Cervical cancer target area automatic sketching system based on deep learning

Abstract

The invention relates to the technical field of medical image processing, in particular to a cervical cancer target area automatic sketching system based on deep learning, which comprises a data acquisition module, a deep learning model construction module, a model training and optimizing module and an automatic sketching and dose calculation module, acquiring and processing a CT image of a patient, an expert sketching result and a dose distribution diagram, and dividing a data set; the deep learning model construction module comprises a segmentation and registration model construction unit which is used for respectively constructing a network structure and setting parameters; the model training and optimizing module is used for training a model by utilizing a training set, expanding data volume by adopting a data transformation strategy and adjusting the model according to a verification set result; and the automatic sketching and dose calculation module performs automatic sketching and dose calculation by using the trained model. According to the method, the sketching efficiency and consistency are improved, the sketching accuracy is improved, the accumulated dose is accurately calculated, the precision and standardization of cervical cancer radiotherapy are facilitated, and the radiotherapy quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to an automatic delineation system for cervical cancer target areas based on deep learning. Background Art

[0002] Cervical cancer is one of the most common malignant tumors in women, which seriously threatens the life and health of patients. Radiotherapy plays an important role in the treatment of cervical cancer, and accurate delineation of target areas and organs at risk and dose calculation are the key links for the success of radiotherapy.

[0003] Traditionally, the delineation of cervical cancer target areas and organs at risk mainly relies on the manual operation of radiotherapists. This process is not only time-consuming and laborious, but also usually requires doctors to spend a lot of time carefully analyzing each CT image. In addition, due to differences in experience and judgment criteria among different doctors, it is difficult to ensure the consistency of the delineation results. This inconsistency may affect the accuracy of the radiotherapy plan and thus affect the treatment effect.

[0004] In terms of dose calculation, accurate calculation of the cumulative dose of internal and external irradiation is crucial for evaluating the efficacy of radiotherapy and reducing complications. However, traditional methods face many challenges in dealing with the dose accumulation of different irradiation methods, such as the registration of images under different scanning conditions, which can easily lead to inaccurate dose calculations, which may result in insufficient dose in the tumor area and affect the treatment effect, or cause excessive irradiation of endangered organs and cause serious complications.

[0005] Therefore, to address the above problems, an automatic cervical cancer target delineation system based on deep learning was proposed. Summary of the invention

[0006] The purpose of the present invention is to provide an automatic cervical cancer target area delineation system based on deep learning to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The deep learning-based automatic delineation system for cervical cancer target areas includes:

[0009] Data acquisition module: The data acquisition module determines the inclusion and exclusion criteria for cervical cancer patients, collects internal and external irradiation positioning CT images of patients, performs expert delineation of target areas and organs at risk in CT images and collects dose distribution maps, and divides the data into training sets, validation sets, and test sets after anonymization.

[0010] Deep learning model building module; Deep learning model building module includes:

[0011] Segmentation model building unit: The segmentation model building unit collects and preprocesses the internal and external irradiation positioning CT images of cervical cancer patients, screens the data and organizes the expert delineation results; based on the V-Net model, builds an input layer to receive 128×128×64 three-dimensional volume data, builds an encoding network including convolution kernels, jump connections and Attention Block modules, and then uses transposed convolution to build a decoding network; randomly initializes model parameters, sets the learning rate, number of iterations and batch size;

[0012] Registration model building unit: The registration model building unit inputs two sets of volume data, the target image and the source image, extracts features through the convolution layer, uses the dilated convolution to expand the receptive field, and then calculates the deformation field based on the objective function including reversible deformation loss, registration similarity measurement loss and deformation field continuity loss. It uses the Adam optimization algorithm combined with the learning rate decay strategy, uses the data transformation strategy to expand the data volume, and adjusts the model based on the training set and the validation set.

[0013] Model training and optimization module: The model training and optimization module uses the training set to train the segmentation model and the registration model, adopts the data transformation strategy to expand the amount of training data, updates the model parameters through the optimization algorithm, and adjusts the model structure and training strategy according to the evaluation results of the validation set;

[0014] Automatic outlining and dose calculation module: Use the trained segmentation model for automatic outlining, evaluate the accuracy of outlining by calculating the DSC and 95HD indicators, use the registration model to register the images to obtain the spatial transformation relationship, apply the spatial transformation relationship to the dose map to perform a unified coordinate system and cumulative calculation of internal and external irradiation doses, and finally obtain the cumulative total dose map.

[0015] As a preferred solution, in the segmentation model construction unit, the encoding network convolution operation formula is: ;

[0016] in, represents the output feature map, Represents the coordinate position in the three-dimensional volume data, represents the convolution kernel size, represents the convolution kernel weight, Represents input data;

[0017] After the convolution operation through the encoding network, it is activated by the ReLU function;

[0018] The Attention Block calculation method is:

[0019] set up and are feature maps of different scales; ,in, is a learnable weight matrix used to calculate Query; ,in, is a learnable weight matrix used to calculate Key; ,in, is a learnable weight matrix used to calculate Value; ,in, for The dimension of Represents the attention matrix, and the softmax function is used to normalize the attention matrix;

[0020] Then the attention-weighted feature map is obtained by the following formula : .

[0021] As a preferred solution, in the registration model construction unit, the reversible deformation loss formula is: ;

[0022] in, Represents the coordinate position of the pixel in the image. Indicates that the source image Align to target image The spatial shape transformation, Indicates that the target graph Align to source image The spatial shape transformation, Represents the calculation norm, which is used to measure the difference in pixel position before and after spatial shape transformation;

[0023] The registration similarity measure loss formula is: ;

[0024] in, represents the target image, represents the source image, Represents the source image After registration transformation After the image, Represents the target image After registration transformation After the image, , , are the edge entropy and joint entropy of the corresponding images respectively. The formulas for calculating the edge probability density and joint probability density are as follows: ;

[0025] in, Representing images The pixel value is The marginal probability density of Representing images The edge entropy of the image The uncertainty or amount of information; ;

[0026] in, Representing images The pixel value is The marginal probability density of Representing images The marginal entropy of ;

[0027] in, Representing images and The joint entropy of

[0028] The deformation field continuity loss formula is: ;

[0029] in, represents the gradient operator, Indicates For the target image, is the deformation field estimated by the registration of the source image, Indicates For the target image, is the deformation field estimated by the registration of the source image, and Denote the deformation field and At coordinate position The smaller the gradient norm, the more continuous the deformation.

[0030] As a preferred solution, in the model training and optimization module, the data transformation strategy includes horizontal image flipping, random extraction, scale transformation and rotation; the parameter update formula of the Adam optimization algorithm is: ;in, Indicates The first-order moment estimate of the iteration, is the attenuation rate, Indicates The first-order moment estimate of the iteration, Indicates The gradient of the iteration; ;in, Indicates The second-order moment estimate of the iteration, is the attenuation rate, Indicates The second-order moment estimate of the iteration, Indicates The square of the gradient of the iteration;

[0031] Using the formula The first-order moment estimate is corrected, where is the result after correcting the first-order moment estimate;

[0032] Using the formula The second-order moment estimate is corrected, where is the result after correcting the second-order moment estimate; ;in, represents the model parameters of the tth iteration, Indicates The model parameters of the iteration, is the learning rate, which determines the step size of parameter update. is a constant.

[0033] As a preferred solution, in the automatic delineation and dose calculation module, the DSC calculation formula is: ;

[0034] Among them, TP represents the number of pixels correctly classified as positive examples, that is, the number of pixels that actually belong to the target area and are classified as pixels in the target area by the algorithm; FP represents the number of pixels incorrectly classified as positive examples, that is, the number of pixels that do not actually belong to the target area but are classified as pixels in the target area by the algorithm; FN represents the number of pixels incorrectly classified as negative examples, that is, the number of pixels that actually belong to the target area but are classified as pixels outside the target area by the algorithm. DSC is used to measure the similarity between the automatic delineation result and the expert delineation gold standard. The closer the value is to 1, the higher the delineation accuracy.

[0035] 95HD calculation first calculates the distance matrix of the predicted result and the actual result edge point set using the following formula: ;

[0036] in, Represents the edge point set of the prediction result, represents the real result edge point set, Indicate point and The distance between points, here we use L2 norm distance or Euclidean distance, Represents the point set Each point-to-point set in The matrix composed of the distances to the nearest points in ;

[0037] The 95th percentile is taken as the 95HD value.

[0038] It can be seen from the technical solution provided by the present invention that the automatic delineation system for cervical cancer target area based on deep learning provided by the present invention has the following beneficial effects: based on the automatic delineation and dose calculation module, the present invention utilizes a trained segmentation model to quickly and automatically delineate a large number of internal and external irradiation CT images of cervical cancer patients, and the automatic delineation function is based on a unified model and algorithm, which is not affected by the subjective factors of the doctor, and can provide consistent delineation results for each patient, thereby improving the consistency of delineation, helping to formulate a more standardized radiotherapy plan, and improving the quality of radiotherapy; the segmentation model adopts a variety of advanced technologies in the construction process; based on the V-Net model, its three-dimensional convolution kernel can effectively process three-dimensional volume data, better adapt to the characteristics of cervical cancer radiotherapy images, extract richer image features, and Attention The application of Block enables the model to focus more on key areas related to the target area and organs at risk; by processing feature maps of different scales and assigning attention weights, the model can highlight important features and suppress the interference of irrelevant information, so as to more accurately outline the target area and organs at risk. By calculating indicators such as DSC and 95HD to evaluate the outline results, the difference between automatic outline and expert outline gold standard can be quantitatively measured; the registration model can accurately register CT images of the same patient at different times, different scanning conditions or different treatment stages to obtain accurate spatial transformation relationships; when dealing with image differences caused by factors such as patient position changes and organ deformation, the registration model can find the best spatial match by calculating reversible deformation loss, registration similarity measure loss and deformation field continuity loss, so as to ensure the accuracy of dose calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram of the overall structure of the automatic delineation system for cervical cancer target areas based on deep learning of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0041] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0042] like Figure 1 As shown, an embodiment of the present invention provides an automatic delineation system for cervical cancer target areas based on deep learning, including a data acquisition module, a deep learning model construction module, a model training and optimization module, and an automatic delineation and dose calculation module.

[0043] In this embodiment, the data acquisition module determines the inclusion and exclusion criteria of cervical cancer patients, collects internal and external irradiation positioning CT images of patients, performs expert delineation of target areas and organs at risk in CT images and collects dose distribution maps, and divides the data into training set, validation set and test set after anonymization processing;

[0044] Furthermore, the data acquisition module specifically includes:

[0045] 1. Determine inclusion and exclusion criteria

[0046] Inclusion criteria:

[0047] Age range: Consider the incidence and biological characteristics of cervical cancer in different age groups. For example, the age range is set between [specific lower age limit, specific upper age limit], because the age group with a high incidence of cervical cancer has certain characteristics, and cases outside this range may have different pathological manifestations or responses to treatment;

[0048] Pathological type: clearly define the included cervical cancer pathological types, such as squamous cell carcinoma, adenocarcinoma and other common types, because different pathological types of cervical cancer may differ in imaging manifestations, tumor growth patterns and sensitivity to radiotherapy;

[0049] Clinical staging: Determine the staging range to be included based on the International Federation of Gynecology and Obstetrics (FIGO) staging system or other widely recognized staging standards. For example, only patients in stages I-III may be included, because early and middle stage patients are more suitable for radiotherapy, and their tumor size, location, and metastasis are relatively clear, which is conducive to research and model training;

[0050] Physical condition assessment: including the patient's general physical condition (such as Karnofsky score above [specific score]), whether there are other serious underlying diseases (such as severe cardiopulmonary insufficiency, uncontrolled diabetes, etc.), because poor physical condition may affect the implementation of radiotherapy and the patient's tolerance to treatment, and may also interfere with imaging images;

[0051] Exclusion criteria establishment:

[0052] Image quality issues: If the CT image has severe artifacts (such as large-area image blur and distortion due to metal implants in the patient's body), low resolution (lower than the minimum resolution standard that can be used for accurate delineation and analysis), incomplete scanning range (not covering all target areas and organs at risk), etc., the patient's data will be excluded;

[0053] Special circumstances: The patient is receiving other treatments that may affect the evaluation of radiotherapy effects (such as other anti-tumor drug treatments or surgeries that may change the tumor morphology or normal tissue status in the recent period), or has special anatomical abnormalities that affect the formulation of radiotherapy plans (such as congenital organ malformations that affect the identification of organs at risk);

[0054] 2. Collect CT images of patients’ internal and external irradiation positioning

[0055] Equipment preparation and calibration:

[0056] Use multi-slice spiral CT scanners that meet medical imaging standards to ensure stable equipment performance. All parameters (such as tube voltage, tube current, scanning layer thickness, pitch, etc.) are calibrated and quality controlled to ensure consistency of image quality;

[0057] Perform regular maintenance and calibration of CT scanners according to the hospital's standard operating procedures, including checking detector performance, X-ray generator output stability, etc. to reduce image artifacts and errors;

[0058] Patient preparation and scanning:

[0059] Patients need to follow the instructions of medical staff to prepare before scanning, such as fasting (specific duration), removing metal objects on the body surface, etc., to avoid artifact interference in the image;

[0060] According to the requirements of the radiotherapy plan, the patient is given an internal and external irradiation positioning CT scan; the scanning range should cover the areas that may be invaded by the cervical cancer tumor and the surrounding organs at risk, such as from the lower edge of the pubic symphysis to above the level of the iliac crest, including the entire pelvic cavity; during the scanning process, ensure that the patient remains still, and use appropriate fixation devices (such as phantoms) to reduce artifacts caused by patient movement;

[0061] 3. Expert delineation of target areas and organs at risk in CT images and collection of dose distribution maps

[0062] Expert team formation:

[0063] Organize an expert team of radiation oncologists and medical physicists who have extensive experience in radiotherapy and are familiar with the anatomical structure and imaging manifestations of cervical cancer to ensure the accuracy and consistency of delineation;

[0064] Train and calibrate expert team members, unify delineation standards and methods through delineation exercises and discussions on known standard cases, and reduce delineation differences between different experts;

[0065] Target volume and organ at risk delineation:

[0066] Target volume delineation: Based on the cervical cancer diagnosis and treatment guidelines and clinical experience, experts accurately delineate the tumor target volume (Gross Tumor Volume, GTV) on the CT image, including the primary tumor and the area it may invade; the clinical target volume (CTV), considering the range of subclinical lesions on the basis of GTV; the planning target volume (PTV), adding the positioning error and the boundary of organ movement on the basis of CTV; during the delineation process, the boundary between the tumor and the surrounding normal tissue must be carefully identified, and a variety of imaging features (such as the density, morphology, and relationship with surrounding tissues of the tumor) must be referred to;

[0067] Organs at risk delineation: Accurately delineate organs at risk that may be affected by radiotherapy, such as the bladder, rectum, sigmoid colon, and femoral head; define each organ at risk according to its anatomical boundaries and functionally important areas, for example, the bladder needs to include the entire bladder wall and intracavitary space, and the rectum is delineated from the rectosigmoid junction to the upper edge of the anorectal ring, etc.;

[0068] Dose distribution map acquisition:

[0069] A professional radiotherapy planning system (TPS) is used to develop a radiotherapy plan based on the delineated target area and organs at risk, combined with the patient's body contour and CT image data. During the radiotherapy plan development process, TPS will calculate the dose distribution of radiation with different fields and different energies in the patient's body, and generate relevant dose information such as the dose-volume histogram (DVH). These dose distribution maps, including the irradiated dose of each target area and organ at risk, the uniformity of dose distribution and other parameters, are collected as an important basis for subsequent model training and evaluation.

[0070] 4. After anonymization, divide the data into training set, validation set and test set

[0071] Data anonymization:

[0072] Remove the patient's personal identification information, such as name, ID number, hospitalization number, etc., and replace it with a unique code to ensure that the patient's privacy is protected; at the same time, process or delete the identifiers in the image data related to the patient's identity (such as hospital name, scanning equipment number, etc., which may reveal the patient's source information);

[0073] Dataset division:

[0074] Random sampling: The anonymized data is divided into training set, validation set and test set by random sampling method. Usually, the data is divided according to a certain ratio, such as 70%-80% for training set, 10%-15% for validation set and 10%-15% for test set. During the sampling process, it is necessary to ensure that each data set has similar distribution in terms of patient characteristics (such as age, pathological type, stage, etc.) and image characteristics (such as image quality, target size and location, etc.) to ensure the effectiveness of model training and evaluation.

[0075] Stratified sampling (optional): If the data has obvious stratified characteristics (such as the number of patients with different pathological types varies greatly), a stratified sampling method can be used; first, the data is divided into different layers according to the stratification factors (such as pathological type), and then samples are proportionally drawn from each layer and allocated to the training set, validation set, and test set. This can better ensure that each data set is representative of data at different levels.

[0076] In this embodiment, the deep learning model construction module includes a segmentation model construction unit and a registration model construction unit;

[0077] The segmentation model building unit collects and preprocesses the internal and external irradiation positioning CT images of cervical cancer patients, screens the data and organizes the expert delineation results; based on the V-Net model, the input layer is built to receive 128×128×64 three-dimensional volume data, and an encoding network including convolution kernels, jump connections and Attention Block modules is constructed, and then a decoding network is constructed using transposed convolution; the model parameters are randomly initialized, and the learning rate, number of iterations and batch size are set; the specific steps include:

[0078] 1. Data collection, preprocessing, screening and sorting

[0079] Data Collection:

[0080] Obtain the patient's internal and external irradiation positioning CT image data processed by the data acquisition module from the hospital's picture archiving and communication system (PACS) or a special cervical cancer radiotherapy database; these data should cover cervical cancer patients of different ages, pathological types, and clinical stages to ensure data diversity;

[0081] Data preprocessing:

[0082] Normalization: Normalize the pixel values ​​of the CT image and map them to a specific range, such as [0,1] or [-1,1]. This can be done by the formula (mapped to [0,1]) or other suitable normalization method, where is the original pixel value, and is the minimum and maximum value of the pixel value in the image; this can make the model training process more stable and avoid adverse effects on model training due to large differences in pixel value ranges;

[0083] Noise removal: Use filtering algorithms (such as Gaussian filtering, median filtering, etc.) to reduce noise in images. For example, Gaussian filtering performs weighted averaging on each pixel in the image and its neighboring pixels, and determines the weights based on the Gaussian function to effectively remove Gaussian noise, making the image smoother while retaining the image's edge and detail information.

[0084] Enhance image contrast: Use histogram equalization or adaptive histogram equalization to improve image contrast; histogram equalization adjusts the grayscale distribution of the image to make the grayscale value of the image more uniform across the entire range, thereby enhancing the visual differences between different tissues in the image and facilitating subsequent feature extraction and model training;

[0085] Data screening and organization:

[0086] Screening according to inclusion criteria: The collected patient data were screened according to the previously established inclusion criteria for cervical cancer patients; data that did not meet the criteria were excluded, such as data with patient age outside the set range, unclear pathological type, or poor image quality that could not accurately delineate the target area and organs at risk;

[0087] Organize expert delineation results: match and organize the target areas (including GTV, CTV, PTV) and contour information of organs at risk drawn by experts on CT images with the corresponding CT images; these delineation results will be used as the gold standard for model training and evaluation to calculate the loss function and evaluate the segmentation accuracy of the model;

[0088] 2. Build a network structure based on the V-Net model

[0089] Build the input layer:

[0090] The input layer is designed to receive 3D volume data of size 128×128×64. This size is determined by comprehensive consideration of computing resources, model complexity, and image information retention. In the input layer, the format and dimensions of the data need to be defined so that subsequent convolution operations can be processed correctly. For example, in deep learning frameworks (such as PyTorch and TensorFlow), corresponding functions can be used to specify the parameters of the input layer to ensure that the shape of the input data meets the requirements.

[0091] Constructing the encoding network

[0092] Convolution operation: Add a three-dimensional convolution kernel for convolution operation; reasonably set the parameters such as convolution kernel size, step size, and padding; for example, use a three-dimensional 5×5×5 convolution kernel with a step size of 1 and a padding of 2 (adjust according to the specific situation); during the convolution process, the convolution kernel performs a dot product operation with the input data in three-dimensional space, and extracts local features through the following formula: ;

[0093] in, represents the output feature map, Represents the coordinate position in the three-dimensional volume data, represents the convolution kernel size, represents the convolution kernel weight, Represents input data; after each layer of convolution, the ReLU function is usually used Activate and introduce nonlinear factors to enable the model to learn more complex image features;

[0094] Skip connection: A skip connection method similar to ResnetBlock is used. This connection method can effectively solve the gradient vanishing problem in deep networks and enable information to propagate more smoothly in the network. Specifically, the input of a certain layer is directly added to the output of a subsequent layer (after appropriate dimension adjustment), so that the network can more easily learn residual information during training and accelerate model convergence.

[0095] Attention Block module: Insert the Attention Block module into the encoding network; and For feature maps of different scales, first use linear transformation is a learnable weight matrix used to calculate Query), is a learnable weight matrix used to calculate Key), is a learnable weight matrix used to calculate Value), and then calculate the attention matrix ( for The dimension of represents the attention matrix, the softmax function is used to normalize the attention matrix), and finally the feature map after attention weighting is obtained ; In this way, the model can pay more attention to the characteristic areas related to the target area and organs at risk, improving the accuracy of segmentation;

[0096] Construct the decoding network:

[0097] Transposed convolution: transposed convolution is used for upsampling operation; transposed convolution increases the size of data by filling zero values ​​in the convolution process, and gradually restores the data resolution; in the transposed convolution process, it is necessary to reasonably set the size, step size and padding parameters of the transposed convolution kernel to achieve the conversion from the low-resolution feature map output by the encoding network to the high-resolution feature map with a size close to that of the original input image; at the same time, the features of the corresponding levels in the encoding network are transferred through jump connections for feature fusion; in this way, the decoding network can make full use of the multi-scale feature information extracted by the encoding network while restoring the resolution, better restore the target area and organ at risk information in the original image, and improve the segmentation effect;

[0098] 3. Model initialization and parameter setting

[0099] Model parameter initialization:

[0100] Random initialization method selection: Randomly initialize the convolution kernel weights, biases and other parameters in the model; you can choose normal distribution initialization, that is, the initial value of the parameter follows the normal distribution (in, is the standard deviation, which can be set according to experience, such as 0.01); it can also be initialized by uniform distribution, and the initial parameter value is randomly generated in a given interval (such as [-0.1, 0.1]); this random initialization method provides a diverse initial state for the model, which is beneficial to the subsequent parameter optimization process;

[0101] Setting hyperparameters:

[0102] Learning rate setting: The learning rate determines the step size of the model parameter update in each iteration; the initial learning rate can be set to a suitable value based on experience or experiments, such as 0.001; during the training process, the learning rate adjustment strategy (such as learning rate decay) can be used to further optimize the model training effect according to the convergence and performance of the model;

[0103] Iteration number setting: The iteration number controls the number of rounds of model training. It needs to be reasonably set according to factors such as the size of the dataset, model complexity, and hardware resources. Generally speaking, you can first try a smaller number of iterations (such as 50) for preliminary training, observe the performance changes of the model on the validation set, and then gradually increase the number of iterations according to the situation, such as adjusting it to 100, 200, etc., until the model performance reaches a stable and satisfactory effect.

[0104] Batch size setting: The batch size determines the number of data samples involved in each training. A smaller batch size (such as 16) may make the model training more stable, but the training speed is slower. A larger batch size (such as 64 or 128) can increase the training speed, but may require more memory resources. It is necessary to balance the training speed and memory usage, and select the appropriate batch size according to the hardware conditions and the characteristics of the dataset.

[0105] Among them, the registration model construction unit inputs two sets of volume data, the target image and the source image, extracts features through the convolution layer, uses the hole convolution to expand the receptive field, and calculates the deformation field based on the objective function including reversible deformation loss, registration similarity measurement loss and deformation field continuity loss. The Adam optimization algorithm is combined with the learning rate decay strategy, the data transformation strategy is used to expand the data volume, and the model is adjusted based on the training set and the validation set. The specific steps include:

[0106] 1. Data input and feature extraction

[0107] Data preparation:

[0108] Acquisition of two sets of volume data, the target image and the source image: The target image and the source image are usually internal and external irradiation positioning CT images of the same cervical cancer patient acquired at different time points, different scanning conditions or different treatment stages; these images have undergone preliminary data acquisition, preprocessing and other steps to ensure that the image quality meets the requirements of subsequent processing, such as noise removal and normalization, and the image resolution, size and other parameters should be mutually adapted to facilitate registration operations;

[0109] Convolutional layer feature extraction:

[0110] Construct convolutional layer: Construct a convolutional layer in the registration model. The parameter settings of the convolutional layer include the convolution kernel size, step size, padding, etc. For example, the convolution kernel size can be set to 3×3×3, the step size to 1, and the padding to 1 (the specific parameters can be adjusted according to the actual situation);

[0111] Convolution calculation: The two sets of volume data of the target image and the source image are input into the convolution layer respectively; the convolution kernel slides on the image data according to the set step size, and extracts features by dot product operation with the image pixel value; for each pixel point in the target image (in, Represents the coordinate position in the three-dimensional volume data), the convolution kernel operates with the pixel value at this point and its neighborhood to generate a new feature value, thereby obtaining the feature map of the target image; similarly, the same operation is performed on the source image to obtain the feature map of the source image; these feature maps contain key feature information such as the edge and texture of the organs in the image, providing a basis for subsequent registration calculations;

[0112] 2. Use dilated convolution to expand the receptive field

[0113] Principle of dilated convolution:

[0114] Dilated Convolution is also called expanded convolution. It expands the receptive field of the convolution kernel by inserting holes between the elements of the convolution kernel (i.e. setting the hole rate) without increasing the actual size of the convolution kernel and the number of parameters of the model. For example, when the hole rate is set to 2, there will be a pixel interval between the adjacent convolution kernel elements, so that the spatial sampling interval of the convolution kernel becomes 2, so that a wider range of contextual information can be captured.

[0115] Applied to feature map:

[0116] The feature maps of the target image and the source image extracted by the ordinary convolution layer are input into the dilated convolution layer for processing; by reasonably setting the dilated convolution rate (such as 2, 3, etc., determined according to actual needs and experiments), the receptive field is further expanded to better estimate the spatial matching position relationship of the pixels between the target image and the source image; this enables the model to take into account the image information farther away when performing registration calculations, thereby improving the accuracy of registration;

[0117] 3. Calculate deformation field based on objective function

[0118] Objective function composition

[0119] Reversible deformation loss: The formula is: ;

[0120] in, Represents the coordinate position of the pixel in the image. Indicates that the source image Align to target image The spatial shape transformation, Indicates that the target graph Align to source image The spatial shape transformation, Represents the calculation norm, which is used to measure the difference in the pixel position before and after the spatial shape transformation; the purpose of this loss function is to ensure that the deformation in the registration process is reversible, that is, the transformations of the forward registration (from the source image to the target image) and the reverse registration (from the target image to the source image) should be able to restore each other. This reversibility is measured by calculating the difference norm of the pixel position before and after the registration and summing them up;

[0121] Registration similarity measure loss: The formula is: ;

[0122] in, represents the target image, represents the source image, Represents the source image After registration transformation After the image, Represents the target image After registration transformation After the image, , , are the edge entropy and joint entropy of the corresponding images respectively. The formulas for calculating the edge probability density and joint probability density are as follows: ;

[0123] in, Representing images The pixel value is The marginal probability density of Representing images The edge entropy of the image The uncertainty or amount of information; ;

[0124] in, Representing images The pixel value is The marginal probability density of Representing images The marginal entropy of ;

[0125] in, Representing images and The joint entropy of

[0126] The deformation field continuity loss formula is: ;

[0127] in, represents the gradient operator, Indicates For the target image, is the deformation field estimated by the registration of the source image, Indicates For the target image, is the deformation field estimated by the registration of the source image, and Denote the deformation field and At coordinate position The smaller the gradient norm, the more continuous the deformation. This loss function is to ensure that the deformation field generated during the registration process is continuous and smooth, avoiding mutations or discontinuities. This continuity is measured by calculating the gradient norm of the deformation field at each pixel and summing them up.

[0128] Compute the deformation field:

[0129] Based on the above objective function including reversible deformation loss, registration similarity measure loss and deformation field continuity loss, the model is trained through optimization algorithms (such as gradient descent algorithm and its variants), and the parameters of the model (such as convolution kernel weights, biases, etc.) are continuously adjusted to minimize the value of the objective function; during the training process, the model will calculate an optimal deformation field based on the input target image and source image feature map; the deformation field is a vector field that describes how each pixel in the source image needs to be displaced relative to the corresponding point in the target image. For example, for a pixel in the source image , the deformation field gives a displacement vector , indicating that the pixel needs to be Move in three directions units to match the corresponding points in the target graph;

[0130] 4. Adopt Adam optimization algorithm combined with learning rate decay strategy

[0131] Adam optimization algorithm principle:

[0132] The Adam optimization algorithm combines the concepts of adaptive learning rate and momentum to update model parameters more effectively during training. It maintains two estimators: the first-order moment estimate and the second moment estimate , to automatically adjust the learning rate;

[0133] The parameter update formula is as follows: ;

[0134] in, Indicates The first-order moment estimate of the iteration, is the attenuation rate, Indicates The first-order moment estimate of the iteration, Indicates The gradient of the iteration, here It is a weighted average of the gradients. As the iterations proceed, it accumulates gradient information and The value of will affect the degree to which it forgets historical gradient information; ;

[0135] in, Indicates The second-order moment estimate of the iteration, is the attenuation rate, Indicates The second-order moment estimate of the iteration, Indicates The square of the gradient of the iteration, It is a weighted average of the squared gradient, which is mainly used to measure the magnitude of the gradient change and helps adjust the step size of the model parameter update together with the first-order moment estimate; ;

[0136] in, To correct the result of the first-order moment estimate, the first-order moment estimate is corrected so that it can reflect the true situation of the gradient more quickly in the early stage of iteration, and the influence of the correction gradually decreases as the number of iterations increases; ;

[0137] The second-order moment estimate is corrected, where is the result after correcting the second-order moment estimate; ;

[0138] in, represents the model parameters of the tth iteration, Indicates The model parameters of the iteration, is the learning rate, which determines the step size of parameter update. is a very small constant to prevent the denominator from being zero. Through this formula, the model updates the parameters according to the current parameters, learning rate, corrected first-order moment estimation and second-order moment estimation, thus achieving effective update of the model parameters.

[0139] Learning rate decay strategy application:

[0140] The learning rate decay strategy is a method of gradually reducing the learning rate as training progresses; for example, an exponential decay method can be used, that is, (in, It is The learning rate of the iteration, is the initial learning rate, is the attenuation factor , is the decay step size (indicates every The learning rate decay strategy is used in combination with the Adam optimization algorithm. This allows the model to use a larger learning rate to quickly learn the general characteristics of the data in the early stage of training, and use a smaller learning rate to fine-tune the model parameters in the later stage of training. This avoids problems such as unstable parameter updates or overfitting caused by excessive learning rates, thereby improving the performance and convergence of the model.

[0141] 5. Use data transformation strategies to expand data volume

[0142] Data transformation strategy introduction:

[0143] Data transformation strategies include image horizontal flipping, random extraction, scale transformation, and rotation operations;

[0144] Image horizontal flipping: Flip the image along the vertical axis, swapping the left and right positions. For example, for a CT image of cervical cancer, the flipped image still semantically represents the same anatomical structure, but the appearance of the image has changed. This transformation can increase the diversity of the data and allow the model to learn features such as the left-right symmetry of the image.

[0145] Random extraction: randomly select a sub-region from the original image as a new data sample; for example, in a larger cervical cancer CT image, randomly extract a local area, which allows the model to learn the local features of the image. Since the location and size of each extraction may be different, the variability of the data is increased, avoiding the model's over-reliance on the global features of the complete image, which helps improve the model's ability to recognize targets of different sizes and locations.

[0146] Scale transformation: Enlarge or reduce an image; for example, enlarging or reducing a cervical cancer CT image by a certain ratio can simulate images of different resolutions, allowing the model to adapt to images under different scanning devices or scanning parameters. It also helps the model learn the characteristics of target objects (such as target areas and organs at risk) at different scales, because in actual clinical applications, the scale of CT images may vary due to various factors;

[0147] Rotation: Rotate the image around a certain center point; for example, rotate a cervical cancer CT image around the center of the image by a certain angle. The rotated image can also increase the diversity of the data, allowing the model to recognize the shape of the target object at different angles. For cervical cancer radiotherapy images, organs and target areas may be presented differently at different scanning angles. By rotating the data, the model can better adapt to these changes.

[0148] Expand data volume operation:

[0149] The above data transformation strategies are applied to the two sets of input volume data, the target image and the source image. For example, for each target image and the source image, there is a certain probability (such as 50%) of performing a horizontal flip operation, a certain probability (such as 30%) of performing a random extraction operation, and a certain probability (such as 20%) of performing a scale transformation or rotation operation (the specific probability can be adjusted according to the actual situation). Through these transformation operations, new image data is generated, thereby expanding the amount of original data, increasing the diversity of data, and helping to improve the generalization ability of the model, so that it can better adapt to cervical cancer radiotherapy image data under different circumstances.

[0150] 6. Adjust the model based on the training set and validation set

[0151] Model training process:

[0152] Use the training set to train the registration model; divide the training data (including the original data and the data generated after the transformation) after data transformation and expansion into several batches according to the set batch size, take a batch of data each time and input it into the registration model for forward propagation to calculate the prediction result, and then calculate the loss function value (based on the objective function composed of the above-mentioned reversible deformation loss, registration similarity measure loss and deformation field continuity loss) according to the predicted result and the true label (such as the manually labeled registration result or the known accurate registration data); according to the calculated loss function value, update the model parameters through the Adam optimization algorithm combined with the learning rate decay strategy, and repeat this process until the set number of iterations is reached or the model performance reaches a certain stability standard;

[0153] Model validation and tuning:

[0154] During the training process, the validation set is used to validate the model regularly. The validation set data is input into the model being trained, and the performance indicators of the model on the validation set are calculated, such as the registration error (which can be measured by calculating a certain distance metric between the source image and the target image after registration, such as mean square error, Hausdorff distance, etc.), the registration similarity index (such as calculating the marginal entropy and the joint entropy related index to measure the similarity after registration), etc. According to the performance indicator results of the validation set, the advantages and disadvantages of the model are analyzed. If problems are found in some aspects of the model, such as excessive registration error or unsatisfactory registration similarity index, the model can be adjusted. The adjustment methods include but are not limited to: adjusting the hyperparameters of the model (such as learning rate, number of iterations, etc.), optimizing the network structure of the model (such as increasing or decreasing network layers, adjusting the size of convolution kernels, etc.), and further optimizing the data transformation strategy (such as changing the probability of data transformation, increasing or decreasing certain transformation operations, etc.). Through these adjustment measures, the performance of the model is continuously improved so that it can better adapt to the needs of cervical cancer radiotherapy image registration in practical applications.

[0155] In this embodiment, the model training and optimization module uses the training set to train the segmentation model and the registration model, adopts the data transformation strategy to expand the amount of training data, updates the model parameters through the optimization algorithm, and adjusts the model structure and training strategy according to the evaluation results of the validation set.

[0156] In this embodiment, the trained segmentation model is used for automatic delineation, the accuracy of delineation is evaluated by calculating the DSC and 95HD indicators, the registration model is used to register the images to obtain the spatial transformation relationship, and the spatial transformation relationship is applied to the dose map to perform a unified coordinate system and cumulative calculation of the internal and external irradiation doses, and finally the cumulative total dose map is obtained, which specifically includes the following steps:

[0157] 1. Automatically outline using the trained segmentation model

[0158] Data preparation and preprocessing:

[0159] Obtain internal and external irradiation CT image data of cervical cancer patients from the hospital image storage system; these image data should undergo the same preprocessing process as the training data, including but not limited to:

[0160] Normalization: Normalize the image pixel values ​​to a specific range, such as [0,1] or [-1,1]; for example, for the grayscale value of each pixel, you can use the formula The normalization method is performed, where is the original pixel value, and is the minimum and maximum pixel values ​​in the image; here Represents the normalized pixel value, which makes the pixel value range of different images uniform and easier to process by the model;

[0161] Image resizing: Ensure that the image size meets the input requirements of the segmentation model. If the model input requires 128×128×64 3D volume data, images that do not meet this size can be resized using methods such as interpolation.

[0162] Noise removal and contrast enhancement: Use filtering algorithms (such as Gaussian filtering) to remove noise in the image. At the same time, use methods such as histogram equalization to enhance image contrast, making structures such as target areas and organs at risk in the image clearer;

[0163] Model loading and settings:

[0164] Load the pre-trained segmentation model into the computing environment; this includes loading the model's network structure parameters (such as convolution kernel weights, biases, etc.), ensuring that the model is in evaluation mode (for example, setting model.eval() in a deep learning framework) to fix the model parameters and turn off some operations used during training (such as random dropout, etc.);

[0165] Automatic outlining process:

[0166] The preprocessed CT image data is input into the segmentation model for forward propagation calculation; the model will process according to its trained encoder-decoder network structure:

[0167] Coding network operation: In the coding network, a convolution operation is performed through a three-dimensional convolution kernel (such as a 5×5×5 convolution kernel) to extract local features of the image according to the convolution formula: ;

[0168] in, represents the output feature map, Represents the coordinate position in the three-dimensional volume data, represents the convolution kernel size, represents the convolution kernel weight, Represents input data;

[0169] here Indicated in coordinates The eigenvalues ​​output at Determines the number of pixels covered by the convolution kernel in each dimension. is the weight value of the convolution kernel at different positions, which is compared with the input data The dot product operation is used to extract features. The ReLU activation function is used after each convolution layer to introduce nonlinear factors. At the same time, through the jump connection and Attention Block module similar to ResnetBlock, the features of different scales are integrated and the attention weights are assigned, so that the model can focus on the key areas related to the target area and the organs at risk.

[0170] Decoding network operation: In the decoding network, upsampling is performed using transposed convolution or interpolation methods to restore data resolution. The features of the corresponding layers of the encoding network are transferred through jump connections for fusion, making full use of multi-scale information, and finally outputting a segmentation mask of the same size as the input image. This mask marks the category information of each pixel point, whether it belongs to the target area, organ at risk, or background, completing the automatic delineation process.

[0171] 2. Evaluate the accuracy of delineation by calculating DSC and 95HD indicators

[0172] Get expert and automatic delineation results:

[0173] Obtain the expert delineation results corresponding to the patient's CT image from the database or storage system; the expert delineation results are the contours of the target area (such as GTV, CTV, PTV) and organs at risk carefully drawn by experienced radiotherapists, which serve as the gold standard for evaluating the accuracy of automatic delineation;

[0174] The automatic delineation result is the segmentation mask just obtained by the segmentation model, which contains the predicted delineation information of the target area and organs at risk;

[0175] Calculate the DSC index:

[0176] Determine TP, FP, and FN values: TruePositive (TP): By comparing the automatic delineation results with the expert delineation results, the number of pixels that actually belong to the target area and are correctly classified as pixels in the target area by the automatic delineation algorithm is counted; this can be determined by traversing each pixel in the image and judging whether it is marked as a target area in both the automatic delineation and the expert delineation; here TP stands for true positive, that is, the number of samples correctly predicted by the model as positive (target area); False Positive (FP): counts the number of pixels that are not actually in the target area but are mistakenly classified as pixels in the target area by the automatic delineation algorithm; that is, the pixels are marked as the target area in the automatic delineation, but do not belong to the target area in the expert delineation; FP stands for false positive, that is, the number of samples where the model mistakenly predicts the negative class (non-target area) as the positive class; False Negative (FN): counts the number of pixels that actually belong to the target area but are mistakenly classified as non-target pixels by the automatic delineation algorithm; that is, the pixels are not marked as target areas in the automatic delineation, but belong to the target area in the expert delineation; FN stands for false negative examples, that is, the number of samples where the model mistakenly predicts the positive class as the negative class;

[0177] Calculate DSC value: Substitute TP, FP, and FN values ​​into the DSC calculation formula: ;

[0178] Get the DSC value; the range of DSC value is 0-1, the closer to 1, the more similar the automatic delineation result is to the expert delineation result, and the higher the accuracy; in this formula, the numerator Emphasizes the importance of true examples, the denominator The similarity between the automatic outline and the expert outline is comprehensively measured by taking into account the false positives, false negatives and true positives.

[0179] Calculate the 95HD indicator:

[0180] Extract edge point sets: Extract edge point sets of the target area and organs at risk from the automatic delineation results and expert delineation results respectively; edge points can be obtained through edge detection algorithms (such as Canny edge detection algorithm) to obtain the automatically delineated edge point sets and the edge point set drawn by the expert ;here and There are two point sets, in which the points represent the edge positions of the target area or organs at risk in the automatic delineation and expert delineation results;

[0181] Calculate the distance matrix: For the edge point set Each point in , calculate its edge point set The distance between all points (Here we use the L2 paradigm distance, that is, the Euclidean distance ,in, and are the three-dimensional coordinates of the point and the point respectively), and find each point arrive The distance between the nearest points in the matrix ; In distance calculation, Indicate point and Point The Euclidean distance between Represents the point set Each point-to-point set in The matrix composed of the distances of the nearest points in reflects the distance relationship between the two point sets;

[0182] Get 95HD value: Sort all distance values ​​in the distance matrix and take the 95th percentile as the 95HD value. The smaller the 95HD value, the smaller the difference between the automatic delineation result and the expert delineation result in the edge part, and the higher the accuracy. The 95th percentile represents the value at the 95th position after all distance values ​​are sorted from small to large. It can be used as a statistic to measure the degree of difference between the edges of two point sets.

[0183] 3. Use the registration model to register images to obtain spatial transformation relationships

[0184] Data preparation and preprocessing:

[0185] Prepare two sets of CT image data of the same cervical cancer patient, as the target image and the source image respectively; the two sets of images may be acquired at different times, with different scanning devices or with different scanning parameters; similarly, preprocess the two sets of images, including normalization, resizing (if necessary), and noise removal, so that they meet the input requirements of the registration model;

[0186] Model loading and settings:

[0187] Load the pre-trained registration model into the computing environment, set the model to evaluation mode (such as model.eval()), fix the model parameters, and prepare for image registration calculation;

[0188] Registration calculation process:

[0189] The target image and the source image are input into the registration model, and the model performs the registration calculation through the following steps:

[0190] Feature extraction: The convolution layer is used to extract features from the target image and the source image. The convolution kernel slides on the image and performs dot product operations with the pixel values ​​to generate a feature map containing key features such as organ edges and textures. Here, the convolution layer extracts features through the interaction between the convolution kernel and the image pixels, providing basic information for subsequent registration calculations.

[0191] Atrous convolution processing: Atrous convolution is used to expand the receptive field. According to the set atrous rate (such as 2 or 3), the convolution kernel can capture a wider range of contextual information and better estimate the spatial matching relationship between the pixels of the target image and the source image. Atrous convolution increases the receptive field by inserting holes between the convolution kernel elements without increasing the convolution kernel size and the number of model parameters. For example, when the atrous rate is 2, the spatial sampling interval of the convolution kernel becomes 2, so that image information farther away can be obtained.

[0192] Calculate deformation field based on objective function: The registration model is calculated based on an objective function that includes reversible deformation loss, registration similarity measure loss, and deformation field continuity loss;

[0193] Calculation of reversible deformation loss: Calculate the reversible deformation loss according to the formula: ;

[0194] in, Represents the coordinate position of the pixel in the image. Indicates that the source image Align to target image The spatial shape transformation, Indicates that the target graph Align to source image The spatial shape transformation, represents the calculation norm, which is used to measure the difference in pixel position before and after spatial shape transformation to ensure that the deformation during the registration process is reversible; in this formula, Used to traverse each pixel in the image. Indicates that the coordinates in the source image are The pixel points are transformed to the position in the target image through registration. Calculate the difference norm of the position of this pixel before and after forward registration. Similarly Calculate the position difference norm of the reverse registration, and add the two to form the reversible deformation loss, which is used to measure the reversibility of the registration;

[0195] Registration similarity measure loss calculation: Calculate the target map And the source image after registration transformation The registration similarity measure loss between : ;

[0196] in, represents the target image, represents the source image, Represents the source image After registration transformation After the image, Represents the target image After registration transformation After the image, , , are the edge entropy and joint entropy of the corresponding images respectively. The formulas for calculating the edge probability density and joint probability density are as follows: ;

[0197] in, Representing images The pixel value is The marginal probability density of Representing images The edge entropy of the image The uncertainty or amount of information; ;

[0198] in, Representing images The pixel value is The marginal probability density of Representing images The marginal entropy of ;

[0199] in, Representing images and The joint entropy of

[0200] Calculation of deformation field continuity loss: According to the formula: ;

[0201] in, represents the gradient operator, Indicates For the target image, is the deformation field estimated by the registration of the source image, Indicates For the target image, is the deformation field estimated by the registration of the source image, and Denote the deformation field and At coordinate position The gradient norm at , the smaller the gradient norm, the more continuous the deformation;

[0202] The above objective function is minimized during the training process through an optimization algorithm (such as the Adam optimization algorithm) to obtain the optimal deformation field; this deformation field represents the displacement transformation that needs to be performed on each pixel point in the source image relative to the corresponding point in the target image, that is, the spatial transformation relationship between the target image and the source image is obtained;

[0203] 4. Apply the spatial transformation relationship to the dose map to unify the coordinate system and cumulative calculation of internal and external radiation doses, and finally obtain the cumulative total dose map

[0204] Dose map acquisition and preprocessing:

[0205] The internal and external radiation dose maps of the patient are obtained from the radiotherapy planning system; the internal and external radiation dose maps are calculated based on the patient's CT images, the target area and the results of the outline of the organs at risk, and the radiotherapy planning parameters, and respectively represent the dose distribution of internal radiation and external radiation at various locations in the patient's body;

[0206] Preprocess the dose map, such as normalizing the dose value to an appropriate range, ensuring that the spatial coordinates of the dose map are consistent with those of the CT image (if they are inconsistent, coordinate conversion is required), and that the size and spatial resolution of the dose map match those of the CT image to facilitate subsequent spatial transformation and accumulation calculation;

[0207] Apply spatial transformation relationships:

[0208] According to the spatial transformation relationship obtained by the registration model, the dose map corresponding to the source image (assuming that the target image is an external irradiation dose map and the source image is an internal irradiation dose map) is spatially transformed; the specific operation is that for each pixel in the source image dose map, according to the displacement vector in the spatial transformation relationship (such as The coordinates of a pixel are adjusted by the displacement in the direction. For example, if the original coordinates of a pixel are , the spatial transformation relationship indicates Direction Move the unit, Direction displacement units, in Direction displacement units, then the transformed coordinates are ; In this way, the internal radiation dose map and the external radiation dose map are aligned in the same coordinate system; Here is the original coordinate of the pixel in the source dose map, It is the displacement vector in the spatial transformation relationship, and the alignment of the dose map is achieved through coordinate transformation;

[0209] Cumulative calculation:

[0210] For the internal and external irradiation dose maps aligned in the same coordinate system, the dose accumulation calculation is performed at each pixel position; the dose values ​​of the internal irradiation dose map and the external irradiation dose map at the same position are added to obtain the cumulative dose value at that position; for example, for the coordinates The external irradiation dose value is The internal radiation dose is , the cumulative dose value is ;here and Respectively represent external and internal irradiation in the coordinate The dose value at is the accumulated dose value;

[0211] The above cumulative calculation is performed on all pixel points in the patient's body, and a cumulative total dose map is finally obtained; this cumulative total dose map shows the total dose received by various areas of the patient's body after comprehensive internal and external radiation treatment. Doctors can evaluate the effectiveness and safety of radiotherapy plans based on the cumulative total dose map, such as whether the tumor has received sufficient radiation dose and whether the endangered organs have been over-irradiated.

[0212] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The deep learning-based automatic delineation system for cervical cancer target area is characterized by: include: Data acquisition module; the data acquisition module determines the inclusion and exclusion criteria of cervical cancer patients, collects internal and external irradiation positioning CT images of patients, performs expert delineation of target areas and organs at risk in CT images and collects dose distribution maps, and divides the data into training set, validation set and test set after anonymization processing; Deep learning model building blocks; The deep learning model building module includes: A segmentation model building unit; the segmentation model building unit collects and preprocesses internal and external irradiation positioning CT images of cervical cancer patients, screens data and organizes expert delineation results; based on the V-Net model, an input layer is built to receive 128×128×64 three-dimensional volume data, an encoding network including convolution kernels, jump connections and Attention Block modules is constructed, and a decoding network is constructed using transposed convolution; model parameters are randomly initialized, and a learning rate, number of iterations and batch size are set; A registration model building unit; the registration model building unit inputs two sets of volume data, namely, a target image and a source image, extracts features through a convolutional layer, uses a dilated convolution to expand the receptive field, calculates the deformation field based on an objective function including a reversible deformation loss, a registration similarity measurement loss, and a deformation field continuity loss, uses an Adam optimization algorithm combined with a learning rate decay strategy, uses a data transformation strategy to expand the data volume, and adjusts the model based on a training set and a validation set; Model training and optimization module: the model training and optimization module uses the training set to train the segmentation model and the registration model, adopts the data transformation strategy to expand the amount of training data, updates the model parameters through the optimization algorithm, and adjusts the model structure and training strategy according to the evaluation results of the validation set; Automatic outlining and dose calculation module: Use the trained segmentation model for automatic outlining, evaluate the accuracy of outlining by calculating the DSC and 95HD indicators, use the registration model to register the images to obtain the spatial transformation relationship, apply the spatial transformation relationship to the dose map to perform a unified coordinate system and cumulative calculation of internal and external irradiation doses, and finally obtain the cumulative total dose map.

2. The deep learning-based automatic delineation system for cervical cancer target area according to claim 1, characterized in that: In the segmentation model construction unit, the encoding network convolution operation formula is: ； in, represents the output feature map, Represents the coordinate position in the three-dimensional volume data, represents the convolution kernel size, represents the convolution kernel weight, Represents input data; After the convolution operation through the encoding network, it is activated by the ReLU function; The Attention Block calculation method is: set up and are feature maps of different scales; ,in, is a learnable weight matrix used to calculate Query; ,in, is a learnable weight matrix used to calculate Key; ,in, is a learnable weight matrix used to calculate Value; ,in, for The dimension of Represents the attention matrix, and the softmax function is used to normalize the attention matrix; Then the attention-weighted feature map is obtained by the following formula : 。 3. The deep learning-based automatic delineation system for cervical cancer target area according to claim 2, characterized in that: In the registration model construction unit, the reversible deformation loss formula is: ； in, Represents the coordinate position of the pixel in the image. Indicates that the source image Align to target image The spatial shape transformation, Indicates that the target graph Align to source image The spatial shape transformation, Represents the calculation norm, which is used to measure the difference in pixel position before and after spatial shape transformation; The registration similarity measure loss formula is: ； in, represents the target image, represents the source image, Represents the source image After registration transformation After the image, Represents the target image After registration transformation After the image, , , are the edge entropy and joint entropy of the corresponding images respectively. The formulas for calculating the edge probability density and joint probability density are as follows: ； in, Representing images The pixel value is The marginal probability density of Representing images The edge entropy of the image The uncertainty or amount of information; ； in, Representing images The pixel value is The marginal probability density of Representing images The marginal entropy of ； in, Representing images and The joint entropy of The deformation field continuity loss formula is: ； in, represents the gradient operator, Indicates For the target image, is the deformation field estimated by the registration of the source image, Indicates For the target image, is the deformation field estimated by the registration of the source image, and Denote the deformation field and At coordinate position The smaller the gradient norm, the more continuous the deformation.

4. The deep learning-based automatic delineation system for cervical cancer target area according to claim 3, characterized in that: In the model training and optimization module, the data transformation strategies include horizontal image flipping, random extraction, scale transformation and rotation; the parameter update formula of the Adam optimization algorithm is: ;in, Indicates The first-order moment estimate of the iteration, is the attenuation rate, Indicates The first-order moment estimate of the iteration, Indicates The gradient of the iteration; ;in, Indicates The second-order moment estimate of the iteration, is the attenuation rate, Indicates The second-order moment estimate of the iteration, Indicates The square of the gradient of the iteration; Using the formula The first-order moment estimate is corrected, where is the result after correcting the first-order moment estimate; Using the formula The second-order moment estimate is corrected, where is the result after correcting the second-order moment estimate; ;in, represents the model parameters of the tth iteration, Indicates The model parameters of the iteration, is the learning rate, which determines the step size of parameter update. is a constant.

5. The deep learning-based automatic delineation system for cervical cancer target area according to claim 4, characterized in that: In the automatic delineation and dose calculation module, the DSC calculation formula is: ； Among them, TP represents the number of pixels correctly classified as positive examples, that is, the number of pixels that actually belong to the target area and are classified as pixels in the target area by the algorithm; FP represents the number of pixels incorrectly classified as positive examples, that is, the number of pixels that do not actually belong to the target area but are classified as pixels in the target area by the algorithm; FN represents the number of pixels incorrectly classified as negative examples, that is, the number of pixels that actually belong to the target area but are classified as pixels outside the target area by the algorithm. DSC is used to measure the similarity between the automatic delineation result and the expert delineation gold standard. The closer the value is to 1, the higher the delineation accuracy. 95HD calculation first calculates the distance matrix of the predicted result and the actual result edge point set using the following formula: ； in, Represents the edge point set of the prediction result, represents the real result edge point set, Indicate point and The distance between points, here we use L2 norm distance or Euclidean distance, Represents the point set Each point-to-point set in The matrix composed of the distances to the nearest points in ; The 95th percentile is taken as the 95HD value.

Images