A radiotherapy target volume automatic segmentation and quality control system and method based on multi-modal medical images

Abstract

The application provides a radiotherapy target automatic segmentation and quality control system and method based on multi-modal medical images, and is applied to the technical field of data processing. CT and MR volume images are acquired and spatial information is extracted, N4 bias field correction is performed on the MR, rigid body coarse registration, bidirectional field of view cutting and fine registration are performed, organ deformation is compensated by combining flexible registration, and a spatially aligned MR-CT image pair is obtained. The double-channel images after registration are input into a U-shaped encoding-decoding network to complete clinical target volume and multi-class organ segmentation. Through multiple reasoning calculations of voxel-level uncertainty, structure-level quality control scores are aggregated in the boundary ring belt to complete hierarchical determination. Finally, multi-layer visualization of images, segmentation results and uncertainty heat maps is realized, and registration logs, quality control scores and structured reports are output.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an automatic segmentation and quality control system and method for radiotherapy target areas based on multimodal medical images. Background Technology

[0002] A key prerequisite for radiotherapy planning is the accurate delineation of the central tumor volume (CTV) and orbital angiography (OARs). Manual delineation is time-consuming and subject to significant observer variability, affecting planning consistency and reproducibility. CT can provide the density information needed for dose calculation, but its soft tissue contrast is limited; MR has advantages in displaying soft tissue boundaries. Therefore, CT-MR fusion analysis is an important direction for improving the accuracy of target definition.

[0003] Existing multimodal automatic segmentation schemes typically include "registration + segmentation", but they still have the following problems: The inconsistency and large differences in field of view (FOV) between CT and MR lead to unstable convergence, redundant calculations, and limited accuracy in full-field registration. The significant movement and deformation of pelvic organs (bladder filling, intestinal peristalsis, etc.) make it difficult to ensure local boundary consistency by relying solely on rigid body alignment. Segmentation models often only output contours / masks, lacking quantitative indicators of the reliability of results and the risk of failure. Clinicians still need to check layer by layer to "where changes are needed," resulting in high implementation costs.

[0004] Therefore, there is an urgent need for a solution for clinical workflows that can output quality control evidence along with segmentation results, including "where the error might be, how serious the error is, and whether a review is needed," and present and report it in an intuitive way.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of this application, a method for automatic segmentation and quality control of radiotherapy target areas based on multimodal medical images is provided, comprising: acquiring CT volumetric images and MR volumetric images of a target user, and reading spatial information including image pixel spacing, orientation, and origin; based on the CT and MR images, first performing N4 field correction on the MR image, constructing a fixed image mask on the CT image, completing initial rigid body alignment through coarse Z-axis search, then achieving bidirectional field cropping of CT and MR images through MR overlay driving, completing fine rigid body registration within the cropping domain, introducing flexible registration to compensate for local organ deformation, generating spatially aligned MR-CT image pairs, and simultaneously outputting registration quality control images and log information; inputting the registered dual-channel CT-MR volumetric images into a U-shaped encoder-decoder network. The system generates segmentation results for clinical target volumes and various organs at risk based on the segmentation model. Multiple inferences are performed on the same case to obtain multiple sets of probability data, and voxel-level uncertainty is calculated using variance or entropy methods. A boundary ring (Bs) is constructed using the segmentation mask (Ms). Voxel-level uncertainty is aggregated in the ring region to obtain structural QC scores. Based on the validation set, thresholds are set to determine Pass / Needs-Review / Fail quality control conclusions. CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, and structural QC scores are integrated and visualized on the terminal interface using multi-layer overlay. A structured report containing registration logs, QC scores for each structure, and quality control judgment results is also output.

[0008] Another aspect of this application discloses an automatic segmentation and quality control system for radiotherapy target areas based on multimodal medical images, comprising: an image data acquisition module for acquiring CT and MR volumetric images of the target user and reading spatial information such as image pixel spacing, orientation, and origin; a multimodal image registration module for performing N4 field correction on the MR images based on the CT and MR images, constructing a fixed image mask on the CT, completing initial rigid body alignment through coarse Z-axis search, achieving bidirectional field cropping through MR coverage driving, completing fine rigid body registration within the cropping domain, introducing flexible registration to compensate for local organ deformation, generating spatially aligned MR-CT image pairs, and simultaneously outputting registration quality control images and log information; and a target area and organ at risk segmentation module for inputting the registered dual-channel CT-MR volumetric images into a U-shaped encoder-decoder network. The system employs a segmentation model to generate segmentation results for clinical target volumes and various organs at risk. A segmentation quality control module, based on the probabilistic output of the segmentation model, performs multiple inferences on the same case to obtain multiple sets of probabilistic data. It calculates voxel-level uncertainty using variance or entropy methods, constructs boundary rings (Bs) using the segmentation mask (Ms), aggregates voxel-level uncertainty in the ring region to obtain structural-level QC scores, and determines Pass / Needs-Review / Fail quality control conclusions based on a validation set threshold. A results visualization and report output module integrates CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, and structural-level QC scores, displaying them in a multi-layer overlay visualization on the terminal interface. It also outputs a structured report containing registration logs, QC scores for each structure, and quality control judgment results.

[0009] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described method for automatic segmentation and quality control of radiotherapy target areas based on multimodal medical images by executing the executable instructions.

[0010] According to another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a second processor, implements the above-described method for automatic segmentation and quality control of radiotherapy target areas based on multimodal medical images.

[0011] This application provides an automatic segmentation and quality control system and method for radiotherapy target areas based on multimodal medical images. The system acquires CT / MR volumetric images and spatial information, performs N4 field correction on the MR images, and compensates for organ deformation through rigid coarse registration, bidirectional field cropping, fine registration, and flexible registration to obtain spatially aligned MR-CT image pairs. The registered dual-channel images are input into a U-shaped encoder-decoder network to segment the clinical target area and various organs at risk. Through multiple inferences combined with variance / entropy methods, voxel-level uncertainty is calculated, and structural-level QC scores are obtained by aggregating at the segmentation mask boundary rings, achieving Pass / Needs-Review / Fail graded quality control. Finally, multi-layer visualization is completed, and a structured report is output. Furthermore, difficult cases reviewed clinically can be included in the sample database, enabling human-machine collaborative iterative optimization of the model.

[0012] This application aims to address the issues of low registration accuracy caused by inconsistencies in CT / MR field of view and organ deformation, thereby improving the accuracy of multimodal image fusion. It integrates CT density information with the contrast advantages of MR soft tissue, combining a U-shaped network to achieve precise segmentation of the target area and organs at risk, reducing observer variability in manual delineation. It quantifies segmentation uncertainty and implements graded quality control, accurately locating areas requiring review, significantly reducing the cost of manual clinical screening and improving the efficiency of radiotherapy planning. Multi-layer visualization and structured reporting enable traceability of the quality control process and quantifiable results, adapting to clinical workflows. A human-machine collaborative model iteration mechanism continuously optimizes segmentation performance, forming a technical closed loop and ensuring the clinical applicability and stability of the treatment plan.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0014] Figure 1 This document illustrates a flowchart of an automatic segmentation and quality control method for radiotherapy target areas based on multimodal medical images, provided in an embodiment of this application. Figure 2 This illustration shows a schematic diagram of an automatic segmentation and quality control system for radiotherapy target areas based on multimodal medical images, provided in one embodiment of this application. Detailed Implementation

[0015] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] The following is combined with Figure 1This application describes an automatic segmentation and quality control method for radiotherapy target areas based on multimodal medical images, according to exemplary embodiments thereof. It should be noted that the following application scenarios are shown only to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited in any way. Rather, the embodiments of this application are applicable to any suitable scenario.

[0017] In one implementation, Figure 1 The diagram illustrates a flowchart of an automatic segmentation and quality control method for radiotherapy target areas based on multimodal medical images, according to an embodiment of this application.

[0018] S101, acquire the CT volumetric image and MR volumetric image of the target user, and read the spatial information including image pixel spacing, orientation, and origin.

[0019] In one implementation, the acquisition of CT and MR volumetric images of the target user and the extraction and reading of image spatial information provide standardized and structured raw image data and spatial reference for subsequent multimodal image registration, target area and organ at risk segmentation, etc. The specific operation and explanation are as follows: Acquiring CT and MR volumetric images of the same subject—both types of images representing scan data of relevant anatomical sites within the same radiotherapy target area—is crucial input data for subsequent multimodal image registration and joint segmentation. The images must conform to clinical radiotherapy scanning specifications in volumetric image format, fully representing the anatomical structures of the target area and surrounding organs at risk. This operation is compatible with clinical radiotherapy workflows. The input CT and MR images are routine clinical scan data prior to radiotherapy planning. CT provides density information necessary for dose calculation, while MR supplements the high-contrast information of soft tissue boundaries. The combination of both provides multimodal information support for subsequent precise segmentation.

[0020] From the acquired CT and MR volumetric images, the core spatial information of the images is accurately extracted, specifically including three key types of information: image pixel spacing, image orientation, and image origin. Among them, pixel spacing represents the actual physical spatial distance corresponding to a single pixel in the image, image orientation defines the scanning orientation of the image in three-dimensional physical space, and image origin is the reference starting point of the image's three-dimensional coordinate system.

[0021] The spatial information read provides a basic spatial coordinate system reference for subsequent multimodal image registration. It is a prerequisite for aligning CT and MR images in the same physical space, ensuring the accuracy of image coordinate mapping, resampling, and other operations during the registration process. At the same time, it provides a standardized spatial basis for the spatial localization and contour extraction of subsequent segmentation results, ensuring that the spatial coordinates of the segmentation results are consistent with those of the original images and meet the spatial accuracy requirements of the radiotherapy plan.

[0022] S102, based on CT and MR images, first performs N4 field correction on MR, constructs a fixed image mask on CT, completes initial rigid body alignment through coarse Z-axis search, and then achieves bidirectional field cropping of CT and MR through MR overlay driving. Fine rigid body registration is completed within the cropping domain, and flexible registration is introduced to compensate for local organ deformation, generating spatially aligned MR-CT image pairs, while outputting registration quality control images and log information.

[0023] In one implementation, based on the practical requirements of radiotherapy target registration, the spatial matching specifications of CT and MR images, and the quality control output requirements of registration results, MRN4 field correction processing data, CT fixed image mask construction data, Z-axis coarse search registration data, bidirectional field cropping execution data, fine rigid body registration optimization data, and flexible registration deformation compensation data are integrated. Guided by the practical requirements of radiotherapy target registration, following industry standards for spatial matching of CT and MR images, and combining the quality control output requirements of registration results, the core processing data of each stage of the entire registration process are systematically integrated. The integrated data sources cover the entire process of registration preprocessing, coarse registration, fine registration, and deformation compensation. Specifically, they include image intensity correction data generated during MRN4 field correction processing, threshold segmentation and connected component filtering data during CT fixed image mask construction, translation and mutual information calculation data for Z-axis coarse search registration, index range and mask construction data for bidirectional field cropping, optimization parameters and similarity measurement data for fine rigid body registration, and displacement field prediction and non-rigid transformation data for flexible registration deformation compensation.

[0024] In the pelvic radiotherapy target registration scenario, the soft tissue intensity homogenization data after MRN4 off-field correction and the CT body contour mask data are integrated to ensure that the spatial coordinate systems of the two types of data are initially matched, providing a unified data foundation for the mutual information calculation of subsequent Z-axis coarse registration, while reserving the registration process data recording field required for quality control output.

[0025] Based on the requirements for the effectiveness of the registration process steps and the accuracy of image spatial alignment, the registration processing logic is designed, clarifying the processing standards for core operational data and associated output data. Core operational data includes MR-corrected images, CT mask images, cropped CT / MR images, and registration transformation parameters. Associated output data includes registration quality control images, registration process log information, and deformation field compensation data. Based on the requirements for the effectiveness of the connection between each step of the registration process and the accuracy of spatial alignment between CT and MR images, a full-link registration processing logic is designed, clarifying the execution order of registration operations, data flow paths, and algorithm calling rules. Simultaneously, standardized processing specifications are formulated for the core operational data and associated output data in the registration process, defining requirements for data format, accuracy, spatial dimensions, and storage methods. Core operational data refers to the key data directly involved in image transformation during the registration process, while associated output data consists of quality control and traceability data generated during the registration process.

[0026] In the core operational data, the MR-corrected image must maintain the same spatial dimension as the original MR image, and the pixel value precision must be retained as floating point. The registration transformation parameters are stored in matrix form, and the matrix dimension matches the three-dimensional space of the image. In the associated output data, the registration quality control image must include fixed types such as checkerboard diagrams and before-and-after alignment slice comparison images. The registration process log information must record the algorithm parameters, execution results, time consumption, and other information for each step according to the timestamp.

[0027] Based on the operational interrelationships of each registration step and the compensation requirements for local organ deformation, registration quality control rules were established, including MR calibration before registration, CT mask construction before constraint, coarse registration before fine-tuning, field of view cropping before optimization, and flexible registration with on-demand compensation. These rules, encompassing the entire registration process, adhere to the core principles of "preprocessing before registration, coarse registration before fine registration, constraint before optimization, and on-demand deformation compensation," ensuring the rationality and necessity of each step while adapting to the deformation characteristics of organs in different anatomical locations. For MR calibration before registration, MRN4 field correction is performed first to eliminate image intensity inhomogeneity. The corrected MR image is then used as a moving image for registration, preventing intensity deviations from affecting the accuracy of similarity measurements such as mutual information. The CT mask is constructed first and then constrained. First, a fixed image mask is constructed on the CT image. Then, the mask is used as a metric constraint and applied to the similarity calculation process of coarse search and fine rigid body registration in the Z-axis to remove background interference.

[0028] Coarse registration is performed first, followed by fine-tuning. Initial rigid body alignment of CT and MR images is achieved through a coarse Z-axis search to obtain a general spatial transformation relationship. This transformation result is then used as initial values ​​for fine rigid body registration within the clipping domain, improving the convergence speed and accuracy of fine registration. Field of view is clipped first, then optimized. Bidirectional field of view clipping of CT and MR images is performed first through MR coverage-driven methods, eliminating redundant areas not covered by each other. Fine rigid body registration is then performed within the clipped image regions, reducing computational load and improving registration accuracy. Flexible registration provides on-demand compensation. For areas with significant organ deformation, such as the pelvis and abdomen, flexible registration is performed on top of rigid registration, using a deformation field prediction network to compensate for local organ deformation. For areas with less deformation, such as the head, the flexible registration module degenerates its output unit deformation field, preserving the rigid registration result.

[0029] The operational requirements, workflow coordination, spatial alignment collaboration needs, and quality control rules for each registration step are integrated to generate basic information for multimodal registration of CT and MR images, including operation type, processing specifications, step-related logic, and registration execution strategy. The specific operational requirements, workflow coordination needs, image spatial alignment accuracy indicators, and established registration quality control rules for each registration step are comprehensively integrated and structured to extract basic information for multimodal registration of CT and MR images. This information serves as the direct basis for subsequent actual registration operations. It is necessary to clearly define the core elements of the entire registration process to ensure the standardization and replicability of the registration operation.

[0030] Clearly define the core operations for each stage, such as MR preprocessing (N4 field correction), CT preprocessing (fixed mask construction), coarse registration (Z-axis coarse search), cropping (bidirectional field cropping), fine registration (fine rigid body registration in the cropping domain), and deformation compensation (flexible registration). Define the technical indicators for each operation, such as the number of iterations for N4 field correction, the threshold range for CT mask construction, the step size and search range for Z-axis coarse search, the physical margin for field cropping, the similarity measurement type for fine rigid body registration (joint histogram mutual information), and the deformation field network structure for flexible registration (U-shaped convolutional encoding-decoding structure). Clearly define the execution order and data flow relationship of each operation, such as "MR N4 field correction → CT fixed mask construction → Z-axis coarse search initial rigid body alignment → MR coverage driving CT cropping → CT cropping region reverse mapping MR cropping → cropping domain fine rigid body registration → flexible registration (on demand) → registration quality control image and log output".

[0031] Develop registration execution plans for different scenarios, such as categorized by anatomical location (head / pelvic / abdomen registration strategy) and by image quality (routine image / low-contrast image registration strategy), clarifying the algorithm parameters, quality control rule adaptation methods, and whether deformation compensation is performed under each strategy. Example of algorithm-business integration: In the execution strategy of fine rigid body registration, the joint histogram mutual information algorithm is combined with the spatial matching specifications of the radiotherapy target area. The voxel sampling range for mutual information calculation is set to the target area and surrounding organs at risk within the CT fixed mask. Simultaneously, the number of pyramid layers in the multi-resolution pyramid iterative optimization algorithm is adapted to the spatial resolution of the image to ensure that the registration accuracy meets the spatial requirements of the radiotherapy plan (error ≤ 1 pixel spacing). In the flexible registration execution strategy, the training objective of the deformation field prediction network is combined with the organ local deformation compensation requirements. The loss function for network training uses a weighted sum of mutual information similarity terms and displacement field smoothing regularization terms, as shown in the formula: , where MI is mutual information, φ is dense displacement field, and ω1 and ω2 are weighting coefficients. This loss function ensures that the predicted displacement field not only maintains the similarity between CT and MR images, but also ensures the smoothness of the deformation field, which conforms to the physiological deformation law of organs.

[0032] S103 inputs the registered dual-channel CT-MR volumetric images into a segmentation model based on a U-shaped encoder-decoder network to generate segmentation results for clinical target volumes and multiple types of organs at risk.

[0033] In one implementation, based on the network modeling requirements for radiotherapy target segmentation, the feature extraction specifications for multimodal images, and the structural segmentation logic of radiotherapy clinicians, multi-scale feature extraction is performed on the registered dual-channel CT-MR volumetric images. This includes spatial texture features, soft tissue boundary features, and anatomical structure morphology features. This generates a multimodal fusion feature set extractable by a shared encoder, hierarchical encoded feature data, and task-related feature mapping information. Using the network modeling requirements for radiotherapy target segmentation as a framework, strictly adhering to the feature extraction specifications for multimodal images, and adapting to the anatomical structure segmentation logic relevant to radiotherapy clinicians, multi-scale feature extraction is performed on the registered dual-channel CT-MR volumetric images. The extraction process covers feature dimensions from the bottom layer to the top layer, specifically including spatial texture features characterizing the grayscale distribution of the image, soft tissue boundary features defining organ edges, and anatomical structure morphology features describing the geometric shape of organs.

[0034] Feature extraction is achieved through a shared encoder in a U-shaped network. This encoder employs a hierarchical convolutional structure to progressively downsample the input dual-channel images, ultimately generating three types of structured feature data: first, a multimodal fusion feature set containing complementary information from CT and MR; second, hierarchical coded feature data reflecting details at different anatomical levels; and third, task-related feature mapping information optimized for target area and organ-at-risk segmentation tasks. This operation deeply fuses the high-resolution texture features of bone and lung tissue in CT images with the clear boundary features of soft tissues such as tumors and nerves in MR images, addressing the clinical pain point of insufficient segmentation of complex anatomical structures by single-modal images and providing a comprehensive feature foundation for subsequent accurate segmentation.

[0035] The segmentation modeling specification based on U-shaped encoder-decoder networks performs structured processing on multimodal fusion feature sets, hierarchical encoded feature data, and task-related feature mapping information. An attention mechanism is used to complete task feature selection and fusion, generating single-channel probability maps of clinical target volume (CTV) and multi-channel probability maps of multiple organ at risk (OAR). Following the U-shaped encoder-decoder network segmentation modeling specification, the three types of feature data extracted above are structured. In the network's decoding stage, skip connections are used to fuse the hierarchical encoded feature data output by the encoder with the upsampled features of the decoder. Simultaneously, an attention mechanism is introduced to perform task feature selection, automatically enhancing key features related to clinical target volume (CTV) and organ at risk (OAR) while suppressing interference from irrelevant background features.

[0036] Through layer-by-layer upsampling and feature reconstruction by the decoder, the fused feature maps are restored to the same spatial resolution as the original images. Finally, two types of probability maps are generated in the network output layer: a CTV single-channel probability map for a single target segmentation task, and a multi-class OAR multi-channel probability map for simultaneous segmentation of multiple different organs at risk.

[0037] During feature fusion, the attention mechanism assigns dynamic weights to task-related feature mapping information, enabling the network to focus on key anatomical regions of clinical interest (such as tumor target areas, spinal cord, heart, etc.), significantly improving the segmentation accuracy of small target areas and adjacent organs.

[0038] Based on the segmentation logic of radiotherapy structures, sigmoid and softmax activation processes are applied to the single-channel probability map of CTV and the multi-channel probability map of OAR (Organ at Risk) for different categories, respectively, to normalize the probability values ​​and determine the categories, generating segmentation masks for clinical target volume and multiple categories of organs at risk. For the single-channel probability map of CTV, the sigmoid activation function is used to map the output value of each voxel to the interval between 0 and 1, representing the probability that the voxel belongs to the clinical target volume. For the multi-channel probability map of OAR for different categories, the softmax activation function is used to normalize the predicted value of each voxel for different categories of organs at risk, ensuring that the sum of the probabilities of each category is 1, achieving segmentation determination of mutual exclusion among multiple categories.

[0039] Based on the probability map after activation processing, binarization is performed by setting a probability threshold. Voxels with probability values ​​higher than the threshold are identified as target structures, and voxels with probability values ​​lower than the threshold are identified as background. Finally, segmentation masks for clinical target volumes and segmentation masks for multiple types of organs at risk are generated.

[0040] CTV single-channel sigmoid activation, specifically... Where z(x) is the original output of the network at voxel x. This represents the probability that the voxel belongs to the CTV.

[0041] Multiple OARsoftmax activations, specifically... ,in, Let be the original output of the network at voxel x corresponding to the k-th class of OAR, and N be the total number of OAR classes. Let be the probability that the voxel belongs to the k-th class of OAR.

[0042] Contour extraction and structural integration were performed on segmentation masks for clinical target volume and multiple organs at risk. Combined with spatial coordinate information from multimodal images, segmentation results for clinical target volume and multiple organs at risk were generated. Medical image post-processing was then performed on the generated clinical target volume and multiple organ at risk segmentation masks. First, the three-dimensional surface contours of each structural mask were extracted using a contour extraction algorithm to visually represent the anatomical morphology of the target area and organs. Subsequently, the extracted contours and the original mask data were structurally integrated to establish spatial relationships between different anatomical structures.

[0043] To ensure the segmentation accuracy of the segmentation model for CTV and multiple types of OAR, and to adapt to the segmentation logic of radiotherapy structures, the model adopts a multi-task weighted training loss function. The total loss is the weighted sum of the losses from multiple tasks. Furthermore, specific loss functions are designed for the segmentation characteristics of CTV and OAR respectively. Total loss function: ; CTV loss is a weighted sum of Dice loss and binary cross-entropy (BCE) loss. ; OAR loss is a weighted sum of multiple types of Dice loss and cross-entropy (CE) loss: Among them, the coefficient , α, β, γ, and δ can be flexibly set according to the class imbalance problem in radiotherapy clinics and the importance of CTV / OAR segmentation tasks. , To predict probabilities for the model, , Label the clinical gold standard.

[0044] During the integration process, the segmentation data is precisely bound to the multimodal image spatial coordinate information registered in the first step, ensuring that each voxel of the segmentation mask corresponds to a unique physical spatial coordinate. This ultimately generates standardized clinical target volumes and segmentation results for various organs at risk. The segmentation results generated by this operation include 3D mask data and visualized contour data, fully conforming to the input standards of the radiotherapy planning system (TPS). Physicians can directly use these results for target area confirmation, dose calculation, and planning evaluation, achieving seamless integration between AI segmentation and clinical workflow.

[0045] S104, based on the probability output of the segmentation model, performs multiple inferences on the same case to obtain multiple sets of probability data, calculates voxel-level uncertainty using variance or entropy methods, constructs boundary rings Bs by combining the segmentation mask Ms, aggregates voxel-level uncertainty in the ring region to obtain structural-level QC scores, and completes the quality control conclusion judgment of Pass / Needs-Review / Fail based on the validation set and the threshold is set.

[0046] In one implementation, the probability data output by the segmentation model, the segmentation mask data of the case images, and the pixel coordinate data of the structural boundaries are classified, extracted, and aligned in dimension to generate multiple sets of inference probability datasets, basic information of the segmentation mask, and structural boundary ring construction parameters. Addressing the core requirements of segmentation quality control, the raw data output by the segmentation model is systematically processed. First, three types of key data are extracted: 1) probability data output from multiple inferences by the segmentation model (including voxel-level probability prediction values ​​of CTV and various OARs); 2) case image segmentation mask data generated based on the probability data (binarized target structure mask Ms); and 3) structural boundary pixel coordinate data representing the edge position of the target structure (three-dimensional spatial coordinates of the voxels at the boundary of the mask Ms).

[0047] After extraction, dimensional alignment is performed to ensure that the three types of data are completely matched in spatial dimension (voxel coordinates), structural dimension (corresponding to the same CTV or OAR), and sample dimension (same case). This results in three standardized types of intermediate data: multiple sets of inference probability datasets, basic information on the segmentation mask (including mask morphology, spatial range, etc.), and structural boundary ring construction parameters (including boundary coordinates, structural type, etc.). This operation resolves the issues of inconsistent output data formats and spatial coordinates in the segmentation model, providing a standardized data foundation for subsequent uncertainty calculations and boundary ring construction, ensuring the accuracy and repeatability of quality control analysis.

[0048] Based on the goal of ensuring the accuracy of segmentation result quality control, multiple sets of inference probability datasets are associated with the uncertainty calculation rules of variance method and entropy method. Combined with voxel-level operational logic, probability fluctuation information from multiple inferences and confidence information of voxel positions are integrated to generate a basis for voxel-level uncertainty calculation. With the core objective of ensuring the accuracy of segmentation result quality control, multiple sets of inference probability datasets are associated with two preset uncertainty calculation rules (variance method and entropy method), and the data is deeply integrated using voxel-level operational logic.

[0049] The integration process focuses on two key types of information: first, the probability fluctuation information of the same voxel during multiple inferences (reflecting the stability of the model's classification of that voxel); and second, the confidence information of each voxel in the fusion probability map (reflecting the reliability of the model's classification of that voxel). By mapping the probability fluctuation information to variance method rules and the confidence information to entropy method rules, a basis for calculating voxel-level uncertainty that can be directly used for calculation is generated, clarifying the uncertainty calculation method and data source for each voxel.

[0050] For a voxel x in CTV, if the variance method is used to calculate the uncertainty, the degree of fluctuation is calculated by using the variance formula based on the probability value sequence of the voxel in multiple sets of inference probability datasets; if the entropy method is used, the information uncertainty is calculated by using the entropy formula based on the confidence value of the voxel in the fusion probability map. Both methods can quantify the reliability of the model's segmentation result for the voxel.

[0051] Based on the aggregation calculation objective of structural-level QC scores, the voxel-level uncertainty data, the dilation and erosion rules of the segmentation mask Ms, and the spatial range of the boundary ring Bs are respectively mapped to the ring aggregation operation model and the structural-level score calculation strategy, generating core parameters for QC score accounting and initial information for aggregation standards. Based on the aggregation calculation objective of structural-level QC scores, a correspondence between multi-dimensional data and the accounting model and strategy is established. First, the voxel-level uncertainty data is associated with the ring aggregation operation model to clarify the aggregation method of uncertainty within the boundary ring. Second, the dilation and erosion rules of the segmentation mask Ms (defining the generation logic of the boundary ring) are associated with the structural-level score calculation strategy to ensure logical consistency between ring construction and score accounting. Finally, combined with the structural boundary ring construction parameters, the spatial range of the boundary ring Bs is clarified (e.g., the value of the ring radius r can be defined by the number of voxels or physical distance). By establishing the above correspondence, two types of key information are generated: first, the core parameters for QC score calculation (including the ring radius r, aggregation operation coefficients, etc.); second, the initial information of the aggregation standard (including the aggregation operation formula, data screening conditions, etc.), which provide a clear basis for the subsequent calculation of structural-level QC scores.

[0052] The generation of the boundary ring Bs strictly follows the formula In this formula, "dilate" represents the expansion operation, "erode" represents the erosion operation, and "\" represents the set difference operation. This formula is incorporated as a core parameter into the QC score calculation system to ensure the standardization of ring construction. The aggregation standard of structural-level QC scores Us follows the formula. It is clearly stated that the average value of the voxel-level uncertainty within the boundary ring zone is used as the structural-level QC score.

[0053] Based on voxel-level uncertainty calculations and the core parameters and initial information of QC score calculation and aggregation standards, key voxels in the core data extracted through classification are labeled and aggregated. Combined with thresholds set in the validation set, Pass / Needs-Review / Fail judgments are completed, generating radiotherapy target segmentation quality control judgment results that include segmentation result quality control conclusions, structural risk levels, and review priorities. The entire quality control judgment process is executed based on voxel-level uncertainty calculations, core parameters of QC score calculation, and initial information of aggregation standards. First, key voxels (mainly voxels within the boundary ring Bs) in the core data extracted from the classification are marked to focus on key quality control areas. Second, aggregation operations are performed on the uncertain data of the marked voxels according to the aggregation criteria to calculate the structure-level QC score Us for each structure (CTV or OAR). Finally, based on the dual thresholds (τ1, τ2) set by the statistical distribution of the validation set, the quality control conclusion is determined: Us < τ1 is judged as Pass (the segmentation result is reliable and no review is required); τ1 ≤ Us < τ2 is judged as Needs-Review (the segmentation result has some uncertainty and requires manual review); Us ≥ τ2 is judged as Fail (the segmentation result is unreliable and there is a risk of failure).

[0054] During the assessment process, morphological outlier rules (such as volume outliers, connected component anomalies, and void anomalies) can be combined to optimize the assessment results, improving the robustness and specificity of quality control. The final result generates a radiotherapy target segmentation quality control assessment that includes the segmentation quality control conclusion, structural risk level (high / medium / low), and review priority (high / medium / low). This operation transforms quantified QC scores into clinically understandable quality control conclusions, clarifying the necessity and priority of reviewing each structure. Physicians can prioritize Needs-Review and Fail structures, significantly reducing the cost of manual review and improving the efficiency of radiotherapy planning.

[0055] The S105 integrates CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, and structural QC scores, displaying them in a multi-layer overlay visualization on the terminal interface. It also outputs a structured report containing registration logs, QC scores for each structure, and quality control judgment results.

[0056] In one implementation, CT / MR volumetric images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, structural QC scores, and registration log data are classified, extracted, and aligned with attributes to generate a multimodal image dataset, segmentation structure visualization information, uncertainty quantification data, and quality control judgment result set. For heterogeneous data generated throughout the radiotherapy target segmentation quality control process, data is classified and extracted according to data type and business purpose, while performing full-dimensional attribute alignment processing to ensure that various types of data can achieve linked matching across spatial, structural, and case dimensions.

[0057] The core data extracted by classification includes five categories: first, CT / MR volumetric images (registered original and corrected images); second, CTV / OAR segmentation contours / masks (binarized masks and three-dimensional contour data of target areas and organs at risk); third, voxel-level uncertainty heatmaps (voxel-level uncertainty quantification data generated based on variance or entropy methods); fourth, structural-level QC scores (aggregated quality control scores for each anatomical structure); and fifth, registration log data (execution parameters, results, and time consumption records for the entire multimodal registration process).

[0058] Attribute alignment processing focuses on three core dimensions: Spatially, it ensures complete consistency in pixel spacing, orientation, and origin among images, masks, and heatmaps, achieving precise mapping of points at the same coordinates; Structurally, it binds each QC score and quality control conclusion to a corresponding CTV or OAR anatomical identifier; and Case-wise, it assigns a unique, unified case identifier to all data, ultimately generating four types of standardized intermediate datasets: multimodal image base dataset, segmentation structure visualization information, uncertainty quantification data, and quality control judgment result set. This operation solves the heterogeneity problem of multimodal images, segmentation results, quality control data, and process logs, establishing a "image-segmentation-quality control-process" linkage, which is a prerequisite for achieving terminal visualization linkage and structured report output.

[0059] Based on the goal of visualizing and reporting radiotherapy target area quality control results, this project associates multimodal image datasets with segmentation structure visualization information and uncertainty quantification data. It integrates the display levels and transparency adaptation information of images, segmentation contours, and heatmaps using layer overlay display logic to generate the basis for constructing the visualization interface. With the goal of "intuitive visualization and convenient interactive verification" of radiotherapy target area quality control results, the standardized multimodal image datasets, segmentation structure visualization information, and uncertainty quantification data are logically linked. Combined with the layer overlay display logic of medical imaging terminals, the core rules and parameters for interface construction are clearly defined.

[0060] The core association and rule definition includes two parts: First, the display layer definition follows the clinical image reading habit of "bottom layer is the image, middle layer is the segmentation contour, and top layer is the uncertainty heatmap," clarifying the switching layer of CT and MR images, the layered display rules of CTV and various OAR contours, and the overlay layer of uncertainty heatmaps. Second, transparency adaptation information sets dynamic transparency ranges for different data types. For example, the transparency of the original image is fixed at 100%, the transparency of the segmentation contour is set to 80% to ensure clear boundaries, and the uncertainty heatmap uses a gradual transparency of 50%-70% to show the quantitative distribution without obscuring the underlying image and segmentation contour. Through the integration of the above rules, the basis for constructing the visualization display interface is generated, clarifying the number of layers, layer order, data mapping relationship, and interaction trigger conditions (such as contour hiding / showing, heatmap switching, and transparency adjustment) of the terminal interface, providing standardized instructions for terminal interface rendering.

[0061] Based on the standardization goals of structured report output, the registered log data, QC scores for each structure, and quality control judgment results are mapped to the report content classification model and data layout and display strategy, respectively, generating initial information for core report output parameters and format standards. In accordance with the goals of standardized and traceable medical document output, a one-to-one correspondence is established between the registered log data, QC scores for each structure, and quality control judgment results and the preset report content classification model and data layout and display strategy, clarifying the core components and format standards of the structured report.

[0062] First, using a report content classification model, the core data is divided into three main report modules: 1) a basic case information module (linked to a unique case identifier); 2) a registration process traceability module (corresponding to registration log data, recording the execution parameters and results of N4 correction, coarse registration, fine registration, and flexible registration); and 3) a segmented quality control results module (corresponding to QC scores and quality control judgment results, presented according to CTV and OAR categories). Second, through data layout and display strategies, the data presentation format for each module is defined: the registration log uses a chronological text description format, QC scores use an itemized format of "structure name - score value - quality control conclusion," and quality control judgment results use a summary conclusion format. Ultimately, two types of key information are generated: core parameters of the report output (including module composition, data entries, and classification identifiers) and initial format standard information (including layout order, font specifications, data accuracy, and conclusion labeling rules).

[0063] Based on the initial information of the visualization interface construction criteria and the core parameters and format standards for report output, key content in the core data extracted by classification is marked, integrated, and formatted to achieve multi-layer overlay visualization on the terminal interface, generating a structured report on radiotherapy target segmentation quality control that includes registration logs, QC scores for each structure, and quality control judgment results. Based on the visualization interface construction criteria and the initial information of the core parameters and format standards for report output, the final operations of terminal presentation and report generation are performed. The first stage is the implementation of multi-layer overlay visualization on the terminal interface: according to the hierarchy and transparency rules of the interface construction criteria, key content in the core data extracted by classification (such as target core layer images, heatmaps of high uncertainty areas, and key OAR segmentation contours) is marked, driving the terminal rendering engine to complete multi-layer overlay rendering, achieving an integrated display of "image base map + contour annotation + heatmap highlighting," supporting clinicians to switch image modalities, hide / show specific structural contours, and adjust heatmap transparency through interactive operations, intuitively verifying segmentation quality and uncertainty distribution.

[0064] The second stage involves the integration, formatting, and generation of a structured report. Following the core parameters and format standards for report output, key data (such as registration key step logs, QC scores for each structure, and overall quality control conclusions) are integrated and formatted. Strict adherence to the standardized requirements of medical documentation is maintained, organically combining the chronological registration process, itemized quality control scores, and summarized judgment results to ultimately generate a structured report on radiotherapy target segmentation quality control. This report comprehensively includes the entire registration process log, QC scores for each CTV / OAR, and quality control judgment results for each structure, ensuring traceability of the quality control process, quantifiability of quality control results, and archiving of report content. The terminal visualization directly serves the real-time review work of clinicians, helping to quickly locate areas of high uncertainty. The structured report meets the compliance requirements for radiotherapy plan review and medical record management, achieving a closed-loop process between AI-assisted segmentation technology and the radiotherapy clinical workflow.

[0065] This application also provides an implementation method that integrates CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, and structural-level QC scores, and displays them in a multi-layer overlay visualization on a terminal interface. It also outputs a structured report containing registration logs, QC scores for each structure, and quality control judgment results. Furthermore, it includes: For structures judged as Needs-Review or Fail, physicians can complete the review and editing through a visual interface. The system records the model output, quality control information, the physician's final outline, and areas of difference in modification, and includes modified or marked difficult cases in a difficult sample library. For CTV / OAR structures judged as Needs-Review (requiring review) or Fail (failed) by quality control, the system provides clinicians with integrated review and editing functions through a terminal visual display interface. Physicians can accurately locate segmentation errors or high uncertainty areas in an overlay view of "multimodal image base map + uncertainty heatmap + model segmentation outline," and complete editing operations such as adding, deleting, and moving outlines to generate a final segmentation outline that conforms to clinical radiotherapy standards.

[0066] After the doctor completes the operation, the system automatically performs full-dimensional data recording and associated storage. The core recorded data includes: the original output probability map and mask of the segmentation model, the voxel-level uncertainty heatmap and structural-level QC score of the structure, the final segmentation contour edited by the doctor, and the mask of the difference region between the original model contour and the doctor's final contour (i.e., the modified difference region). For samples marked by the doctor as difficult cases (such as rare anatomical variations, tumors with special morphology, or extremely poor image quality), the system binds them to the above full-volume recorded data and uniformly includes them in the difficult sample library for management, providing core incremental data for subsequent model iterations. This operation transforms the professional knowledge and experience of clinicians into quantifiable and reusable model optimization data, solving the pain point of missing "difficult samples" in traditional model training. At the same time, by recording and modifying the difference region, it accurately locates the performance shortcomings of the model, achieving targeted model optimization.

[0067] Based on a historical training set and a difficult sample database, the segmentation model undergoes periodic offline joint training or fine-tuning. Higher training weights are assigned to regions of difference and boundary neighborhoods modified by doctors to complete iterative model optimization. Simultaneously, gray-scale evaluation, deployment, and rollback management of model versions are implemented, forming a human-machine collaborative model optimization closed loop. Using a combined historical training set and a difficult sample database as the data foundation, offline optimization of the segmentation model is performed according to a preset cycle (e.g., monthly, quarterly). Optimization methods include joint training (for scenarios with ample difficult sample database data) and model fine-tuning (for scenarios with limited difficult sample database data). The core strategy focuses on key regions modified by doctors through weighted training, achieving precise improvement in model performance.

[0068] Historical training data is merged with a difficult sample database to maintain the basic proportion of historical training data and preserve the model's generalization ability. Samples from the difficult sample database are used as the primary training data. For each difficult sample, a spatial weight mask is constructed based on the modified difference regions recorded by the system: voxels within the modified difference regions, as well as the boundary neighbor voxels surrounding these regions, are assigned training weights significantly higher than those for the background and correctly segmented regions; for regions originally correctly segmented by the model, basic training weights are set, ensuring that the training process focuses on the model's weak points.

[0069] The model is still trained based on the original U-shaped encoder-decoder network architecture, without changing the network's modules, layers, and connections; only the network's weight parameters are updated. If joint training is used, the fused full dataset is divided into training and validation sets according to a preset ratio. The original model is used as the initial weights, and the complete training process is executed to optimize the network's fully connected and convolutional layer parameters. If model fine-tuning is used, the parameters of the network's low-level feature extraction layer are fixed (maintaining the ability to extract basic features of multimodal images), and only the parameters of the encoder's higher layers, decoder, and attention mechanism layers are updated. Training is performed with a small learning rate to avoid overfitting.

[0070] The loss function used during training still employs the standard loss for segmentation tasks, while incorporating a spatial weight mask to tilt the loss calculation towards modifying discrepancies, ensuring improved segmentation accuracy in key regions. Key training parameters include: the learning rate during fine-tuning (1 / 10 to 1 / 5 of the original training learning rate), the weight decay coefficient (maintaining consistency with the original model training), the number of training epochs (stopping when the validation set loss no longer decreases), and the weight coefficients for modifying discrepancies (set to 5-10 times the base weights). This weighted training strategy enables the model to quickly learn the doctor's correction logic and specifically optimize weak areas such as boundary segmentation and small target region segmentation.

[0071] To ensure the safety and stability of the iteratively optimized model in clinical applications, a standardized model version lifecycle management process has been established. This process includes three core stages: gray-scale evaluation, official deployment, and version rollback, forming a closed-loop model management mechanism. The iteratively optimized new model version is deployed to a gray-scale environment, selecting select clinical departments and specific case types (such as those containing difficult cases) for gray-scale testing. Evaluation metrics include: the average QC score across the entire structure, the Needs-Review / Fail rate, and the time spent editing by physicians. The evaluation results of the new model are compared with those of the original model. Only when the new model's core metrics are significantly better than the original model (e.g., an increase in the average QC score and a decrease in the review rate) can it proceed to the official deployment stage.

[0072] After gray-scale evaluation, the new model version is deployed to the production environment in a smooth transition, replacing the original model to perform segmentation tasks. Simultaneously, the new model's version number, training data, optimization parameters, and evaluation results are fully archived to ensure model version traceability. If any anomalies are detected in clinical applications after the new model is launched (such as a sudden increase in segmentation error rate or system compatibility issues), the system can trigger an emergency rollback mechanism to quickly switch to the previous stable model version, ensuring the continuity and safety of radiotherapy target segmentation. This management process ensures the effectiveness of model iteration while mitigating the clinical risks associated with launching a new model, achieving a smooth transition between AI model optimization and clinical workflow, ultimately forming a human-machine collaborative closed loop of "clinical feedback driving model optimization, and the optimized model serving clinical practice."

[0073] In one implementation, such as Figure 2 As shown, this application also provides an automatic segmentation and quality control system for radiotherapy target areas based on multimodal medical images, comprising: The image data acquisition module 201 is used to acquire the CT volumetric image and MR volumetric image of the target user, and read spatial information such as image pixel spacing, orientation, and origin. The multimodal image registration module 202 is used to perform N4 field correction on MR based on CT and MR images, construct a fixed image mask on CT, complete the initial rigid body alignment through coarse Z-axis search, realize bidirectional field cropping through MR coverage drive, complete fine rigid body registration within the cropping domain, introduce flexible registration to compensate for local organ deformation, generate spatially aligned MR-CT image pairs, and output registration quality control images and log information at the same time. The target volume and organs at risk segmentation module 203 is used to input the registered dual-channel CT-MR volume images into a segmentation model based on a U-shaped encoder-decoder network to generate segmentation results for clinical target volumes and multiple types of organs at risk. The segmentation quality control module 204 is used to perform multiple inferences on the same case based on the probability output of the segmentation model to obtain multiple sets of probability data. It calculates voxel-level uncertainty through variance method or entropy method, constructs boundary ring Bs in combination with segmentation mask Ms, aggregates voxel-level uncertainty in the ring region to obtain structural level QC score, and completes the quality control conclusion judgment of Pass / Needs-Review / Fail based on the validation set and the threshold is set. The results visualization and report output module 205 is used to integrate CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps and structural QC scores, and display them in a multi-layer overlay visualization on the terminal interface. At the same time, it outputs a structured report containing registration logs, QC scores for each structure, and quality control judgment results.

[0074] The computer-readable storage medium provided in the above embodiments of this application and the automatic segmentation and quality control method for radiotherapy target area based on multimodal medical images provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application programs stored therein.

[0075] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the automatic segmentation and quality control method, system, electronic device, and readable storage medium for radiotherapy targets based on multimodal medical images are basically similar to the embodiments of the automatic segmentation and quality control method for radiotherapy targets based on multimodal medical images described above, and are therefore described simply. Relevant parts can be referred to in the descriptions of the embodiments of the automatic segmentation and quality control method for radiotherapy targets based on multimodal medical images described above.

Claims

1. A method for automatic segmentation and quality control of radiotherapy target areas based on multimodal medical images, characterized in that, include: Acquire CT and MR volumetric images of the target user and read spatial information including image pixel spacing, orientation, and origin. Based on CT and MR images, N4 field correction is first performed on MR, a fixed image mask is constructed on CT, initial rigid body alignment is completed through coarse Z-axis search, bidirectional field cropping of CT and MR is achieved through MR coverage drive, fine rigid body registration is completed within the cropping domain, flexible registration is introduced to compensate for local organ deformation, spatially aligned MR-CT image pairs are generated, and registration quality control images and log information are output at the same time. The registered dual-channel CT-MR volumetric images are input into a segmentation model based on a U-shaped encoder-decoder network to generate segmentation results for clinical target volumes and multiple types of organs at risk. Based on the probability output of the segmentation model, multiple inferences are performed on the same case to obtain multiple sets of probability data. Voxel-level uncertainty is calculated by variance method or entropy method. Boundary rings Bs are constructed by combining the segmentation mask Ms. Voxel-level uncertainty is aggregated in the ring region to obtain structural QC scores. Based on the validation set, the threshold is set to complete the quality control conclusion judgment of Pass / Needs-Review / Fail. It integrates CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, and structural QC scores, and displays them in a multi-layer overlay visualization on the terminal interface. It also outputs a structured report that includes registration logs, QC scores for each structure, and quality control judgment results.

2. The method as described in claim 1, characterized in that, Based on CT and MR images, N4 field correction is first performed on the MR image. A fixed image mask is then constructed on the CT image. Initial rigid body alignment is achieved through coarse Z-axis search. Next, bidirectional field cropping of CT and MR images is achieved through MR overlay driving. Fine rigid body registration is then performed within the cropping domain. Flexible registration is introduced to compensate for local organ deformation, generating spatially aligned MR-CT image pairs. Simultaneously, registration quality control images and log information are output, including: Based on the practical requirements of radiotherapy target registration, the spatial matching specifications of CT and MR images, and the quality control output requirements of registration results, MRN4 off-field correction processing data, CT fixed image mask construction data, Z-axis coarse search registration data, bidirectional field cropping execution data, fine rigid body registration optimization data, and flexible registration deformation compensation data are integrated. Based on the collaborative requirements of the effectiveness of the registration process steps and the accuracy of image spatial alignment, the registration processing logic is designed, and the processing standards of core operational data and associated output data are clarified. The core operational data includes MR corrected images, CT mask images, cropped CT / MR images, and registration transformation parameters. The associated output data includes registration quality control images, registration process log information, and deformation field compensation data. Based on the operational correlation of each registration step and the compensation requirements for local organ deformation, we set up registration quality control rules: MR calibration before registration, CT mask construction before constraint, coarse registration before fine adjustment, field of view cropping before optimization, and flexible registration with compensation as needed. The operational requirements, workflow connections, spatial alignment coordination needs, and quality control rules for each step of the registration process are integrated to generate basic information for multimodal registration of CT and MR images, including operation type, processing specifications, step-related logic, and registration execution strategy.

3. The method as described in claim 1, characterized in that, The registered dual-channel CT-MR volumetric images are input into a segmentation model based on a U-shaped encoder-decoder network to generate segmentation results for clinical target volumes and multiple types of organs at risk, including: Based on the network modeling requirements for radiotherapy target segmentation, the feature extraction specifications for multimodal images, and the structural segmentation logic of radiotherapy clinical practice, multi-scale feature extraction is performed on the registered dual-channel CT-MR volumetric images, including spatial texture features, soft tissue boundary features, and anatomical structure morphology features, generating a multimodal fusion feature set, hierarchical coded feature data, and task-related feature mapping information that can be extracted by the shared encoder. The segmentation modeling specification based on U-shaped encoder-decoder network performs structured processing on multimodal fusion feature set, hierarchical encoded feature data, and task-related feature mapping information. It completes task feature selection and fusion through attention mechanism to generate CTV single-channel probability map and multi-class OAR multi-channel probability map. Based on the segmentation logic of radiotherapy structure, sigmoid and softmax activation processing are applied to the CTV single-channel probability map and the multi-class OAR multi-channel probability map respectively to complete the probability value normalization and class determination, and generate segmentation masks for clinical target volume and segmentation masks for multiple types of organs at risk. Contour extraction and structural integration are performed on the segmentation masks of clinical target volume and multiple organs at risk. Combined with the spatial coordinate information of multimodal images, the segmentation results of clinical target volume and multiple organs at risk are generated.

4. The method as described in claim 1, characterized in that, Based on the probability output of the segmentation model, multiple inferences are performed on the same case to obtain multiple sets of probability data. Voxel-level uncertainty is calculated using the variance method or entropy method. A boundary ring Bs is constructed by combining the segmentation mask Ms. Voxel-level uncertainty in the ring region to obtain the structural-level QC score. Based on the validation set, a threshold is set to complete the quality control conclusion judgment of Pass / Needs-Review / Fail, including: The probability data output by the segmentation model, the segmentation mask data of the case images, and the pixel coordinate data of the structural boundary are classified, extracted, and aligned in dimension to generate multiple sets of inference probability datasets, basic information of the segmentation mask, and construction parameters of the structural boundary ring. Based on the accuracy target of quality control judgment of segmentation results, multiple sets of inference probability datasets are associated with the uncertainty calculation rules of variance method and entropy method. Combined with voxel-level operation logic, the probability fluctuation information of multiple inferences and the confidence information of voxel positions are integrated to generate the basis for voxel-level uncertainty calculation. Based on the aggregation calculation objective of structural-level QC scores, the voxel-level uncertainty data, the expansion and corrosion rules of the segmentation mask Ms, and the spatial range of the boundary ring Bs are respectively mapped to the ring aggregation operation model and the structural-level score calculation strategy, generating the core parameters for QC score accounting and the initial information of the aggregation standard. Based on the calculation basis of voxel-level uncertainty and the initial information of QC score accounting core parameters and aggregation standards, key voxels in the core data extracted by classification are marked and aggregated. Combined with the threshold set by the validation set, the Pass / Needs-Review / Fail judgment is completed, and the radiotherapy target area segmentation quality control judgment result including the segmentation result quality control conclusion, structural risk level, and review priority is generated.

5. The method as described in claim 4, characterized in that, Integrating CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, and structural QC scores, the system provides multi-layer overlay visualization on the terminal interface. It also outputs a structured report containing registration logs, QC scores for each structure, and quality control assessment results, including: Classification, extraction, and attribute alignment processing are performed on CT / MR volumetric images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, structural QC scores, and registration log data to generate a multimodal image basic dataset, segmentation structure visualization information, uncertainty quantification data, and quality control judgment result set. Based on the goal of visualizing and reporting the quality control results of radiotherapy target areas, the multimodal image dataset is associated with the visualization information of segmentation structure and uncertainty quantification data. The display hierarchy and transparency adaptation information of images, segmentation contours and heat maps are integrated by combining the layer overlay display logic to generate the basis for constructing the visualization display interface. Based on the standardization goal of structured report output, the registered log data, QC scores of each structure, and quality control judgment results are respectively matched with the report content classification model and data layout and display strategy to generate the initial information of the core parameters and format standards for report output; Based on the initial information of the visualization interface construction basis and the core parameters and format standards of the report output, the key content in the core data extracted by classification is marked and integrated for layout, realizing the visualization display of multiple layers on the terminal interface, and generating a structured report of radiotherapy target area segmentation quality control that includes registration logs, QC scores of each structure, and quality control judgment results.

6. The method as described in claim 1, characterized in that, Integrating CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, and structural QC scores, the system provides multi-layer overlay visualization on the terminal interface. It also outputs a structured report containing registration logs, QC scores for each structure, and quality control assessment results. Further details include: For structures that are judged as Needs-Review or Fail, doctors can complete the review and editing on the visual interface. The system records the model output, quality control information, the doctor's final outline and the areas of difference in modification, and includes modified or marked difficult cases into the difficult sample library. Based on historical training sets and a database of difficult samples, the segmentation model is periodically trained or fine-tuned offline. Doctors modify the difference regions and boundary neighborhoods and set higher training weights to complete the iterative optimization of the model. At the same time, the gray-scale evaluation, online and rollback management of the model version is realized, forming a human-machine collaborative model optimization closed loop.

7. An automatic segmentation and quality control system for radiotherapy target areas based on multimodal medical imaging, characterized in that, The system includes: The image data acquisition module is used to acquire CT volumetric images and MR volumetric images of the target user, and read spatial information such as image pixel spacing, orientation, and origin. The multimodal image registration module is used to perform N4 field correction on MR based on CT and MR images, construct a fixed image mask on CT, complete the initial rigid body alignment through coarse Z-axis search, realize bidirectional field cropping through MR coverage drive, complete fine rigid body registration within the cropping domain, introduce flexible registration to compensate for local organ deformation, generate spatially aligned MR-CT image pairs, and output registration quality control images and log information at the same time. The target volume and organs at risk segmentation module is used to input the registered dual-channel CT-MR volume images into a segmentation model based on a U-shaped encoder-decoder network to generate segmentation results for clinical target volumes and multiple types of organs at risk. The segmentation quality control module is used to perform multiple inferences on the same case based on the probability output of the segmentation model to obtain multiple sets of probability data. It calculates voxel-level uncertainty through variance or entropy methods, constructs boundary rings Bs in combination with the segmentation mask Ms, aggregates voxel-level uncertainty in the ring region to obtain structural-level QC scores, and completes the quality control conclusion judgment of Pass / Needs-Review / Fail based on the validation set and the threshold is set. The results visualization and report output module integrates CT / MR images, CTV / OAR segmentation contours / masks, voxel-level uncertainty heatmaps, and structural QC scores, and displays them in a multi-layer overlay visualization on the terminal interface. It also outputs a structured report containing registration logs, QC scores for each structure, and quality control judgment results.

8. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the automatic segmentation and quality control method for radiotherapy target area based on multimodal medical images according to any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the automatic segmentation and quality control method for radiotherapy target area based on multimodal medical images as described in any one of claims 1 to 6.

Images