.

[0003] The relevant literature (Zhang D, Yang Z, Jiang S, Zhou Z, Meng M, Wang W. Automatic segmentation and applicator reconstruction for CT-based brachytherapy of cervical cancer using 3D convolutional neural networks. J Appl Clin Med Phys. 2020 Oct;21(10):158-169. doi: 10.1002 / acm2.13024. Epub 2020 Sep 29. PMID:32991783; PMCID: PMC7592978) introduces an automatic organ and applicator reconstruction method based on three-dimensional convolutional neural networks. This article describes gynecological tumor cases treated with the clinically common Fletcher applicator. Each case underwent CT scanning, and the CT data was enhanced before model training. An improved 3D convolutional neural network was proposed to segment the input CT data into corresponding contour organs and the applicator. The applicator's automatic reconstruction accuracy was high, with Descein coefficients exceeding 88% and Jed coefficients exceeding 80%. However, the article does not mention the assembly details of the applicator used, specifically the assembly of the uterine cavity retaining clips within the uterine canal. The applicator reconstruction described in this article uses skeletalization and nonlinear fitting, a relatively simple method, but in practical use, it is difficult to distinguish metal bolts and other components, and certain reconstruction biases exist.

[0004] The method described in Chinese patent CN116152437A, "Applicator Reconstruction Method, Apparatus, Electronic Device, and Computer-Readable Storage Medium," first uses methods such as rotating and cutting CT images, and then applies a U-net network to obtain and generate a three-dimensional mask. This method does not mention the assembly of the applicator used, meaning it does not explain the assembly of the uterine cavity limiting clips within the uterine cavity canal. Furthermore, the method does not explain the identification and classification of the left and right oocytes during the reconstruction process, which could lead to potential applicator reconstruction errors. The method employs a three-dimensional mask, but does not explicitly state whether the mask is based on theoretical conditions or the actual applicator structure. Summary of the Invention

[0005] The purpose of this invention is to propose an applicator reconstruction method to achieve rapid and accurate reconstruction of the applicator used in clinical treatment.

[0006] The technical solution of the present invention is as follows: a method for reconstructing an applicator, which obtains thin-slice CT data of each channel of the applicator through thin-slice CT scanning, fits the curve equation of each channel, and establishes an applicator template library based on thin-slice CT.

[0007] An improved sparse cascade network model was established and trained using CT images of applicators under conventional scanning conditions. This model was then used to classify clinical applicators into single-tube or multi-tube types. In the case of multi-tube applicators, the tubes were split based on the classification results to obtain multiple single tubes. Single-tube matching was performed on applicators within the applicator template library based on thin-slice CT.

[0008] The CT images in the applicator template library based on thin-slice CT are registered with the CT images of clinical applicators using an improved physical structure-driven optimization multi-registration method. Using the applicator fitting curve established based on the thin-slice CT applicator template library as the skeleton, the fitting curve of the center point set extracted from the CT image of the clinical applicator is optimized by segmented fitting and physical constraint methods to complete the accurate reconstruction of the applicator.

[0009] The process of obtaining the applicator template library based on thin-slice CT is as follows: obtain thin-slice CT images of the applicator with only the tubular portion; the tubular portion includes the left oval tube, the right oval tube, the 15-degree uterine cavity tube, the 30-degree uterine cavity tube, the 45-degree uterine cavity tube, and the 60-degree uterine cavity tube;

[0010] Based on the above thin-slice CT images, the center point set extraction and curve fitting of each type of applicator were performed using hybrid geometric features; the above thin-slice CT images and fitting curve results were stored as a thin-slice CT-based applicator template library.

[0011] The specific implementation process of the donor center point set extraction and curve fitting based on hybrid geometric features is as follows:

[0012] Gaussian filtering noise reduction method is used to preprocess thin-slice CT images;

[0013] The three-dimensional center points of the applicator channel in thin-slice CT images were extracted using a dynamic thresholding method to obtain a set of center points.

[0014] Point cloud sorting is performed along the donor path, the endpoints of the donor center point set are identified, and the number of point sets is increased at the donor endpoints.

[0015] Perform spline function-based curve fitting on the above center point set to obtain the fitted curve function;

[0016] Based on the above fitted curve function, curvature analysis and segmentation are performed, and the segmentation equation is as follows;

[0017] (1)

[0018] The equation is a partitioning equation; It is the equation of the fitting curve of the three-dimensional center point of the source feeder pipeline; Let the arc length be , and They are The first and second derivatives;

[0019] Based on curve fitting of the source geometry, the three-segment smooth connection of the curve is enhanced, resulting in geometric features of straight line segment, curve segment, and straight line segment in sequence.

[0020] The thin-slice CT image is CT image data with a slice thickness of 0.6 mm.

[0021] The improved sparse cascaded network model includes a dynamic sparse activation module, a light branching module for simple regions, a deep branching module for complex regions, and a cascaded decision-maker module. The data processed by the deep branching and light branching modules are feature-concatenated. The concatenated data then passes through a cascaded decision-maker module; high-confidence results are output directly, while low-confidence results are output after passing through a transformer module. The input is a CT image of a clinical applicator. Applicator types include: single-tube applicators (15-degree, 30-degree, 45-degree, and 60-degree uterine cavity tubes); two-tube applicators (applicators used in combination with the left and right oocyte ducts); and three-tube applicators (applicators used in combination with the left and right oocyte ducts at various angles of the uterine cavity tube).

[0022] The input data of the improved sparse cascaded network model first passes through a dynamic sparse activation module; the dynamic sparse activation module generates a mask by a gated convolutional layer, with a dynamic threshold set to 30%, which is used to identify key regions.

[0023] In the above process, the mask generated by the gated convolutional layer is regarded as a weight map, which works together with the dynamic threshold to process the input data. After processing the weight map, two data points of the same size are output. They are arranged in descending order of weight value. The values ​​of points before 30% of the weight value are retained, and the values ​​of the remaining points are set to 0. This region is defined as a complex region. The values ​​of points after 30% of the weight value are retained, and the values ​​of the remaining points are set to 0. This region is defined as a simple region.

[0024] Simple regions output by the dynamic sparse activation module are used to extract features using the lightweight branching module; complex regions output by the dynamic sparse activation module are used to extract features using the deep branching module.

[0025] The features obtained from the lightweight branching module and the features obtained from the deep branching module are concatenated and then passed through a cascaded decision-maker module. The first stage of the cascaded decision-maker module is global average pooling connected to a fully connected layer, and the second stage is a Transformer encoder block. If the confidence score obtained after the first stage is ≥0.85, the classification result is given directly without going through the second stage. If the confidence score is <0.85, the final classification is completed by the Transformer encoder block in the second stage, and the classification result is output.

[0026] The registration process specifically involves: after preprocessing the CT image of the clinical applicator, matching the corresponding channel data in the applicator template library based on thin-slice CT according to the names of each channel of the applicator, and using a multi-registration method based on improved physical structure-driven optimization to complete the three-dimensional spatial registration between the CT image of the clinical applicator and the CT image of the corresponding applicator in the applicator template library based on thin-slice CT, thus establishing the correspondence between the two applicators in three-dimensional space.

[0027] The reconstruction of the clinical applicator specifically involves: calling the center point set fitting curve function of the applicator template library based on thin-slice CT; correcting the center point set on the CT image of the clinical applicator based on the above registration results; and creating the applicator outline on the CT image of the clinical applicator according to the physical size of the applicator to complete the reconstruction of the clinical applicator.

[0028] The specific implementation process of the multi-registration method based on improved physical structure-driven optimization is as follows:

[0029] Initial image registration: The CT images of the corresponding applicators in the applicator template library based on thin-slice CT were approximately registered with the clinical applicator CT images using the bone-based registration method.

[0030] The fitted curve obtained from the applicator template library based on thin-slice CT is used as a "shape skeleton" to guide the point set formed by the clinical applicator CT image to fit within the target range;

[0031] Clinical applicator center point set extraction and fitting, enhancing the three-segment structural features; the three-segment structural features are, in order, a straight segment, a curved segment, and a straight segment;

[0032] Weighted registration: For the enhancement weight of the curved segment, the registration of the curved segment is enhanced in the multi-segment registration, and finally the three-dimensional spatial registration of the clinical applicator and the corresponding applicator in the applicator template library based on thin-slice CT is realized. The objective function of the relevant registration is given by formula (2).

[0033] The contour reconstruction of the physical dimensions of the pipeline is performed by constrained fitting of the clinical applicator center point set;

[0034] (2)

[0035] in, These are the constraint weights of the line segments. This represents the distance from the point corresponding to the straight line segment extracted from the clinical applicator CT image after displacement to the straight line segment in the CT image based on the thin-slice CT applicator template library. It is the point corresponding to the line segment after displacement. It is a straight line segment in thin-section CT. It is the constraint weight of the bending segment. It is the point corresponding to the curved line segment after displacement. It is a set of corresponding points in the curved segments of CT images based on the applicator template library of thin-slice CT. This represents the sum of distances from the points corresponding to the straight line segments extracted from the clinical applicator CT image after displacement to the straight line segments in the thin-slice CT-based applicator template library CT image. This represents the sum of the minimum distances from the point corresponding to the curved segment extracted from the clinical applicator CT image after displacement to the curved segment in the applicator template library CT image based on thin-slice CT.

[0036] Formula (2) is the objective function of weighted registration. Its purpose is to optimize the rigid transformation parameters so that the set of center points extracted from the clinical applicator CT image is aligned with the piecewise geometric features of the fitting curve corresponding to the applicator template library based on thin-slice CT.

[0037] (3)

[0038] Formula (3) is the energy function of the smooth spline, which is used for piecewise constraint fitting. Under the condition of known piecewise structure and transition point constraints, smooth curve fitting is performed.

[0039] in: It is a curve fitting function based on the applicator template library of thin-slice CT. Fitting function for the center point set of clinical applicator CT images. It is a parameter; It refers to weighting, with greater weighting at curved sections; It is the smoothing coefficient; It is a point in the clinical applicator CT image where the center point is located within a straight line segment. This indicates that the center point set in the applicator template library based on thin-slice CT is located at the intersection of a straight line segment and a curved line segment. A function representing the set of center points of a clinical applicator CT image at the intersection of a straight line and a curve and a straight line. It is the constraint strength coefficient;

[0040] Apply position constraints at the transition point:

[0041]

[0042] in, and It is the position of the transition point of the fitting curve of the applicator template library based on thin-slice CT.

[0043] Compared with existing technologies, the technical solution proposed in this invention acquires high-resolution three-dimensional data of the applicator through thin-slice CT (0.6mm) scanning, which is more efficient, faster, and more accurate in subsequent image registration and curve fitting optimization. In actual clinical use, applicators generally require multiple channels to be used in combination according to treatment needs. This invention selects applicator channel CT data from a thin-slice CT-based applicator template library according to the type and model of the clinical applicator. The two are registered at multiple levels, using the fitted curve obtained from the applicator in the thin-slice CT-based applicator template library as the skeleton and combining it with the geometric characteristics of the applicator. During the registration and fitting process, conditional constraints are used to optimize the fitting of the clinical applicator center point set, achieving accurate reconstruction of the clinically used applicator. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the overall technical solution of the present invention.

[0045] Figure 2 A flowchart for establishing an applicator template library based on thin-slice CT;

[0046] Figure 3 Flowchart for improving the sparse cascaded network model;

[0047] Figure 4 To improve the sparse cascaded network model structure diagram;

[0048] Figure 5 This is a structural diagram of the dynamic sparse activation module;

[0049] Figure 6 This is a diagram of the lightweight branch module structure.

[0050] Figure 7 This is a diagram of the deep branching module structure;

[0051] Figure 8 This is a structural diagram of the cascaded decision-making module;

[0052] Figure 9 For registration flowchart. Detailed Implementation

[0053] like Figure 2 As shown, an applicator template library based on thin-slice CT is established:

[0054] The present invention acquires thin-slice CT data of the applicator without accessories such as metal bolts, i.e. only the tubular parts (left oval tube, right oval tube, 15-degree uterine cavity tube, 30-degree uterine cavity tube, 45-degree uterine cavity tube, 60-degree uterine cavity tube). The present invention uses CT image data with a slice thickness of 0.6 mm.

[0055] Based on the above thin-slice CT images, the center point set of the applicator based on hybrid geometric features was extracted and fitted for each type of applicator.

[0056] Save the above thin-slice CT images and fitting curve results as XML format files and store them as a template library for applicators based on thin-slice CT.

[0057] like Figure 3 As shown, an improved sparse cascade network model is established. The applicators used in clinical treatment are assembled according to different treatment purposes. This invention uses an improved sparse cascade network to identify and classify the applicators used in clinical treatment.

[0058] Preprocessing of CT images for clinical single-channel / multi-channel applicators: After the applicator is identified and classified, the CT image is processed using a three-dimensional feature extraction method while retaining the original geometric characteristics. This includes removing noise and foreign objects. However, when the applicator is detected to be multi-channel, the applicator needs to be split into multiple single-channel applicator CT image data.

[0059] Calling on the applicator CT images from the applicator template library based on thin-slice CT, the CT image registration and applicator reconstruction are completed:

[0060] After the CT image of the clinical applicator is preprocessed, the corresponding channel data of the applicator in the applicator template library based on thin-slice CT is matched according to the name of each channel of the applicator. The registration is performed by a multi-registration method based on improved physical structure driven optimization. The three-dimensional spatial registration of the CT image of the clinical applicator and the corresponding CT image of the applicator in the thin-slice CT template library is completed, and the correspondence between the two applicators in three-dimensional space is established.

[0061] The function of fitting the center point set of the applicator template library based on thin-slice CT is called. Based on the above registration results, the center point set on the clinical applicator CT image is corrected, and the applicator outline is created on the clinical applicator CT image according to the physical size of the applicator, thus completing the reconstruction of the clinical applicator.

[0062] It should be clarified that the applicator reconstruction of the present invention is a method for extracting and correcting the position of the center point of the applicator channel on the CT image of the metal applicator obtained by the CT image scanning scheme during clinical use.

[0063] The specific implementation scheme for applicator reconstruction proposed in this invention is carried out according to the following steps: Figure 1 Here is a flowchart illustrating the process:

[0064] Establish an applicator template library based on thin-slice CT:

[0065] The present invention acquires thin-slice CT data of a single tube (including: left oval tube, right oval tube, 15-degree uterine cavity tube, 30-degree uterine cavity tube, 45-degree uterine cavity tube, and 60-degree uterine cavity tube) without metal bolts or other accessories. The CT slice thickness is set to 0.6 mm. This slice thickness CT can be called thin-slice CT.

[0066] Based on the aforementioned thin-slice CT data, the applicator was subjected to applicator center point set extraction and curve fitting based on hybrid geometric features, as detailed in the following procedures: Figure 2 As shown. The specific implementation steps are as follows: (1) First, the CT image is preprocessed using the Gaussian filtering noise reduction method; (2) Then, the three-dimensional center points of the applicator channel in the thin-layer CT image are extracted using the dynamic threshold method. Since the applicator channel has a certain length and the scanning range is generally greater than 5cm, the number of center points extracted is large. For accurate description, it is called the center point set; (3) Point cloud sorting is carried out along the applicator path to identify the endpoints of the applicator center point set, and the number of point sets is enhanced at the applicator endpoints; (4) Curve fitting based on spline function is performed on the above center point set to obtain the fitted curve function; curvature analysis and segmentation are performed based on the above fitted curve function equation; (5) Based on the curve fitting of the applicator geometry, the three-segment smooth connection of the curve is strengthened, namely the geometric features of the straight line segment, the curve segment, and the straight line segment;

[0067] Save the above thin-slice CT images and fitting curve results into XML format files and store them as applicator templates with corresponding names to complete the establishment of an applicator template library based on thin-slice CT.

[0068] Based on the improved sparse-cascade network model, we can quickly identify and accurately classify clinically used applicators (combined forms with assembly accessories such as metal clips). Figure 3 This is a flowchart of the improved sparse cascade network model. The improved sparse cascade network model can achieve rapid (<1s) and accurate (recognition rate >95%) identification of the type of applicator combination and the angle of the uterine canal used clinically. The framework of the improved sparse cascade network model is shown below. Figure 4 The specific implementation process is as follows:

[0069] The data input for the improved sparse cascade network model is CT images of clinical applicators. The dataset used for data modeling and validation in this invention covers the following applicator usage methods: single-tube applicators (15-degree uterine cavity tube, 30-degree uterine cavity tube, 45-degree uterine cavity tube, 60-degree uterine cavity tube), two-tube applicators (applicators that include the left and right oocyte ducts), and three-tube applicators (applicators that combine the left and right oocyte ducts with various angles of the uterine cavity tube (15-degree uterine cavity tube, 30-degree uterine cavity tube, 45-degree uterine cavity tube, 60-degree uterine cavity tube), thus there are a total of 9 categories;

[0070] The improved sparse cascaded network model first passes the input data through a dynamic sparse activation module, see... Figure 5 The dynamic sparse activation module generates a weight map using a gated convolutional layer (Gated Conv), sets a dynamic threshold of 30%, and outputs the weight map to simple and complex regions respectively.

[0071] For simple regions, a lightweight branching module is used to further extract features, such as... Figure 6 The network structure shown is a lightweight branching module; for complex regions, a deep branching module is used to further extract features, such as... Figure 7 The network structure of the deep branching module is shown below.

[0072] After concatenating the features obtained from the lightweight branching module and the features obtained from the deep branching module, the data is then passed through a cascaded decision-maker module. (See...) Figure 8 As shown: The first stage is global average pooling, used to obtain the confidence score, and the second stage is a Transformer encoder block. If a high confidence score (≥0.85) is obtained after the first stage, a fully connected layer is connected to directly give the classification result without going through the second stage. However, when the confidence score is low (<0.85), the Transformer encoder block in the second stage is needed to complete the final classification and output the classification result.

[0073] Preprocessing of clinical applicator CT images: After the clinical applicator CT images are identified and classified by the improved sparse cascade network model that has been trained above, the CT images are further processed using three-dimensional feature extraction methods, including removing noise, foreign objects (image enhancers, IUDs or calcifications, etc.) and splitting the applicator while preserving the physical morphological features of the applicator.

[0074] The specific method for extracting three-dimensional features is as follows:

[0075] Adaptive thresholding: A dynamic thresholding algorithm based on local grayscale statistics, combined with the cascading use of morphological closing and opening operations (two iterations of closing operation followed by one iteration of opening operation), eliminates isolated noise points and other foreign objects caused by metal artifacts, while preserving the complete structure of the applicator.

[0076] Obtaining a center point set based on geometric features: Different center point set extraction strategies are used depending on the number of applicator ducts in this process. When the applicator is a single tube, i.e., a single-tube applicator, this invention uses anisotropic morphological filtering: a combination strategy of 3×3×3 cubic structure kernel and 2×2×2 spherical kernel is used respectively, and a skeletal smoothing technique based on curvature constraint is used to forcibly constrain the curvature change rate threshold (<1.5cm) to avoid anatomical distortion caused by excessive smoothing; when the applicator is multi-tube, i.e., multi-tube applicators (two-tube applicators and three-tube applicators), after the center point set is extracted, multiple ducts are separated to complete the applicator duct splitting, and the left and right oval ducts and uterine cavity ducts are identified based on the physical location of the applicator (for three-tube applicators). For example, a three-channel applicator with a 45-degree uterine cavity canal, left and right oval ducts, after being split, will yield three sets of CT images. Each set of CT images retains only the corresponding single channel of the applicator. In this example, the three sets of CT images respectively contain only the 45-degree uterine cavity canal, the left oval duct, and the right oval duct. The corresponding CT images will be matched with the applicator data of the corresponding channel in the applicator template library based on thin-slice CT in subsequent processes. During the channel splitting process of multi-channel applicators, the maximum connected component is extracted using image feature labeling parameters (3x3x4, all-1 three-dimensional matrix) tailored to clinical scanning conditions. The number of maximum connected components is limited according to the number of channels to complete the channel splitting.

[0077] After the CT images of the split clinical applicator are matched with the corresponding single-channel applicators in the applicator template library based on thin-slice CT, this invention uses a physical structure-driven optimization-based multi-registration technique to complete the spatial correspondence between the two sets of CT images and ultimately reconstruct the clinical applicator. This process can be seen in the following figure. Figure 9 The specific implementation process is as follows:

[0078] Initial image registration: Approximate registration of the two CT scans was completed using a bone-based registration method;

[0079] The fitted curve obtained from the applicator template library based on thin-slice CT is used as a "shape skeleton" to guide the point set formed by the clinical applicator CT image to fit within the target range;

[0080] Clinical application of applicator center point set extraction and fitting to enhance the three-segment structural features;

[0081] Smoothing based on physical structure, with hard constraints on transition points;

[0082] Weighted registration: Enhancement weight is applied to the curved segment, and the registration of the curved segment is enhanced in the multi-segment registration, so as to achieve three-dimensional spatial registration between the clinical applicator and the corresponding applicator in the applicator template library based on thin-slice CT.

[0083] Based on the constraint fitting of the clinical applicator center point set, the contour reconstruction of the physical dimensions of the pipeline is performed.

[0084] The relevant equations and objective function formulas used in this invention are as follows:

[0085] (1)

[0086] This formula is a curvature analysis equation used for segmenting fitted curves. Wherein... The equation for the fitting curve of the three-dimensional center point of the source device pipeline (in arc length) is... (for parameters) and These are the first and second derivatives (tangent vector and curvature vector), respectively.

[0087] (2)

[0088] This formula is the objective function for second-order registration (geometric constraint optimization), aiming to optimize the rigid transformation parameters so that the point set formed by the transformed clinical applicator CT image matches the piecewise geometric features of the thin-slice CT image curve. straight line segments and Align the curved sections.

[0089]

[0090] This formula is the energy function for smooth splines (used for piecewise constraint fitting), aiming to perform smooth curve fitting under known piecewise structure (straight segments and curved segments) and transition point constraints. Wherein: It is a curve fitting function based on the applicator template library of thin-slice CT. It is a parameter. It is a weight (the weight is greater in curved sections). The first term is the smoothing coefficient. The second term is the smoothing regularization term (penalty curvature). Positional constraints are applied to the transition points (the start and end points of the curved segment): .in, and This refers to the location of the thin-layer curve at the transition point. The third item is the constraint point limitation (reinforced bending segment segmentation point). It is the constraint strength coefficient.

[0091] In summary, this invention obtains CT data of each channel of the applicator through thin-slice CT scanning and fits the curve equation of each channel to establish an applicator template library based on thin-slice CT.

[0092] An improved sparse cascade network model for rapid identification and classification was established, taking into full account the efficiency and accuracy requirements in clinical work. The improved sparse cascade network model can quickly and accurately classify applicators, providing a basis for further matching the corresponding applicator channels in the applicator template library based on thin-section CT.

[0093] Based on thin-slice CT, the CT-CT image registration between CT data in the thin-slice CT applicator template library and CT data of clinical applicators is more efficient and faster. Using the applicator fitting curve established by thin-slice CT data as the skeleton, the fitting curve of the center point set extracted from the CT data of clinical applicators is optimized by means of segmented fitting and physical constraints, thereby completing the accurate reconstruction of the applicator.

Claims

1. A method for reconstructing an applicator, characterized in that, Thin-slice CT data of each channel of the applicator are obtained by thin-slice CT scanning, the curve equation of each channel is fitted, and an applicator template library based on thin-slice CT is established. An improved sparse cascade network model was established and trained using CT images of applicators under conventional scanning conditions. This model was then used to classify clinical applicators into single-tube or multi-tube types. In the case of multi-tube applicators, the tubes were split based on the classification results to obtain multiple single tubes. Single-tube matching was performed on applicators within the applicator template library based on thin-slice CT. The CT images in the applicator template library based on thin-slice CT are registered with the CT images of clinical applicators using an improved physical structure-driven optimization multi-registration method. Using the applicator fitting curve established based on the thin-slice CT applicator template library as the skeleton, the fitting curve of the center point set extracted from the CT image of the clinical applicator is optimized by segmented fitting and physical constraint methods to complete the accurate reconstruction of the applicator.

2. The applicator reconstruction method according to claim 1, characterized in that, The process of obtaining the applicator template library based on thin-slice CT is as follows: obtain thin-slice CT images of the applicator with only the tubular portion; the tubular portion includes the left oval tube, the right oval tube, the 15-degree uterine cavity tube, the 30-degree uterine cavity tube, the 45-degree uterine cavity tube, and the 60-degree uterine cavity tube; Based on the above thin-slice CT images, the center point set extraction and curve fitting of each type of applicator were performed using hybrid geometric features; the above thin-slice CT images and fitting curve results were stored as a thin-slice CT-based applicator template library.

3. The applicator reconstruction method according to claim 2, characterized in that, The specific implementation process of the donor center point set extraction and curve fitting based on hybrid geometric features is as follows: Gaussian filtering noise reduction method is used to preprocess thin-slice CT images; The three-dimensional center points of the applicator channel in thin-slice CT images were extracted using a dynamic thresholding method to obtain a set of center points. Point cloud sorting is performed along the donor path, the endpoints of the donor center point set are identified, and the number of point sets is increased at the donor endpoints. Perform spline function-based curve fitting on the above center point set to obtain the fitted curve function; Based on the above fitted curve function, curvature analysis and segmentation are performed, and the segmentation equation is as follows; (1) The equation is a partitioning equation; It is the equation of the fitting curve of the three-dimensional center point of the source feeder pipeline; Let the arc length be , and They are The first and second derivatives; Based on curve fitting of the source geometry, the three-segment smooth connection of the curve is enhanced, resulting in geometric features of straight line segment, curve segment, and straight line segment in sequence.

4. The applicator reconstruction method according to claim 2, characterized in that, The thin-slice CT image is CT image data with a slice thickness of 0.6 mm.

5. The applicator reconstruction method according to claim 1, characterized in that, The improved sparse cascaded network model includes a dynamic sparse activation module, a light branch module for processing simple regions, a deep branch module for processing complex regions, and a cascaded decision module. The data processed by the deep branch module and the light branch module are concatenated for features. The concatenated data is then passed through a cascaded decision module. The high-confidence results are output directly, while the low-confidence results are then passed through a transformer module before being output. Input CT images of clinical applicators. Applicator types include: single-tube applicators: 15-degree uterine cavity tube, 30-degree uterine cavity tube, 45-degree uterine cavity tube, and 60-degree uterine cavity tube; two-tube applicators: applicators used in combination with the left and right oocyte ducts; and three-tube applicators: applicators used in combination with the left and right oocyte ducts at various angles of the uterine cavity tube.

6. The applicator reconstruction method according to claim 5, characterized in that, The input data of the improved sparse cascaded network model first passes through a dynamic sparse activation module; The dynamic sparse activation module generates a mask using a gated convolutional layer, with a dynamic threshold set to 30% to identify key regions. In the above process, the mask generated by the gated convolutional layer is regarded as a weight map, which works together with the dynamic threshold to process the input data. After processing the weight map, two data points of the same size are output. They are arranged in descending order of weight value. The values ​​of points before 30% of the weight value are retained, and the values ​​of the remaining points are set to 0. This region is defined as a complex region. The values ​​of points after 30% of the weight value are retained, and the values ​​of the remaining points are set to 0. This region is defined as a simple region. The simple regions output by the dynamic sparse activation module are used to extract features using the lightweight branching module. Features are extracted from complex regions output by the dynamic sparse activation module using the deep branching module; The features obtained from the lightweight branching module and the features obtained from the deep branching module are concatenated and then passed through a cascaded decision-maker module. The first stage of the cascaded decision-maker module is global average pooling connected to a fully connected layer, and the second stage is a Transformer encoder block. If the confidence score obtained after the first stage is ≥0.85, the classification result is given directly without going through the second stage. If the confidence score is <0.85, the final classification is completed by the Transformer encoder block in the second stage, and the classification result is output.

7. The applicator reconstruction method according to claim 1, characterized in that, The registration specifically refers to: After the CT image of the clinical applicator is preprocessed, the corresponding channel data in the applicator template library based on thin-slice CT is matched according to the name of each channel of the applicator. Then, the CT image of the clinical applicator is registered in three-dimensional space with the CT image of the corresponding applicator in the applicator template library based on thin-slice CT, and the correspondence between the two applicators in three-dimensional space is established. The reconstruction of the clinical applicator specifically involves: calling the center point set fitting curve function of the applicator template library based on thin-slice CT; correcting the center point set on the CT image of the clinical applicator based on the above registration results; and creating the applicator outline on the CT image of the clinical applicator according to the physical size of the applicator to complete the reconstruction of the clinical applicator.

8. The applicator reconstruction method according to claim 7, characterized in that, The specific implementation process of the multi-registration method based on improved physical structure-driven optimization is as follows: Initial image registration: The CT images of the corresponding applicators in the applicator template library based on thin-slice CT were approximately registered with the clinical applicator CT images using the bone-based registration method. The fitted curve obtained from the applicator template library based on thin-slice CT is used as a "shape skeleton" to guide the point set formed by the clinical applicator CT image to fit within the target range; Clinical applicator center point set extraction and fitting, enhancing the three-segment structural features; the three-segment structural features are, in order, a straight segment, a curved segment, and a straight segment; Weighted registration: For the enhancement weight of the curved segment, the registration of the curved segment is enhanced in the multi-segment registration, and finally the three-dimensional spatial registration of the clinical applicator and the corresponding applicator in the applicator template library based on thin-slice CT is realized. The objective function of the relevant registration is given by formula (2). The contour reconstruction of the physical dimensions of the pipeline is performed by constrained fitting of the clinical applicator center point set; （2） in, These are the constraint weights of the line segments. This represents the distance from the point corresponding to the straight line segment extracted from the clinical applicator CT image after displacement to the straight line segment in the CT image based on the thin-slice CT applicator template library. It is the point corresponding to the line segment after displacement. It is a straight line segment in thin-section CT. It is the constraint weight of the bending segment. It is the point corresponding to the curved line segment after displacement. It is a set of corresponding points in the curved segments of CT images based on the applicator template library of thin-slice CT. This represents the sum of distances from the points corresponding to the straight line segments extracted from the clinical applicator CT image after displacement to the straight line segments in the thin-slice CT-based applicator template library CT image. This represents the sum of the minimum distances from the point corresponding to the curved segment extracted from the clinical applicator CT image after displacement to the curved segment in the applicator template library CT image based on thin-slice CT. Formula (2) is the objective function of weighted registration, which aims to optimize the rigid transformation parameters so that the set of center points extracted from the clinical applicator CT image is aligned with the piecewise geometric features of the fitting curve corresponding to the applicator template library based on thin-slice CT. （3） Formula (3) is the energy function of the smooth spline, which is used for piecewise constraint fitting. Under the condition of known piecewise structure and transition point constraints, smooth curve fitting is performed. in: It is a curve fitting function based on the applicator template library of thin-slice CT. Fitting function for the center point set of clinical applicator CT images. It is a parameter; It refers to weighting, with greater weighting at curved sections; It is the smoothing coefficient; It is a point in the clinical applicator CT image where the center point is located within a straight line segment. This indicates that the center point set in the applicator template library based on thin-slice CT is located at the intersection of a straight line segment and a curved line segment. A function representing the set of center points of a clinical applicator CT image at the intersection of a straight line and a curve and a straight line. It is the constraint strength coefficient; Apply position constraints at the transition point: in, and It is the position of the transition point of the fitting curve of the applicator template library based on thin-slice CT.

Images