A Deep Learning-Based Automatic Delineation Method and System for High-Risk Clinical Target Areas and Organs at Risk in Cervical Cancer Brachytherapy

Abstract

This invention provides a method and system for automatically delineating high-risk clinical target areas and organs at risk in brachytherapy for cervical cancer based on deep learning, applied in the field of data processing technology. This application processes CT and MR images of cervical cancer patients using a localization network to generate 3D position and morphological information of the applicator, as well as image feature sequence information; classifies CT images of cervical cancer patients using the localization network to generate prior knowledge labels; processes the prior knowledge labels using a segmentation network to generate applicator localization target factors; processes the applicator localization target factors, the 3D position and morphological information of the applicator, and the image feature sequence information using a segmentation network to generate high-risk clinical target areas and organ segmentation parameters; and processes the high-risk clinical target areas and organ segmentation parameters to generate the automatically delineated results of high-risk clinical target areas and organs at risk for brachytherapy of cervical cancer.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an automatic delineation method and system for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning. Background Technology

[0002] In the treatment of cervical cancer, brachytherapy is an important method, and the localization of the applicator and the delineation of high-risk clinical target areas and organs at risk are crucial. However, current technology has significant shortcomings. The localization and delineation stages mainly rely on manual operation by physicians. This method is not only inefficient, but also results in inconsistent accuracy and significant individual variations due to differences in physician experience and expertise.

[0003] In image reconstruction, interference factors affect image quality, and the ability to distinguish soft tissues is also poor, increasing the difficulty of accurate localization and delineation. Existing automatic segmentation algorithms are also not ideal, failing to meet clinical requirements in terms of accuracy, requiring significant manual intervention, and consuming large amounts of hardware resources during computation, thus limiting their practical application. Overall, a new technical solution is urgently needed to address these issues, improving the accuracy and efficiency of delineating high-risk clinical target areas and organs at risk in brachytherapy for cervical cancer, reducing the workload of physicians, and mitigating the impact of individual differences.

[0004] 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 includes information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for automatically delineating high-risk clinical target areas and organs at risk in cervical cancer brachytherapy based on deep learning. This method overcomes, to some extent, the problems existing in current technologies. By utilizing a 3D U-Net network structure, the segmentation process is divided into two parts: a localization network and a segmentation network. The localization network extracts features through a region classifier, multi-layer surfaces, and pooling layers, and processes these features through fully connected layers to achieve 3D localization and reconstruction of the applicator, generating image feature sequences and prior knowledge labels. Next, the segmentation network generates applicator localization target factors based on the prior knowledge labels. Combining this with the 3D position, morphology, and image feature sequence information of the applicator, and through feature extraction, fusion, and decoder processing, it generates segmentation parameters for high-risk clinical target areas and organs. Finally, the segmentation parameters are format-converted, visualized, and automatically delineated using pre-defined image post-processing and edge detection algorithms to generate the results. This achieves full automation from image input to segmentation results, reducing manual intervention and improving segmentation accuracy and efficiency.

[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 automatically delineating high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning is provided. The method includes: acquiring CT and MR image information of cervical cancer patients, delineation requirements for high-risk clinical target areas and organs at risk, and an automatic delineation model for cervical cancer, wherein the automatic delineation model for cervical cancer includes a localization network and a segmentation network; processing the CT and MR image information of cervical cancer patients based on the localization network to generate 3D position and morphological information of the applicator, as well as image feature sequence information; classifying the CT image information of cervical cancer patients based on the localization network to generate prior knowledge labels; processing the prior knowledge labels based on the segmentation network to generate applicator localization target factors; processing the applicator localization target factors, the 3D position and morphological information of the applicator, and the image feature sequence information based on the segmentation network to generate segmentation parameter information for high-risk clinical target areas and organs; and processing the segmentation parameter information for high-risk clinical target areas and organs to generate automatic delineation results for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer.

[0008] Another aspect of this application discloses an automatic delineation device for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, characterized by comprising: an acquisition module for acquiring CT and MR image information of cervical cancer patients, delineation requirements for high-risk clinical target areas and organs at risk, and an automatic delineation model for cervical cancer, wherein the automatic delineation model for cervical cancer includes a localization network and a segmentation network; and a processing module for processing the CT and MR image information of cervical cancer patients based on the localization network to generate 3D position and morphological information of the applicator. The system generates high-risk clinical target areas and organ segmentation parameters by processing the CT image information of cervical cancer patients, including image feature sequence information, and using a localization network to classify and process CT image information of cervical cancer patients and generate prior knowledge labels. It also processes the prior knowledge labels using a segmentation network to generate applicator localization target factors, the 3D position and morphological information of the applicator, and image feature sequence information. Finally, it processes the high-risk clinical target areas and organ segmentation parameters to generate automatic delineation results of high-risk clinical target areas and organs at risk for brachytherapy of cervical cancer.

[0009] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for automatically delineating high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning.

[0010] This application provides a method and system for automatically delineating high-risk clinical target areas and organs at risk in cervical cancer brachytherapy based on deep learning. The server utilizes a 3D U-Net network structure, dividing the segmentation process into a localization network and a segmentation network. The localization network extracts features through a region classifier, multi-layer surfaces, and pooling layers, and processes these features through fully connected layers to achieve 3D localization and reconstruction of the applicator, generating image feature sequences and prior knowledge labels. Next, the segmentation network generates applicator localization target factors based on the prior knowledge labels. Combining this with the applicator's 3D position, morphology, and image feature sequence information, it generates high-risk clinical target area and organ segmentation parameters through feature extraction, fusion, and decoder processing. Finally, the segmentation parameters are format-converted, visualized, and automatically delineated using pre-defined image post-processing and edge detection algorithms to produce the high-risk clinical target areas and organs at risk for cervical cancer brachytherapy. This achieves full automation from image input to segmentation result, reducing manual intervention and improving segmentation accuracy and efficiency, thus possessing significant clinical application value.

[0011] 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

[0012] Figure 1 The flowchart illustrates an embodiment of the present application of an automatic delineation method for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning.

[0013] Figure 2 This illustration shows a schematic diagram of an automatic delineation device for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, according to an embodiment of this application. Detailed Implementation

[0014] 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.

[0015] The following is combined with Figure 1 This application describes an automatic delineation method for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, according to exemplary embodiments of this application. In one embodiment, this application also proposes an automatic delineation method and system for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning. Figure 1 This illustration schematically shows a flowchart of an automatic delineation method for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, according to an embodiment of this application. Figure 1 As shown, this method is applied to a server and includes:

[0016] S101 acquires CT and MR image information of cervical cancer patients, information on the delineation requirements of high-risk clinical target areas and organs at risk, and an automatic delineation model of cervical cancer.

[0017] In one implementation, a patient diagnosed with cervical cancer was admitted to the radiotherapy department of a large general hospital. To develop a precise brachytherapy plan, the medical team first arranged for the patient to undergo CT and MR scans. The CT scan used a multi-slice spiral CT scanner, with the scanning range extending from the pubic symphysis to the iliac crest level. Scanning parameters were set to a tube voltage of 120 kV, a tube current of 250 mA, and a slice thickness of 1 mm. This yielded images with high density resolution, clearly displaying structures such as bones and metal applicators, providing crucial information for subsequently determining the applicator's location and morphology. Simultaneously, an MR scan was performed using a 3.0T MRI scanner with a T2-weighted imaging sequence. Scanning parameters were a repetition time (TR) of 4000 ms, an echo time (TE) of 100 ms, and a slice thickness of 1 mm. MR images clearly distinguish soft tissues and provide excellent visualization of the boundaries between the cervical tumor and surrounding normal tissues, such as the uterus, vagina, rectum, and bladder, helping to accurately define high-risk clinical target areas and the extent of organs at risk. These two scanning methods were used to obtain CT and MR images of the patient's pelvic region. These image data were stored digitally in the hospital's Picture Archiving and Communication System (PACS) for further processing.

[0018] The information required for delineating high-risk clinical target areas and organs at risk is determined based on the patient's specific condition, tumor stage, physical status, and clinical treatment standards. The medical team will comprehensively analyze the patient's medical records, including pathology reports, tumor size, location, and degree of invasion. If the patient has stage II cervical cancer, and the tumor has invaded part of the parametrial tissue but not reached the pelvic wall, the high-risk clinical target area (HR-CTV) includes the cervix, uterine body, part of the vagina, and the invaded parametrial tissue. Organs at risk mainly include the bladder, rectum, sigmoid colon, and small intestine. Based on this information, the physician will clarify the specific requirements for delineating high-risk clinical target areas and organs at risk, such as the boundary range of the high-risk clinical target area and the areas of organs at risk that require special protection. This information is recorded in the form of electronic medical records and linked to the patient's image information, providing clinical guidance for subsequent automated delineation.

[0019] In another implementation, obtaining an automatic cervical cancer delineation model includes: acquiring a preset automatic delineation model, a training sample set, and data training-related information. The preset automatic delineation model includes a localization network and a segmentation network. The training sample set includes CT and MR image information of cervical cancer patients at different stages and with different body types. The data training-related information is used to characterize the delineation needs of high-risk clinical target areas and organs at risk for cervical cancer patients at different stages and with different body types. First, a preset automatic delineation model framework is constructed. The localization network of this preset model adopts a 3D-CNN (three-dimensional convolutional neural network) architecture, which has powerful image feature extraction capabilities. The model includes a localization network and a segmentation network. The localization network aims to accurately identify key information such as the location of the applicator from the patient's images, while the segmentation network is responsible for accurately segmenting high-risk clinical target areas and organs at risk based on the localization results and other relevant information.

[0020] The hospital utilized its extensive clinical data resources to collect CT and MR images of 300 cervical cancer patients at different stages and with varying body types. These patients covered early, intermediate, and late-stage cervical cancer cases, and their body types also differed, including thin, normal, and obese individuals. For example, Patient A was an early-stage cervical cancer patient with a thin build; Patient B was in the intermediate stage with a normal build; and Patient C was an late-stage patient with an obese build. This diverse data provided the model with a wide range of learning materials, enabling it to adapt to various clinical situations. Based on clinical experience and medical standards, detailed information on high-risk clinical target areas and organ delineation requirements was developed for each patient. For early-stage patients, the high-risk clinical target area was relatively limited, possibly only including the cervix and a small surrounding area; while for intermediate and late-stage patients, the high-risk clinical target area expanded as the tumor spread. Organs at risk, such as the bladder and rectum, also had different delineation requirements depending on their relative position to the tumor and the degree of potential radiation exposure. This information is organized into a structured data table, which is then linked to the corresponding CT and MR image information to form data training-related information.

[0021] The training sample set was processed according to a preset classification ratio to generate a training set and a validation set. The training sample set included CT and MR images of cervical cancer patients at different stages and body types. The training sample set was processed according to a preset classification ratio of 7:3. From 300 patient data points, 210 cases were randomly selected as the training set, and the remaining 90 cases were used as the validation set. The training set was used to train the model, allowing it to learn the relationship between image features and high-risk clinical target areas and organs at risk. The validation set was used to evaluate the model's performance on unseen data. During the partitioning process, it was ensured that both the training and validation sets included patient data from different stages and body types to guarantee the model's generalization ability under different conditions.

[0022] Data augmentation, random rotation, and flipping were performed on the training set to generate preprocessed CT and MR image information. To further enrich the data diversity of the training set and improve the robustness of the model, data augmentation operations were performed. For each CT and MR image in the training set, random rotation and flipping were applied. A CT image was randomly rotated by a certain angle (e.g., a random angle between 5° and 15°), or flipped horizontally or vertically. This process expanded the original 210 training data points, allowing the model to learn image features from more different angles and directions, enhancing its adaptability to various image variations.

[0023] The pre-set automatic delineation model is processed based on the training set and relevant training information to generate an automatic delineation prediction feature vector. This feature vector includes segmentation prediction target information for high-risk clinical target areas and organs at risk. During training, the model processes the data layer by layer through a localization network and a segmentation network. The localization network first extracts features from CT and MR images, identifying potential features of the applicator, high-risk clinical target areas, and organs at risk. The segmentation network then uses these features, combined with the delineation requirements for high-risk clinical target areas and organs at risk, to attempt to predict the segmentation results for these areas, generating the automatic delineation prediction feature vector. This vector contains the model's segmentation prediction target information for different regions, such as a preliminary judgment of the location, size, and shape of a tumor region. Based on the pre-processed CT and MR image information and the automatic delineation prediction feature vector, the pre-set automatic delineation model is processed to generate a trained automatic delineation model. The model continuously adjusts its parameters and optimizes its performance through backpropagation based on the differences between the predicted results and the actual delineation requirements for high-risk clinical target areas and organs at risk. After multiple rounds of iterative training, the model gradually learns to more accurately identify and segment high-risk clinical target areas and organs at risk, forming a trained automatic delineation model.

[0024] The trained automatic delineation model is processed using a validation set to generate validation results. These validation results are then used to further refine the trained model, resulting in an automated cervical cancer delineation model. The validation set is input into the trained model, which processes data from each patient in the validation set to generate corresponding automatic delineation results. These results constitute the validation set. The model's predicted high-risk clinical target areas and organ-at-risk contours are compared with standard results manually drawn by physicians to evaluate the model's accuracy, precision, and recall on the validation set. Based on the validation results, the trained automatic delineation model is evaluated and adjusted. If the model's performance on the validation set meets the expected standards (e.g., accuracy above 90%, precision, and recall meeting clinical requirements), then this model is designated as the final automated cervical cancer delineation model. If the model performs poorly, the reasons are further analyzed, training parameters are adjusted, or the model structure is improved, and training and validation are repeated until the model's performance meets the requirements. The final automated cervical cancer delineation model will be applied in actual clinical practice, providing accurate delineation support for high-risk clinical target areas and organ-at-risk contours in the treatment of cervical cancer patients. After multiple rounds of training and validation, the model was optimized based on the validation results, resulting in an automatic cervical cancer delineation model. This model is stored on the hospital's server, awaiting use to automatically delineate image data from new patients.

[0025] S102, based on the localization network, processes the CT and MR image information of cervical cancer patients to generate the 3D position and morphological information of the applicator, as well as the image feature sequence information.

[0026] In one implementation, a region classifier based on a localization network, along with multi-layered surfaces and pooling layers, performs feature extraction processing on CT and MR images of cervical cancer patients to generate target features and image feature sequences related to the applicator. A cervical cancer patient was admitted to the department. Before radiotherapy, the patient underwent CT and MR scans, obtaining the patient's CT and MR image information. The region classifier of the localization network then begins to work, acting as an intelligent filter to identify specific regions in the CT and MR images related to cervical cancer treatment, particularly information related to areas where the applicator might be present. For example, it focuses on the pelvic region where the applicator might appear in the image, initially distinguishing the approximate location of the applicator.

[0027] Next, multiple layers of surface and pooling layers begin to process these initially selected regions in depth. The surface layer analyzes the spatial structure in the image, transforming and extracting complex shape features such as bending and folding of the applicator. The pooling layer is responsible for filtering and integrating these features, removing some unimportant details while retaining key features, thus reducing the amount of data while highlighting the important features of the applicator. For example, through pooling layer processing, the key shape and position features of the applicator are highlighted, while some subtle noise or irrelevant tissue features are weakened. After this series of operations, target features and image feature sequence information related to the applicator are generated. These target features contain key information such as the material, shape, and position of the applicator, while the image feature sequence information records the feature changes related to the applicator at different levels and scales, laying the foundation for subsequent accurate determination of the applicator's location.

[0028] A fully connected layer based on a localization network processes target features associated with the applicator to generate the applicator's location information in CT and / or MR images. The fully connected layer of the localization network receives the previously generated applicator-associated target features. The neurons in the fully connected layer are interconnected with all neurons in the previous layer, allowing it to fully utilize the relationships between target features. Through complex calculations and analysis of these target features, the fully connected layer generates the applicator's location information in the CT and MR images.

[0029] Because CT and MR images each have their own advantages—CT images clearly display high-density materials such as metals, while MR images have a stronger ability to resolve soft tissues—the localization network references information from both CT and MR images to determine the applicator's location. For example, CT images can clearly show the approximate outline and location of the applicator, but may not clearly show the boundary between the applicator and the surrounding soft tissue; while MR images display soft tissue boundaries well, the applicator within them may be relatively blurry. The fully connected layer combines the advantages of both types of images, accurately calculating the precise location information of the applicator in both CT and MR images through analysis of target features, providing accurate data support for subsequent 3D reconstruction.

[0030] Based on the localization network, the applicator's location information in CT and / or MR images is segmented, delineated, and reconstructed in 3D to generate its 3D location and morphological information. After obtaining the applicator's location information in CT and MR images using the localization network, segmentation, delineation, and 3D reconstruction are performed. This process is explicitly performed simultaneously based on the location information from both CT and MR images. The localization network accurately delineates the applicator's outline in the images based on its location information, segmenting it from surrounding tissues and organs. For example, by comparing the high-density features of the applicator in the CT image with the differences in soft tissue in the MR image, the applicator's boundaries are determined more precisely.

[0031] Then, 3D reconstruction is performed using the results of these segmentations and delineations. The localization network integrates the two-dimensional positional information of the applicator in CT and MR images, and constructs a three-dimensional model of the applicator using a specific algorithm. In this process, the positional changes of the applicator at different image layers and its spatial relationship with surrounding tissues are fully considered, ultimately generating 3D positional and morphological information of the applicator. This 3D information can intuitively show the accurate position and shape of the applicator in the patient's body. Based on this information, the placement and dose distribution of the radiation source can be planned more precisely, thereby improving the effectiveness of radiotherapy and reducing damage to normal tissues.

[0032] S103, classifies and processes CT image information of cervical cancer patients based on a localization network to generate prior knowledge labels.

[0033] In one implementation, a localization network is used to classify CT images of cervical cancer patients, generating information on different anatomical structures and tissue types within the CT images. The convolutional layers in the localization network act like a set of detectors with different "fields of view," sliding across the image to capture features at various scales. Convolutional layers with smaller kernels capture subtle features such as vascular textures, while those with larger kernels focus on macroscopic features like organ outlines. Pooling layers downsample the feature maps output by the convolutional layers, reducing data volume while preserving key features and improving processing efficiency. Fully connected layers integrate the feature vectors processed by convolution and pooling to determine the anatomical structure or tissue type of each image region. Through this series of complex operations, the localization network can identify various anatomical structures and tissue types in the patient's CT images, such as distinguishing between different structures like the uterus, bladder, rectum, cervical tumors, and bones. This information is then labeled on the image using a specific encoding method, providing a foundation for subsequent processing.

[0034] The localization network acquires target features related to the applicator, high-risk clinical target areas, and organs at risk. Applicator-related target features characterize the applicator's location and morphology in CT images, while features related to high-risk clinical target areas and organs at risk characterize the inherent properties of these regions. During previous feature extraction and processing of CT images, the localization network has already acquired a large amount of feature information. Applicator-related target features include key information such as the applicator's material, shape, and location. Since the applicator is made of metal, it exhibits unique high-density features in CT images. The localization network can identify and extract these features to determine the applicator's location and approximate morphology in the image. For features related to high-risk clinical target areas and organs at risk, the localization network focuses on the tumor region and surrounding organs requiring special protection. For high-risk clinical target areas, boundary features are extracted, such as grayscale changes and texture features in the transition region between the tumor and normal tissue. For organs at risk, such as the bladder, features such as shape, location, and contrast with surrounding tissues are considered. Once these features are extracted, they are stored and represented in vector form, which facilitates subsequent integrated processing with anatomical structure and tissue type information.

[0035] Based on the target features related to the applicator, high-risk clinical target areas, and organs at risk, the localization network processes different anatomical and tissue type information in CT images to generate classification results. After acquiring various feature information, the localization network fuses the target features related to the applicator, high-risk clinical target areas, and organs at risk with the different anatomical and tissue type information in the previously generated CT images. Specifically, the localization network analyzes the positional relationship between the applicator and surrounding anatomical structures, determining whether the applicator is located near a high-risk clinical target area and its distance from organs at risk. For a specific region in a CT image, if it has the characteristics of a high-risk clinical target area for tumors and is close to the applicator, this region will be given a higher weight in the classification results and may be classified as a key area closely related to radiotherapy; while if a region belongs to an organ at risk and is close to the applicator, it will be marked as an area requiring special attention and protection. In this way, the localization network comprehensively evaluates and classifies each region in the CT image, generating a preliminary classification result that roughly divides the importance and potential risks of different regions during radiotherapy.

[0036] The classification results are processed to generate prior knowledge labels. The localization network refines and optimizes the classification results, removing unreasonable classifications caused by image noise or feature extraction errors. Thresholding methods are used to exclude regions with feature strengths below a certain threshold from key areas; or clustering algorithms are employed to merge adjacent regions with similar features into a single entity. After these processes, prior knowledge labels are generated based on the final classification results. These labels are marked on the CT image in a specific format, such as using different colors or values ​​to represent different region categories. Red indicates high-risk areas, possibly high-risk clinical target areas of tumors near the applicator; blue indicates areas of endangered organs requiring special protection; and green indicates relatively safe normal tissue areas. These prior knowledge labels provide important reference information for subsequent applicator localization, high-risk clinical target area and endangered organ segmentation, helping the segmentation network to more accurately identify and delineate the boundaries of each region, improving the accuracy and efficiency of radiotherapy planning.

[0037] S104, based on the segmentation network, processes the prior knowledge labels to generate the agent location target factor.

[0038] In one implementation, a segmentation network is used to extract features from prior knowledge labels and information on the delineation requirements of high-risk clinical target areas and organs at risk. Applicator-related features, high-risk clinical target area-related features, and organ-at-risk features are extracted from different angles and scales using waveform layers and pooling layers, respectively. Applicator-related features include the position and morphology of the applicator, and its material and density. High-risk clinical target area-related features include the boundary and extent of the high-risk clinical target area, and its tissue characteristics. Organ-at-risk features include the position and morphology of the organ at risk, and its relative position to the high-risk clinical target area. The waveform layers can be viewed as a set of special filters with different shapes and frequency responses, capable of "scanning" the input prior knowledge labels and information on the delineation requirements of high-risk clinical target areas and organs at risk from different angles.

[0039] For extracting the location and morphological features of the applicator, the waveform layer analyzes the applicator's imaging characteristics in CT or MR images. Applicators exhibit specific shapes and location distributions in images, and the waveform layer, through specific convolution operations, acts like matching different shaped "templates" on the image. If the applicator is elongated, convolution kernels with similar shapes in the waveform layer will produce a strong response, thus capturing the applicator's shape information. By adjusting parameters such as the size and stride of the convolution kernel, the applicator can be observed at different scales. Larger convolution kernels can capture the overall location and approximate shape of the applicator, while smaller kernels can focus on detailed features such as surface texture. Regarding the material and density features of the applicator, since its material exhibits unique grayscale or signal intensity characteristics in the image, the waveform layer can extract these features through convolution operations sensitive to changes in image grayscale values. Metallic applicators appear as high-density areas in CT images, and the waveform layer can detect the boundaries and internal grayscale changes of these high-density areas, thereby obtaining material and density-related features.

[0040] To extract the boundary and extent features of high-risk clinical target areas, the waveform layer leverages its sensitivity to image edges and textures. Differences in grayscale or signal intensity exist between high-risk clinical target areas and surrounding normal tissue. The waveform layer's convolutional kernels generate larger output values ​​at these points of difference, thus outlining the boundaries of the high-risk clinical target area. By using convolutional kernels of different scales, the extent of the high-risk clinical target area can be determined at both macroscopic and microscopic levels. Larger-scale kernels can determine the approximate extent of the high-risk clinical target area, while smaller-scale kernels can refine the boundary, capturing subtle changes at the edge of the high-risk clinical target area. For the tissue characteristics of high-risk clinical target areas, such as tissue density and metabolic activity, the waveform layer can extract these features by analyzing the grayscale distribution patterns of different regions in the image and their contrast with surrounding tissues. Tumor tissue and normal tissue have different metabolic activities, manifesting as different grayscale or signal characteristics in the image. The waveform layer can identify these differences in features, thereby obtaining tissue characteristic information of the high-risk clinical target area.

[0041] When extracting the location and morphological features of organs at risk, the waveform layer operates based on the shape characteristics and positional distribution of the organs in the image. The bladder exhibits a specific shape and location in the image; the waveform layer captures its contour information using convolutional kernels that match the bladder's shape, while simultaneously utilizing convolutional kernels of different scales to determine its size and position. For the relative positional features of organs at risk and high-risk clinical target areas, the waveform layer comprehensively analyzes the positional information of both. By comparing their coordinate positions in the image and their relationships such as distance and angle, relative positional features are extracted. The positional information of organs at risk and high-risk clinical target areas is converted into vector representations, and the waveform layer can extract feature vectors reflecting their relative positional relationships through specific operations.

[0042] The pooling layer follows the waveform layer, further processing the features extracted by the waveform layer. The main function of the pooling layer is to reduce data volume while preserving key features, improving computational efficiency and model robustness. For features of the applicator, high-risk clinical target area, and organs at risk, the pooling layer uses operations such as max pooling or average pooling. Max pooling selects the maximum value within a local region as the output, thus highlighting the most significant features. In the applicator's feature map, max pooling can retain the most representative shape or positional features of the applicator, ignoring some subtle changes, thereby reducing data volume while preserving key information. Average pooling calculates the average value within a local region as the output, which can smooth the features and reduce the impact of noise. Through the processing of the pooling layer, the relevant features of the applicator, high-risk clinical target area, and organs at risk extracted from different angles and scales are further filtered and integrated, preparing for subsequent feature fusion.

[0043] Feature fusion processing is performed on applicator-related features, high-risk clinical target area-related features, and organ-at-risk-related features to generate applicator localization target factors. These target factors characterize the relationship between applicator localization and the delineation of high-risk clinical target areas and organs at risk. After processing with waveform layers and pooling layers, relevant features of the applicator, high-risk clinical target area, and organs at risk are obtained. Next, these features need to be fused to generate the applicator localization target factors. The feature fusion process is like piecing together different puzzle pieces to form a more complete and representative "image" to accurately characterize the relationship between applicator localization and the delineation of high-risk clinical target areas and organs at risk.

[0044] This process begins by combining applicator-related features, high-risk clinical target area-related features, and organ-at-risk-related features. For example, feature vectors corresponding to the applicator's location, morphology, material, and density are combined with the boundary and extent features and tissue characteristic feature vectors of the high-risk clinical target area, as well as the location and morphology features of the organ-at-risk-relationship with the high-risk clinical target area. Then, a specific fusion algorithm is used for processing. The fusion method can employ a weighted summation approach, assigning different weights to each feature vector based on its importance in determining applicator location and the delineation relationship between the high-risk clinical target area and the organ-at-risk-relationship. Features with a significant impact on applicator location and delineation are given higher weights, while those with less impact are given lower weights. For instance, if the boundary and extent features of the high-risk clinical target area are crucial for accurate applicator location in the current case, then these feature vectors will have relatively higher weights during the fusion process.

[0045] Let A be the location and morphology feature vector of the source device, B be the material and density feature vector, C be the boundary and extent feature vector of the high-risk clinical target area, D be the tissue characteristic feature vector, E be the location and morphology feature vector of the organ at risk, and F be the relative position feature vector with respect to the high-risk clinical target area. After weighting, assuming the weight of A is ω1, the weight of B is ω2, the weight of C is ω3, the weight of D is ω4, the weight of E is ω5, and the weight of F is ω6, then the fused feature vector G can be expressed as: G = ω1×A + ω2×B + ω3×C + ω4×D + ω5×E + ω6×F.

[0046] Besides weighted summation, more complex neural network structures, such as multilayer perceptrons (MLPs), can be used for fusion. MLPs can learn more complex nonlinear relationships between features, further improving the fusion effect. These combined feature vectors are input into the MLP, processed by multiple layers of neurons, and output as a new feature vector. This vector is the donor's target factor for localization.

[0047] Applicator localization target factors integrate multiple aspects of information about the applicator, high-risk clinical target area, and organs at risk. They clearly reflect the positional relationship, spatial arrangement, and mutual influence between the applicator and these elements. For example, applicator localization target factors can determine whether the applicator is accurately located near the high-risk clinical target area and how far it maintains a safe distance from organs at risk. This provides crucial information for accurately determining the applicator's location and delineating the high-risk clinical target area and organs at risk, helping physicians develop more precise and safer radiotherapy plans.

[0048] S105 processes the applicator localization target factors, the applicator's 3D position and morphological information, and image feature sequence information based on the segmentation network to generate high-risk clinical target areas and organ segmentation parameter information.

[0049] In one implementation, a segmentation network is used to extract and fuse features from applicator localization target factors, the applicator's 3D position and morphological information, and image feature sequence information to generate an automatically delineated predictive feature vector. The segmentation network acts as an intelligent analysis system, receiving "clues" such as applicator localization target factors, applicator's 3D position and morphological information, and image feature sequence information. Applicator localization target factors include information about the relationship between the applicator and high-risk clinical target areas and organs at risk; 3D position and morphological information clearly define the applicator's specific location and shape within the patient's body; and image feature sequence information encompasses various feature details extracted from the image. The segmentation network then performs feature extraction and feature fusion on this information.

[0050] During feature extraction, convolutional layers in the network act like "detectors," meticulously analyzing the input information. For the material characteristics of the applicator, convolutional layers can capture the unique representation of the applicator's material in the image through sensitive computations on changes in image grayscale values, such as the grayscale difference between a metal applicator and surrounding tissue. For high-risk clinical target areas and organs at risk, convolutional layers identify the unique features of these regions based on their boundaries, textures, and other characteristics. Just as when identifying the boundaries of high-risk clinical target areas, convolutional layers can keenly perceive changes in grayscale or signal intensity between high-risk clinical target areas and normal tissue.

[0051] Next, feature fusion is performed, much like piecing together different puzzle pieces to form a more complete "puzzle." The segmentation network integrates the extracted features of the applicator, high-risk clinical target area, and organs at risk. For example, it combines the positional features of the applicator with the boundary features of the high-risk clinical target area to determine their relative positional relationship; it fuses the morphological features of the organs at risk with their relative positional features to the high-risk clinical target area to comprehensively understand the spatial layout of the organs at risk and the high-risk clinical target area. Through this fusion, a novel automatically delineated predictive feature vector is generated. This vector differs from the automatically delineated predictive feature vectors generated during the model training phase; it is generated based on the specific image and related information of the current patient, making it more targeted and able to reflect the individual tumor and organ characteristics of the patient.

[0052] The segmentation network-based decoder processes the automatically delineated predicted feature vector through upsampling and convolution to generate segmentation parameters for high-risk clinical target areas and organs. After obtaining the automatically delineated predicted feature vector for the specific patient, the segmentation network's decoder begins its work. The decoder can be viewed as a "restore and refine" tool, processing the automatically delineated predicted feature vector through upsampling and convolution. Upsampling is like enlarging a reduced image; it expands the low-resolution information in the automatically delineated predicted feature vector to a resolution close to that of the original image, restoring the image's size information. In this process, some detailed information may be lost, so convolution is needed for further refinement.

[0053] Convolution operations perform in-depth analysis and processing of the expanded features. It re-scans these features, using previously learned knowledge to more precisely define the boundaries of high-risk clinical target areas and organs at risk, and to more accurately determine their shape and location. After multiple processing steps of upsampling and convolution, segmentation parameters for high-risk clinical target areas and organs are finally generated. These parameters contain detailed information such as the location, shape, and size of high-risk clinical target areas and organs at risk, including boundary coordinates of the high-risk clinical target areas and contour descriptions of the organs at risk. This parameter information is crucial for subsequent automatic delineation, allowing doctors to accurately delineate the contours of high-risk clinical target areas and organs at risk on the patient's images, providing vital support for developing precise radiotherapy plans.

[0054] S106 processes the segmentation parameter information of high-risk clinical target areas and organs to generate automatic delineation results of high-risk clinical target areas and organs at risk for brachytherapy of cervical cancer.

[0055] In one implementation, the segmentation parameters of high-risk clinical target areas and organs are converted into formats and visualized to generate tumor contour images and images of organs at risk. The tumor contour images and images of organs at risk are highlighted with lines of different colors. After obtaining the segmentation parameters, the system first performs format conversion. Since the segmentation parameters are initially in the form of a computer-readable digital matrix, they cannot be directly understood by doctors and therefore need to be converted into a format suitable for image display. This is analogous to translating a complex code into human-readable text. For example, the digital information describing the location and shape of high-risk clinical target areas and organs at risk is converted into data conforming to common image formats (such as DICOM format).

[0056] After format conversion, the system performs visualization. Based on the converted information, the system draws tumor outlines and images of organs at risk on the image. For easy differentiation, the tumor outline is highlighted with red lines, while organs at risk are highlighted with blue lines. In this patient's case, when the system receives the segmentation parameters for the high-risk clinical target area, it precisely outlines the area where the tumor may exist on the corresponding CT or MR image with red lines, clearly showing the approximate shape and extent of the tumor; for organs at risk such as the bladder and rectum, their outlines are depicted with blue lines. In this way, doctors can easily distinguish the location of the tumor and organs at risk on the image.

[0057] Based on preset image post-processing and edge detection algorithms, tumor contour images and images of organs at risk are processed to generate high-risk clinical target areas and automatically delineated organs at risk for cervical cancer brachytherapy. The preset image post-processing algorithm is mainly used to optimize image quality and improve the accuracy and reliability of the automatic delineation results. The preset image post-processing algorithm can employ median filtering. This algorithm processes each pixel in the image to remove noise interference. In this patient's image, due to equipment acquisition or transmission issues, there may be some isolated noise points that could affect the accurate judgment of tumor and organ at risk boundaries. The median filtering algorithm selects a neighborhood of a specific size (e.g., a 3×3 or 5×5 pixel matrix) centered on each pixel, sorts the pixel values ​​within the neighborhood, and takes the median value as the new value for that pixel. In this way, the image can be effectively smoothed, noise removed, and the contours of the tumor and organs at risk made clearer.

[0058] Morphological processing algorithms, including erosion and dilation operations, can also be employed. For tumor contour images, erosion can remove small protrusions that may be caused by noise or image artifacts, making the contour smoother and more accurate. Dilation can expand regions that may be narrowed or discontinuous due to noise or other reasons, ensuring a complete representation of the tumor. When processing the tumor contour image of this patient, erosion was performed first to remove small protrusions that should not be present, followed by dilation to make the tumor contour more continuous and complete. A similar approach is used for images of organs at risk to ensure the accuracy of their contours.

[0059] Edge detection algorithms play a crucial role in determining the precise boundaries of tumors and organs at risk. The Canny edge detection algorithm is a commonly used method. It first applies a Gaussian filter to the image to further smooth it and reduce the impact of noise on edge detection. Then, it calculates the gradient intensity and direction of each pixel in the image. In this patient's image, there are differences in grayscale or signal intensity between the tumor and normal tissue, and between the organs at risk and surrounding tissue; these differences manifest as large gradient values ​​in the gradient calculation. Based on this gradient information, the Canny algorithm refines the edges using non-maximum suppression, retaining only pixels with significant gradient changes as edge points. Finally, it determines the final edges through double thresholding and hysteresis tracking. This allows for accurate identification of the boundaries between the tumor and organs at risk, providing a precise basis for automatic delineation.

[0060] After processing with preset image post-processing and edge detection algorithms, the system integrates this processed image information to generate automatic delineation results of high-risk clinical target areas and organs at risk for cervical cancer brachytherapy. In this result, the boundaries of high-risk clinical target areas and organs at risk are more precise and clear, allowing doctors to directly formulate radiotherapy plans based on these results. The system presents the automatically delineated results to doctors in an intuitive format, such as displaying the outlines of different tissues in a layered manner within an electronic medical record system. Doctors can clearly see the positional relationship between the tumor and surrounding organs at risk, thereby more accurately planning the radiotherapy dose and irradiation range, improving the effectiveness of radiotherapy, and minimizing damage to normal tissues.

[0061] The server utilizes a 3D U-Net network structure to divide the segmentation process into two parts: a localization network and a segmentation network. It acquires CT and MR image information of cervical cancer patients, clinical delineation requirements, and an automatic delineation model. The localization network extracts features through a region classifier, multi-layer surfaces, and pooling layers, which are then processed by fully connected layers to achieve 3D localization and reconstruction of the applicator, generating image feature sequences and prior knowledge labels.

[0062] Next, the segmentation network generates applicator localization target factors based on prior knowledge labels. Then, combining the applicator's 3D position, morphology, and image feature sequence information, and through feature extraction, fusion, and decoder processing, it generates high-risk clinical target areas and organ segmentation parameters. Finally, the segmentation parameters are format-converted, visualized, and, using pre-defined image post-processing and edge detection algorithms, automatically delineate high-risk clinical target areas and organs at risk for cervical cancer brachytherapy.

[0063] In terms of model training, a pre-defined model, training sample set, and relevant information are acquired. The model is divided into training and validation sets, and the training set is augmented before training. After multiple rounds of training and validation optimization, the final automatic cervical cancer delineation model is obtained. This approach achieves full automation from image input to segmentation results, reducing manual intervention and improving segmentation accuracy and efficiency, thus possessing significant clinical application value.

[0064] In one implementation, such as Figure 2 As shown, this application also provides an automated delineation device for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, comprising:

[0065] The acquisition module 201 is used to acquire CT image information and MR image information of cervical cancer patients, information on the delineation requirements of high-risk clinical target areas and organs at risk, and an automatic delineation model of cervical cancer. The automatic delineation model of cervical cancer includes a localization network and a segmentation network.

[0066] The processing module 202 is used to process CT and MR image information of cervical cancer patients based on a localization network to generate 3D position and morphological information of the applicator, as well as image feature sequence information; classify CT image information of cervical cancer patients based on a localization network to generate prior knowledge labels; process the prior knowledge labels based on a segmentation network to generate applicator localization target factors; process the applicator localization target factors, 3D position and morphological information of the applicator, and image feature sequence information based on a segmentation network to generate high-risk clinical target areas and organ segmentation parameter information; and process the high-risk clinical target areas and organ segmentation parameter information to generate automatic delineation results of high-risk clinical target areas and organs at risk for brachytherapy of cervical cancer.

[0067] All 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 of the automatic delineation method, electronic device, electronic device, and readable storage medium for evaluating high-risk clinical target areas and organs at risk in deep learning-based brachytherapy of cervical cancer are basically similar to the above-described embodiments of the automatic delineation method for high-risk clinical target areas and organs at risk in deep learning-based brachytherapy of cervical cancer, so the description is relatively simple. Relevant parts can be referred to in the description of the above-described embodiments of the automatic delineation method for high-risk clinical target areas and organs at risk in deep learning-based brachytherapy of cervical cancer.

Claims

1. A method for automatically delineating high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, characterized in that, include: Acquire CT and MR image information of cervical cancer patients, information on the delineation requirements of high-risk clinical target areas and organs at risk, and an automatic delineation model for cervical cancer, which includes a localization network and a segmentation network. Based on the localization network, CT and MR image information of cervical cancer patients are processed to generate 3D position and morphological information of the applicator, as well as image feature sequence information; This paper describes a method for classifying CT images of cervical cancer patients using a localization network to generate prior knowledge labels. The method includes classifying CT images of cervical cancer patients based on the localization network, generating different anatomical structures and tissue types in the CT images; acquiring target features related to the applicator, high-risk clinical target areas, and organs at risk; using the applicator-related target features to characterize the applicator's location and morphology in the CT images, and using the high-risk clinical target areas and organs at risk-related features to characterize the features of these areas; processing the different anatomical structures and tissue types in the CT images based on the applicator-related target features, high-risk clinical target areas, and organs at risk-related features to generate classification results; and processing the classification results to generate prior knowledge labels. This process utilizes a segmentation network to process prior knowledge labels and information on the delineation requirements of high-risk clinical target areas and organs, generating applicator positioning target factors. This includes feature extraction from prior knowledge labels and information on the delineation requirements of high-risk clinical target areas and organs at risk using a segmentation network. Applicator-related features, high-risk clinical target area-related features, and organ-at-risk-related features are extracted from different angles and scales using waveform layers and pooling layers. Applicator-related features include the applicator's location and morphology, and its material and density. High-risk clinical target area-related features include the boundary and extent features and tissue characteristics of the high-risk clinical target area. Organ-at-risk-related features include the organ's location and morphology, and its relative position to the high-risk clinical target area. Feature fusion processing is then performed on these applicator-related features, high-risk clinical target area-related features, and organ-at-risk-related features to generate applicator positioning target factors. These applicator positioning target factors characterize the relationship between applicator positioning and the delineation of high-risk clinical target areas and organs at risk. Based on the segmentation network, the applicator localization target factors, the 3D position and morphological information of the applicator, and the image feature sequence information are processed to generate high-risk clinical target areas and organ segmentation parameter information. The high-risk clinical target area and organ segmentation parameter information are processed to generate automatic delineation results of high-risk clinical target areas and organs at risk for brachytherapy of cervical cancer.

2. The method as described in claim 1, characterized in that, Obtain an automatically drawn model of cervical cancer, including: The system acquires a preset automatic delineation model, a training sample set, and data training-related information. The preset automatic delineation model includes a localization network and a segmentation network. The training sample set includes CT and MR image information of cervical cancer patients at different stages and with different body sizes. The data training-related information is used to characterize the high-risk clinical target areas and organs at risk delineation requirements of cervical cancer patients at different stages and with different body sizes. The training sample set is processed based on a preset classification ratio to generate a training set and a validation set. The training sample set includes CT and MR image information of cervical cancer patients with different stages and body types. Data augmentation, random rotation, and flipping are performed on the training set to generate preprocessed CT and MR image information. The pre-set automatic delineation model is processed based on the training set and data training related information to generate an automatic delineation prediction feature vector. The automatic delineation prediction feature vector includes segmentation prediction target information for high-risk clinical target areas and organs at risk. The pre-processed CT and MR image information and the automatic delineation prediction feature vector are used to process the preset automatic delineation model to generate the trained automatic delineation model. The trained automatic drawing model is processed based on the validation set to generate validation results; The trained automatic delineation model is processed based on the validation results to generate an automatic delineation model for cervical cancer.

3. The method as described in claim 1, characterized in that, Based on a localization network, CT and MR image information of cervical cancer patients are processed to generate 3D position and morphological information of the applicator, as well as image feature sequence information, including: A region classifier based on a localization network, along with multi-layered curved surfaces and pooling layers, is used to extract features from CT and MR images of cervical cancer patients, generating target features and image feature sequence information related to the applicator. The fully connected layer based on the localization network processes the target features related to the applicator to generate the location information of the applicator in CT and / or MR images; Based on the localization network, the position information of the applicator in CT and / or MR images is segmented, delineated, and reconstructed in 3D to generate the 3D position and shape information of the applicator.

4. The method as described in claim 1, characterized in that, Based on the segmentation network, the applicator localization target factors, the applicator's 3D position and morphological information, and image feature sequence information are processed to generate high-risk clinical target areas and organ segmentation parameters, including: Based on the segmentation network, feature extraction and feature fusion processing are performed on the applicator localization target factor, the 3D position and morphological information of the applicator and the image feature sequence information to generate an automatically delineated prediction feature vector. The segmentation network-based decoder processes the automatically delineated predicted feature vectors through upsampling and convolution to generate segmentation parameter information for high-risk clinical target areas and organs.

5. The method as described in claim 1, characterized in that, The high-risk clinical target area and organ segmentation parameter information are processed to generate automatic delineation results of high-risk clinical target areas and organs at risk for cervical cancer brachytherapy, including; The high-risk clinical target area and organ segmentation parameter information are format converted and visualized to generate tumor contour image information and organ at risk image information. The tumor contour image information and organ at risk image information are highlighted with lines of different colors. The tumor contour image information and the image information of organs at risk are processed to generate high-risk clinical target areas and organs at risk for brachytherapy of cervical cancer.

6. An automated delineation device for high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, characterized in that, For implementing the method of claim 1, the apparatus includes: The acquisition module is used to acquire CT and MR image information of cervical cancer patients, information on the delineation requirements of high-risk clinical target areas and organs at risk, and an automatic delineation model of cervical cancer. The automatic delineation model of cervical cancer includes a localization network and a segmentation network. The processing module is used to process CT and MR image information of cervical cancer patients based on a localization network to generate 3D position and morphological information of the applicator, as well as image feature sequence information; classify CT image information of cervical cancer patients based on a localization network to generate prior knowledge labels; process the prior knowledge labels based on a segmentation network to generate applicator localization target factors; process the applicator localization target factors, 3D position and morphological information of the applicator, and image feature sequence information based on a segmentation network to generate high-risk clinical target areas and organ segmentation parameters; and process the high-risk clinical target areas and organ segmentation parameters to generate automatic delineation results of high-risk clinical target areas and organs at risk for cervical cancer brachytherapy.

7. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; Wherein, the first processor is configured to execute the method for automatically delineating high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, as described in any one of claims 1 to 5, by executing the executable instructions.

8. 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 method for automatically delineating high-risk clinical target areas and organs at risk in brachytherapy of cervical cancer based on deep learning, as described in any one of claims 1 to 5.

Images