Cervical cancer internal and external irradiation dose superposition system based on deep learning

Abstract

The invention relates to the technical field of medical image processing and radiotherapy, in particular to a cervical cancer internal and external irradiation dose superposition system based on deep learning, which comprises a data input module, a deformation registration module, a dose calculation module, a dose superposition module and a result analysis module. The data input module verifies and preprocesses internal and external irradiation CT image data; the deformation registration module adopts a V-Net framework to be combined with an algorithm based on Transform to accurately align images; the dose calculation module calculates dose distribution by using a convolutional neural network; the dose superposition module performs weighted superposition on the internal and external irradiation doses to generate a comprehensive dose volume histogram; and the result analysis module evaluates the dose conditions of the target region and the normal tissue by using statistical and visual tools. According to the method, the accuracy of dose-related operation is improved, the working efficiency is improved, personalized precise radiotherapy is assisted, and the method has positive significance in promoting development and clinical application of the radiotherapy technology.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing and radiotherapy, and specifically to a cervical cancer internal and external irradiation dose superposition system based on deep learning. Background Art

[0002] Cervical cancer is one of the most common malignant tumors in women worldwide, and radiotherapy plays a vital role in its comprehensive treatment. At present, a combination of internal and external radiation therapy is often used in clinical practice, aiming to maximize the radiation dose to the tumor target area while minimizing damage to surrounding normal tissues. However, in the actual radiotherapy process, there are many challenges in accurately evaluating and optimizing the superposition effect of internal and external radiation doses.

[0003] On the one hand, CT images acquired under different irradiation modes have differences in spatial position and morphology. When traditional image registration methods perform spatial alignment on internal and external irradiation CT images, it is often difficult to accurately capture the complex features and subtle changes in the images, which is prone to registration errors, thus affecting the accuracy of subsequent dose calculation and superposition. For example, the registration algorithm based on manual features and simple transformation models has limited processing capabilities for image differences caused by factors such as changes in patient position and deformation of tissues and organs.

[0004] On the other hand, dose calculation relies on complex physical models and empirical formulas, and needs to consider a variety of factors related to the interaction between radiation and human tissue, such as tissue density, atomic composition, and radiation energy and angle. However, the diversity and individual differences of human tissue make it difficult for traditional dose calculation methods to accurately reflect the actual dose distribution, resulting in a certain deviation between the calculated results and the actual situation, which is not conducive to the formulation of accurate radiotherapy plans.

[0005] Limitations of existing technologies:

[0006] Traditional image registration technologies are mostly based on rigid transformation or simple elastic deformation models, lacking in-depth understanding of image semantic information and unable to effectively deal with the complex anatomical changes and tissue deformation in CT images of cervical cancer with internal and external irradiation. At the same time, manual feature extraction is time-consuming and laborious, and the feature expression ability is limited, making it difficult to meet the needs of high-precision registration, which in turn affects the accurate superposition of images under different irradiation methods and the subsequent dose analysis based on accurate images.

[0007] Existing dose calculation methods are mainly based on physical methods such as analytical algorithms and Monte Carlo simulation. Although these methods have a certain theoretical basis, they require a lot of computing resources and time costs when facing complex human tissue structures and individual differences, and their accuracy still cannot reach the ideal level in some cases. Moreover, for dose analysis, traditional methods mainly rely on a limited number of dosimetric indicators and lack comprehensive and intuitive analysis methods. It is difficult to provide clinicians with sufficiently detailed and easy-to-understand dose distribution information, which is not conducive to timely and accurate adjustment of radiotherapy plans.

[0008] Therefore, to address the above problems, a deep learning-based internal and external irradiation dose superposition system for cervical cancer was proposed. Summary of the invention

[0009] The purpose of the present invention is to provide a cervical cancer internal and external irradiation dose superposition system based on deep learning to solve the problems raised in the above background technology.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] The deep learning-based cervical cancer internal and external irradiation dose superposition system includes:

[0012] Data input module: The data input module receives the internal and external irradiation CT image data, verifies the integrity and accuracy of the internal and external irradiation CT image data, and performs denoising and standardization preprocessing operations;

[0013] Deformable registration module: The deformable registration module includes a deep learning-based deformable registration algorithm. The deformable registration algorithm uses a V-Net architecture combined with a Transformer-based algorithm to spatially align the internal and external irradiation CT image data output by the data input module, achieve the superposition of images under different irradiation modes, and obtain the registered image;

[0014] Dose calculation module: The dose calculation module uses a deep learning model to calculate the dose distribution. The deep learning model is a convolutional neural network, which is used to process the registered image output by the variable registration module and output the distribution results of the internal and external irradiation doses.

[0015] Dose superposition module: The dose superposition module performs weighted superposition on the internal and external irradiation doses output by the dose calculation module to generate a comprehensive dose volume histogram;

[0016] Result analysis module: The result analysis module uses statistical analysis methods and visualization tools to analyze the comprehensive dose volume histogram and dose data generated by the dose superposition module to evaluate the dose distribution of the target area and normal tissue.

[0017] As a preferred solution, in the data input module, the algorithm used for the denoising operation is the Wiener filter algorithm, and the calculation formula of the Wiener filter algorithm is: ;

[0018] in, Indicates the coordinates of the image after denoising The pixel value at is the point spread function that causes image blur, Its complex conjugate, and represent the noise power spectral density and the power spectral density of the original image signal respectively, is the original image with noise at coordinates The pixel value at , that is, the input image function to be denoised;

[0019] The method used for standardization is Z-score standardization, and its calculation formula is: ;in, Represents the corresponding data value after standardization, is the original data value, is the mean of the original data, is the standard deviation of the original data.

[0020] As a preferred solution, the working mechanism of the deformation registration module is as follows:

[0021] The V-Net architecture consists of multiple sequentially connected convolutional layers, downsampling layers, upsampling layers, and skip connection structures, which are used to extract and fuse image features and restore image spatial resolution;

[0022] The convolutional layer assumes that the input feature map is , the output feature map is , the convolution kernel size is , the step length is , the convolution kernel weight is , biased to , a single convolutional layer is calculated according to the following formula: ;

[0023] in, Output feature map Pixel coordinate position Represents the input feature map Corresponding position pixel value;

[0024] In the downsampling layer, the pooling window size and step size are set to , the downsampling layer calculation formula is: ;

[0025] in, The output feature map after downsampling is at coordinate The pixel value at , that is, from the input feature map Take the maximum value within the corresponding pooling window as output;

[0026] In the upsampling layer, the transposed convolution kernel size is , the step length is , the upsampling layer calculation formula is: ;

[0027] in, , are the weights and biases of the transposed convolution kernel respectively;

[0028] The jump connection fuses the feature maps at different levels. The low-resolution feature map is , the high-resolution feature map is , the fused feature map The calculation formula is: ;

[0029] In the Transformer-based algorithm, for the input image feature sequence , represents the sequence length, Represents the feature dimension, and generates query, key, and value through linear transformation respectively. , the calculation formula is as follows: ;in, , , is the learnable weight matrix, and its dimensions are ,in, ;

[0030] The formula for calculating the attention score is: ;

[0031] Among them, the softmax function is used to normalize the attention score to the probability distribution form. yes The transpose of

[0032] The image features extracted by V-Net are input into the Transformer-based module. After the self-attention mechanism is used to capture the relationship between features, they are then sent back to the subsequent layers of V-Net to continue feature fusion and upsampling operations, and finally the aligned image is output.

[0033] As a preferred solution, the training mechanism of the deformable registration module is as follows:

[0034] The network parameters were optimized and adjusted by training on a training data set containing a large number of cervical cancer cases with internal and external irradiation CT images and their standard registration results. The loss function used in the training process was the mean square error loss function, and the calculation formula was: ;

[0035] in, is the number of training samples, Indicates The true label value of the sample, Represents the network prediction The output value of each sample;

[0036] The optimization algorithm is the stochastic gradient descent algorithm, and its basic formula for updating parameters is: ;

[0037] in, Indicated in The model parameters at the iteration, Indicates The updated model parameters at the iteration, is the learning rate, which controls the step size of each parameter update. Represents the loss function About parameters The gradient of , indicating the direction of parameter update.

[0038] As a preferred solution, in the dose calculation module, the convolutional neural network uses the dose distribution data of cervical cancer cases verified by physical dose calculation as the reference standard for supervised learning during training. The network's hyperparameters include the convolution kernel size k, step size s, and number of layers l. The hyperparameters are adjusted and optimized based on the training results and the accuracy of the validation set.

[0039] As a preferred solution, in the dose superposition module, the weighted coefficient of the dose superposition algorithm is set based on clinical experience and the dosimetric characteristics of internal and external irradiation of cervical cancer, and is used to allocate the weights of internal and external irradiation doses in the process of generating the comprehensive dose volume histogram. Assuming that the internal irradiation dose distribution function is , the external radiation dose distribution function is , set the internal illumination weight to , the external radiation weight is ,and The comprehensive dose distribution The calculation formula is: ;

[0040] in, Represents the spatial coordinate position based on the comprehensive dose distribution To generate a comprehensive dose-volume histogram.

[0041] As a preferred solution, in the result analysis module, the statistical analysis method at least includes the statistical operations of calculating the dose mean, maximum, minimum and dose coverage volume. The visualization tool can generate dose cloud maps, dose curves and comprehensive dose volume histogram comparison diagrams that intuitively display the dose distribution. The specific calculation and display methods are as follows:

[0042] Assume that the dose distribution function is In the area of ​​interest Mean dose The calculation formula is: ;

[0043] in, represents a volume differential element, and the integral is over the entire region of interest The numerator represents the total volume integral of the dose in the area, and the denominator represents the area The total volume of the sample is divided by the average dose.

[0044] Given a dose distribution function Within the covered space, by traversing all coordinate positions Corresponding dose values, find the maximum value and minimum value ,Right now: ; ;

[0045] in, Indicates the entire spatial range involved in the dose distribution;

[0046] For a given dose threshold , calculate the volume whose dose is greater than or equal to the threshold , by determining the dose distribution function Whether the coordinates are satisfied at each position ,Then count the space volume formed by the coordinate positions that meet the conditions, and the calculation formula is expressed as: ;

[0047] Among them, the integral range is to meet the dose greater than or equal to the threshold The spatial region composed of all coordinate positions of .

[0048] It can be seen from the technical solutions provided by the present invention that the deep learning-based cervical cancer internal and external irradiation dose superposition system provided by the present invention has the following beneficial effects:

[0049] 1. The present invention adopts V-Net architecture combined with Transformer-based algorithm for deformation registration. V-Net can effectively extract different levels of features in internal and external irradiation CT images through its multi-layer convolution, downsampling, upsampling and jump connection structures, and can accurately capture everything from local image texture to overall tissue and organ structure features. The self-attention mechanism of Transformer can further capture the long-distance dependencies between image features, overcoming the limitations of traditional convolutional neural networks in understanding global image information, making the spatial alignment of CT images under different irradiation modes more accurate, reducing the dose calculation and analysis errors caused by inaccurate registration, and thus laying a solid foundation for subsequent accurate dose superposition and evaluation;

[0050] 2. Utilizing the powerful nonlinear fitting ability of convolutional neural networks, the network is trained with dose distribution data of cervical cancer cases verified by physical dose calculation as the reference standard for supervised learning. The network can learn the mapping relationship between complex dose distribution and image features from accurately registered images, and accurately calculate the dose distribution results of internal and external irradiation. Compared with traditional dose calculation methods based on physical models or empirical formulas, it can better cope with the complexity of human tissues, individual differences, and nonlinear characteristics of the interaction between radiation and tissues, effectively improve the accuracy of dose calculation, and provide reliable data support for subsequent accurate dose superposition;

[0051] 3. By reasonably setting the weighting coefficients based on clinical experience and dosimetric characteristics for dose superposition, and using a variety of statistical analysis methods and visualization tools to conduct a comprehensive and in-depth analysis of the comprehensive dose volume histogram (DVH) and related dose data, the dose distribution of the target area and normal tissue can be more accurately evaluated. For example, by accurately calculating key indicators such as the dose mean, maximum, minimum, and dose coverage volume, the actual dose level and dose coverage degree received by different tissue areas can be carefully grasped, avoiding dose assessment deviations caused by rough analysis, and helping doctors make radiotherapy decisions that are more in line with the actual situation of the patient. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a schematic diagram of the overall structure of the cervical cancer internal and external irradiation dose superposition system based on deep learning of the present invention. DETAILED DESCRIPTION

[0053] 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.

[0054] 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.

[0055] like Figure 1 As shown, an embodiment of the present invention provides a cervical cancer internal and external irradiation dose superposition system based on deep learning, including a data input module, a deformation registration module, a dose calculation module, a dose superposition module and a result analysis module.

[0056] In this embodiment, the data input module receives the internal and external irradiation CT image data, verifies the integrity and accuracy of the internal and external irradiation CT image data, and performs denoising and standardization preprocessing operations, which specifically include the following steps:

[0057] Data reception:

[0058] Establishing data interface:

[0059] Create an input interface that is compatible with multiple medical image formats, with a focus on ensuring good support for the DICOM format (Digital Imaging and Communications in Medicine). This interface can be connected to data sources such as the hospital's Pictures Archiving and Communication System (PACS) and Treatment Planning System (TPS) to obtain internal and external irradiation CT image data of cervical cancer patients;

[0060] Data transmission and caching: Through the established interface, in accordance with the set communication protocol (such as the transmission rules specified in the DICOM standard protocol), the internal and external irradiation CT image data are received frame by frame or in batches from the data source, and the received data are temporarily stored in the system cache area, waiting for subsequent verification and preprocessing operations, and the relevant meta-information of the received data, such as patient identification, image acquisition time, equipment number, etc., is recorded to facilitate subsequent tracking and management;

[0061] Integrity check:

[0062] File format verification:

[0063] For each CT image data file received, first check its file header information and confirm whether the file header identifier is correct according to the standard specifications of the corresponding format such as DICOM. For example, a DICOM file should start with a specific byte sequence (such as "DICM"). If it does not meet the format requirements, it will be marked as format error data.

[0064] Parse the file structure, check whether the number, order and length of data elements are consistent with the standard format definition, and ensure that all necessary data elements exist and are in the correct format. For example, key elements such as image dimension information and pixel data length must be complete and accurate. Missing or abnormally formatted data elements are recorded and corresponding error messages are prompted.

[0065] Data volume verification:

[0066] According to the image acquisition parameters (such as the preset number of scanning layers, the pixel matrix size of each layer of the image, etc.) and the data volume range agreed in the transmission protocol, check whether the actual amount of internal and external irradiation CT image data received is within a reasonable range; if the data volume is significantly lower than or exceeds the expected range, it may mean that problems such as loss or redundancy occur during the data transmission process. In this case, the data is marked as questionable integrity and the abnormal data volume is recorded for subsequent investigation;

[0067] For multi-frame image sequences (such as CT images obtained by dynamic scanning), it is also necessary to check the continuity and integrity of the data between frames to ensure that there are no frame drops or duplicate frames in the image sequence, and to perform comparison and verification through information such as frame numbers and timestamps;

[0068] Image logo verification:

[0069] Check the patient identification information (such as name, medical record number, etc.) attached to each CT image data and the unique identifier (UID) of the image itself to ensure that the internal and external irradiation CT image data of the same patient can be accurately associated and that the data of different patients will not be confused with each other. Maintain the uniqueness and accuracy of data identification in the entire system. If there is an identification conflict or mismatch, issue a warning in time and suspend the subsequent processing process, waiting for manual intervention confirmation;

[0070] Accuracy check:

[0071] Gray value range calibration:

[0072] According to the imaging principle of CT images and the experience of the grayscale value range of normal clinical images, check whether the grayscale value of each image pixel is within a reasonable range; for example, for conventional human tissue CT images, the grayscale value is generally distributed in a specific Hounsfield Unit (HU) range, and bone tissue usually presents a higher HU value, while soft tissue corresponds to different typical HU value ranges. If there are pixels with abnormally high or low grayscale values ​​(beyond the reasonable HU range corresponding to normal tissue), the proportion of abnormal pixels is counted. If the proportion exceeds the set threshold, it is considered that the image may have accuracy problems, such as imaging abnormalities caused by incorrect scanning parameter settings, equipment failure, etc., and the image is marked as pending further confirmation;

[0073] At the same time, the logical relationship of the grayscale values ​​of the corresponding areas of the CT images under different irradiation methods (internal irradiation and external irradiation) of the same patient is compared. Under normal circumstances, the grayscale values ​​of similar tissues in different irradiation images should show a certain correlation and reasonable difference. If there is a situation that obviously violates this logic (such as the grayscale value difference of a certain area in the internal and external irradiation images is too large and does not conform to medical common sense), it indicates that there may be problems with data accuracy, which requires manual verification or other auxiliary means (such as referring to other imaging examination results of the same patient) to make judgments;

[0074] Geometric feature verification:

[0075] Analyze the geometric size, shape and other features of the internal and external irradiation CT images, and verify whether the geometric features of the images meet the normal human anatomical structure and the imaging geometry requirements of the scanning equipment by extracting the image contour, calculating the aspect ratio of the image, and checking the spatial position relationship between each layer of the image (such as whether the spacing between adjacent layers is uniform, whether the image has abnormal conditions such as rotation and translation), etc. For example, the CT image of the normal human pelvic area should show relatively continuous and anatomically consistent shape changes at different levels. If the image geometry is severely distorted or the layer spacing is abnormal, it may mean that the image accuracy is impaired due to problems such as patient position movement and equipment calibration deviation during the scanning process. The relevant images should be marked and the abnormal geometric feature information should be recorded;

[0076] Data consistency check:

[0077] For multiple internal and external irradiation CT image acquisitions of the same patient (such as reexamination images at different time points, etc.), check whether the key anatomical structure markers, lesion characteristics and other information in the images are consistent and coherent; if the positions of the originally stable anatomical landmarks change significantly in the before and after images, or the size and morphology of the known lesions change in a way that does not conform to the logic of the treatment process, it is necessary to doubt the accuracy of the data, indicating that there may be problems such as data entry errors, image registration deviations or misjudgment of the condition, so that targeted verification and processing can be carried out later;

[0078] Denoising operation:

[0079] Parameter estimates:

[0080] For each internal and external irradiation CT image received, the parameters required by the Wiener filter algorithm are first estimated; by analyzing the noise characteristics of the image, the power spectrum density estimate of the image is calculated. Some classic power spectrum estimation methods, such as the periodogram method and the autocorrelation method, can be used to obtain the power spectrum density of the original image signal. and noise power spectral density ;

[0081] At the same time, the point spread function that causes image blur is determined based on the imaging model of the image acquisition device or by analyzing a known standard image. The point spread function can be estimated or directly obtained based on the physical properties of the device, the imaging principle, and the previous calibration data. can be calculated accordingly;

[0082] Filter calculation:

[0083] According to the Wiener filtering algorithm formula:

[0084] For each coordinate position in the image The pixel value at (i.e. the original image pixel value containing noise), substitute the estimated parameters , , and , calculate the pixel value of the denoised image at this position , traverse all the pixels of the entire image to complete the filtering operation, so as to obtain the denoised internal and external irradiation CT image, reduce the noise components introduced by factors such as equipment electronic noise and scanning environment interference in the image, improve the image quality, and facilitate subsequent processing;

[0085] Standardized operations:

[0086] Statistical feature calculation:

[0087] For each internal and external irradiation CT image, calculate the mean grayscale value of the image pixels and standard deviation ; By traversing all the pixels of the image, accumulating the pixel values ​​and recording the number of pixels, according to the mean calculation formula (in, is the pixel value, is the total number of pixels) to calculate the mean; then calculate the standard deviation according to the standard deviation calculation formula to obtain the statistical characteristics of the grayscale value distribution of the image pixels. The standard deviation calculation formula is:

[0088] Standardization transformation:

[0089] According to the Z-score standardization formula , for each pixel value in the image , substitute the calculated mean and standard deviation , calculate the corresponding data value after standardization , all pixel values ​​of the entire image are subjected to such a standardized transformation, so that the pixel grayscale value distribution of different internal and external irradiation CT images has a similar mean (usually 0) and standard deviation (usually 1), eliminating the differences in image data magnitude and distribution caused by different scanning equipment, scanning parameters and other factors, and providing a more unified and standardized data foundation for subsequent deep learning-based module processing;

[0090] After the above-mentioned integrity and accuracy verification as well as denoising and standardization preprocessing operations, the data input module passes the qualified internal and external irradiation CT image data to the deformation registration module and enters the subsequent dose superposition processing flow of the system.

[0091] In this embodiment, the deformable registration module includes a deformable registration algorithm based on deep learning. The deformable registration algorithm uses a V-Net architecture combined with a Transformer-based algorithm to spatially align the internal and external irradiation CT image data output by the data input module, thereby realizing the superposition of images under different irradiation modes and obtaining a registered image. Specifically, the following steps are included:

[0092] Data preparation and preprocessing:

[0093] Receiving input data:

[0094] Obtaining internal and external irradiation CT image data from the data input module after integrity and accuracy verification, denoising, and standardization preprocessing; these image data correspond to the internal and external irradiation conditions of cervical cancer patients, and are stored in a suitable data structure (such as a pixel matrix representing an image in the form of a multidimensional array, etc.), while retaining image-related meta-information (such as patient number, image acquisition time, image space coordinate information, etc.) for subsequent processing and result traceability;

[0095] Data division and grouping:

[0096] According to clinical needs and the convenience of deep learning model training and reasoning, the received internal and external irradiation CT image data are divided and grouped according to certain rules; for example, they can be grouped according to individual patients to ensure that the internal and external irradiation images of the same patient appear in pairs, forming image pair data sets. Each image pair represents the corresponding CT image of the patient under different irradiation modes, providing basic data units for subsequent spatial alignment and registration;

[0097] V-Net architecture feature extraction:

[0098] Initial convolutional layer processing:

[0099] The divided internal and external irradiation CT image pairs are input into the initial convolutional layer of the V-Net architecture respectively; taking a single image as an example (assuming that the input feature map is , the output feature map is , the convolution kernel size is , the step length is , the convolution kernel weight is , the bias is ), feature extraction is performed according to the following convolution calculation formula: ;

[0100] in, Output feature map Pixel coordinate position Represents the input feature map The pixel value at the corresponding position, that is, the pixel intensity information at the corresponding position of the original image data when it is sampled by the convolution kernel sliding. It participates in feature calculation as the basic data of the convolution operation;

[0101] Through this layer of convolution operation, we start to extract preliminary image features from the original CT image, such as local features such as edges and textures, and generate new feature maps with different feature representations, preparing for subsequent deep feature extraction.

[0102] Downsampling and multi-scale feature extraction:

[0103] Downsampling layer operation (taking the maximum pooling as an example, assuming that the window size and step size are ): Down-sampling operation is performed on the feature map after the initial convolution layer. The calculation formula is: ;

[0104] in, The output feature map after downsampling is at coordinate The pixel value at , that is, from the input feature map The maximum value in the corresponding pooling window is taken as the output. In this way, the size of the feature map can be reduced while retaining the main features of the image, achieving data dimensionality reduction and expanding the receptive field, which helps to capture more macroscopic image features, extract feature information at different scales, and simulate the human visual system's perception process of the image from local to overall.

[0105] Repeat the convolution and downsampling process:

[0106] The V-Net architecture usually contains a combination of multiple convolutional layers and downsampling layers connected in sequence. The above convolution and downsampling operations are repeated. As the network level deepens, more abstract and semantically informative multi-scale features are gradually extracted, allowing the model to understand the structures and patterns at different levels in the image. For example, it gradually transitions from simple texture features to high-level features such as the shape and position relationship of tissues and organs, providing rich feature basis for subsequent accurate spatial alignment.

[0107] Upsampling and feature fusion (via skip connections):

[0108] Upsampling layer (taking transposed convolution as an example, the transposed convolution kernel size is , the step length is ): In the second half of the network, upsampling operation begins to restore the spatial resolution of the feature map, and its calculation formula is: ;

[0109] in, , are the weight and bias of the transposed convolution kernel, and the meanings of the other symbols are similar to those of the common convolution layer. Through the transposed convolution operation, the feature map that becomes smaller after downsampling is gradually enlarged to the spatial size corresponding to the original image, so as to restore the spatial resolution of the final registered image.

[0110] Skip connection (taking the addition method as an example, the low-resolution feature map is , the high-resolution feature map is ):

[0111] During the upsampling process, the feature maps at different levels are fused using jump connections to enhance the feature expression capability. The calculation formula is: ;

[0112] This fusion method can replenish some of the detail features lost during the downsampling process, so that the final restored feature map contains both the abstract semantic features extracted by the deep network and the local detail features in the original image, improving the richness and accuracy of the features and helping to achieve more accurate spatial alignment of images.

[0113] Transformer-based feature enhancement:

[0114] Feature sequence conversion:

[0115] Convert the image features extracted after processing by each layer of the V-Net architecture into a feature sequence form suitable for Transformer processing; usually, the image can be divided into multiple patches (small blocks) according to certain rules, and the feature representation of each patch is combined into a sequence (in, Indicates the sequence length, which is related to the number of patches divided into images. Represents feature dimension), that is, converting the two-dimensional image feature map into a one-dimensional feature sequence so as to be input into the Transformer structure for subsequent operations;

[0116] Self-attention mechanism calculation:

[0117] Generate query, key, and value: Generate query, key, and value through linear transformation, denoted as , the calculation formula is as follows: ; ; ;

[0118] in, , , is the learnable weight matrix, and its dimensions are (generally ); These linear transformation operations are designed to map the input feature sequence to different representation spaces to facilitate the subsequent calculation of the correlation between features;

[0119] Calculate the attention scores and weighted fusion:

[0120] The formula for calculating the attention scores is: ;

[0121] Among them, the softmax function is used to normalize the attention score to the probability distribution form. yes Through this self-attention mechanism calculation process, Transformer can capture the long-distance dependency between any two positions in the image feature sequence. Unlike traditional convolutional neural networks that are limited to the local receptive field, it can better understand the association between features in different regions of the image, further enhance the features extracted by V-Net and model the relationship, so that the model can more comprehensively consider the overall structural information of the image, which is very critical for accurately finding the spatial correspondence between internal and external irradiation CT images.

[0122] Feature return and fusion:

[0123] The feature sequence processed by Transformer based on the self-attention mechanism is converted back to the feature map form suitable for processing by the subsequent layers of V-Net, and is sent back to the subsequent layers of V-Net (such as the subsequent upsampling layer, etc.) to continue feature fusion, upsampling and other operations, and the global features enhanced by Transformer are fused with the original features of V-Net, making full use of the advantages of both, further optimizing the feature representation, and providing a more accurate feature basis for generating high-quality registered images;

[0124] Spatial alignment and image superposition generate registered images:

[0125] Spatial transformation parameter calculation:

[0126] Based on the fusion features obtained after V-Net combined with Transformer processing, the spatial transformation parameters between the internal and external irradiation CT images are calculated through a specific network output layer (which can be a fully connected layer or other structure); these parameters are used to describe how to align the internal irradiation image and the external irradiation image in space, such as transformation parameters such as translation, rotation, and scaling. The specific parameter form and number depend on the adopted spatial transformation model (such as affine transformation model, thin plate spline transformation model, etc.). Taking affine transformation as an example, a matrix parameter representing a two-dimensional or three-dimensional space affine transformation may be output to determine the image translation amount, rotation angle, scaling ratio and other information in each direction;

[0127] Spatial transformation operations:

[0128] According to the calculated spatial transformation parameters, one of the internal and external irradiation CT images (usually one of the irradiation methods is selected as the reference image, for example, the external irradiation image is used as the reference to transform the internal irradiation image) is subjected to the corresponding spatial transformation operation to achieve spatial alignment of the images; for example, for the affine transformation of a two-dimensional image, assuming that the spatial transformation matrix is , the coordinates of a point in the image are , the transformed coordinates It can be calculated by the following matrix multiplication: ;

[0129] By performing such coordinate transformation on all pixels, the internal illumination image is adjusted to a position that accurately corresponds to the external illumination image in space, thereby achieving spatial alignment of images under different illumination modes;

[0130] Image superposition and fusion (optional step, based on actual needs): After completing the spatial alignment, if it is necessary to further superimpose and fuse the internal and external irradiation images, some appropriate fusion strategies can be used, such as simple weighted average, linear combination based on pixel values, or more complex fusion algorithms (selected according to image characteristics and clinical requirements). The aligned internal and external irradiation images are superimposed to generate a fused image as the final registered image, so that they accurately correspond in space and contain comprehensive information under the two irradiation modes, which is convenient for the subsequent dose calculation module to calculate dose distribution and other operations;

[0131] Through the above complete steps, the deformable registration module uses the V-Net architecture combined with the Transformer-based algorithm to spatially align the internal and external irradiation CT image data output by the data input module, realize the superposition of images under different irradiation methods, and obtain the registered image, providing an accurate registration image basis for the subsequent process of the entire cervical cancer internal and external irradiation dose superposition system.

[0132] In this embodiment, the dose calculation module uses a deep learning model to calculate the dose distribution. The deep learning model is a convolutional neural network, which is used to process the registered image output by the variable registration module and output the distribution results of the internal and external irradiation doses. Specifically, the following steps are included:

[0133] Data preparation and preprocessing:

[0134] Receive the registered image data:

[0135] Obtain the registered image data after spatial alignment and superposition (if any superposition operation) from the deformable registration module; these images contain comprehensive information of the internal and external irradiation CT images of cervical cancer patients after precise registration, and are stored in a suitable data structure (such as a multi-dimensional array to represent the image pixel matrix and related spatial coordinate information, etc.), and are accompanied by corresponding patient identification and other meta-information to facilitate subsequent processing and result recording;

[0136] Data grouping and annotation:

[0137] Data grouping: Rationally group the registered image data according to individual patients and different irradiation scenarios (internal irradiation, external irradiation); for example, group the registered images acquired at different time points of the same patient together, or further subdivide the groups according to different irradiation doses, irradiation angles and other conditions, to ensure that each group of data has a certain correlation and similarity in terms of dose distribution characteristics, so as to facilitate subsequent targeted training and analysis;

[0138] Preparation of annotation information: Collect dose distribution data of cervical cancer cases verified by physical dose calculation (these data can come from past clinical practice, physical experiment simulation or output results of professional dose calculation software, etc.) as dose distribution annotation information (i.e. real label) of the corresponding registered images; the annotation information is also sorted in a way corresponding to the image data grouping, and the actual distribution values ​​of the internal and external irradiation doses corresponding to each registered image in different tissue areas (such as target area, normal tissue, etc.) are clarified to form the corresponding data set of "image-dose distribution annotation", which provides a basis for the supervised learning of convolutional neural networks;

[0139] Convolutional neural network model construction and initialization:

[0140] Network architecture selection and design:

[0141] Select a suitable convolutional neural network architecture as the basis, such as the classic VGG series, Res Net series, or design a network structure that suits the characteristics of the dose calculation task; determine the number of network layers, the size of each layer's convolution kernel, the step size, the padding method and other hyperparameters. The selection of these hyperparameters usually takes into account factors such as image size, resolution, the complexity of dose distribution, and computing resources;

[0142] Parameter initialization:

[0143] Initialize the weights (convolution kernel weights, etc.) and biases of each layer of the convolutional neural network. Common initialization methods include random initialization (such as normal distribution random initialization, uniform distribution random initialization, etc.), initialization of pre-trained models based on specific distributions (if there are pre-trained models for similar tasks available), etc. Reasonable initialization helps the network converge quickly and learn effective feature representations in the early stages of training;

[0144] Feature extraction and dose calculation based on convolutional neural network:

[0145] Initial convolutional layer feature extraction:

[0146] The registered image data is input into the initial convolutional layer of the convolutional neural network. Taking a single image channel as an example (assuming that the input feature map is , the output feature map is , the convolution kernel size is , the step length is , the convolution kernel weight is , the bias is ), feature extraction is performed according to the following convolution calculation formula: ;

[0147] in: Output feature map The pixel coordinate position in the image represents the coordinates corresponding to the new feature value generated at this position after the convolution operation; Represents the input feature map The pixel value at the corresponding position, that is, the pixel intensity information at the corresponding position of the registered image data when it passes through the convolution kernel sliding sampling, which participates in feature calculation as the basic data of the convolution operation;

[0148] Through this layer of convolution operation, we start to extract preliminary image features from the original registered image, such as boundary features of different tissue regions, density-related features, etc., and generate new feature maps with different feature representations, preparing for subsequent deep feature extraction.

[0149] Multi-layer convolution and pooling operations: Convolutional neural networks usually contain multiple sequentially connected convolutional layers and pooling layers (such as maximum pooling, average pooling, etc.) to further extract more abstract and semantically informative features;

[0150] Convolution layer: Repeat the above convolution operation. As the network level deepens, more complex image features are gradually extracted. For example, simple tissue boundary features are gradually transitioned to high-level features that can reflect the physical properties of different tissues such as radiation absorption and scattering. The convolution kernel size, step size and other parameters of different layers can be adjusted as needed to meet the requirements of feature extraction at different scales.

[0151] Pooling layer (taking the maximum pooling as an example, assuming that the pooling window size and step size are ): Its calculation formula is: ;

[0152] here The output feature map after pooling is at coordinate The pixel value at, which means from the input feature map The maximum value in the corresponding pooling window is taken as the output. The pooling operation can reduce the size of the feature map while retaining the main features of the image, achieve data dimensionality reduction, and expand the receptive field, which helps to capture more macroscopic image features and simulate the human visual system's perception process of the image from local to overall, while reducing the amount of calculation and avoiding problems such as overfitting.

[0153] Fully connected layer and dose prediction:

[0154] After extracting features through multiple convolution and pooling layers, the obtained high-level feature map is flattened and converted into a one-dimensional vector form, and then input into the fully connected layer network structure; the fully connected layer performs nonlinear transformation and integration on the flattened feature vector through multiple neurons, and outputs the predicted dose distribution results according to the pre-set number of network output nodes (corresponding to the relevant dimensions of dose distribution, such as dose values ​​in different tissue regions, dose values ​​at different spatial coordinate positions, etc.);

[0155] Taking the prediction of dose values ​​at multiple spatial coordinate positions in a certain tissue region as an example, assuming that the number of output nodes of the fully connected layer is consistent with the number of spatial coordinate points to be predicted, the value of each output node represents the predicted dose value at the corresponding coordinate position, and the entire output vector constitutes the dose distribution prediction result in the tissue region; this process is essentially a regression task, which continuously optimizes the network parameters by training the network to minimize the difference between the predicted dose distribution and the true annotated dose distribution (such as measured by loss functions such as mean square error), so that it can accurately learn the mapping relationship of dose distribution from the registered image, thereby outputting high-precision internal and external irradiation dose distribution results;

[0156] Model training and optimization (for supervised learning):

[0157] Loss function definition: Select a suitable loss function to measure the difference between the dose distribution results predicted by the model and the actual annotated dose distribution. Commonly used loss functions include the mean squared error (MSE) loss function, and its calculation formula is as follows: ;

[0158] in: is the number of training samples, i.e., the number of groups of registered images and their corresponding annotated dose distribution data participating in one training iteration; Indicates The true label value of each sample (i.e., the actual dose value of the corresponding tissue area and the corresponding spatial coordinate position in the standard dose distribution annotation information); Represents the network prediction The output value of each sample (i.e. the predicted dose value at the corresponding position output by the network);

[0159] In addition to the mean square error, other loss functions suitable for regression tasks can also be selected according to actual conditions, such as the mean absolute error (MAE). The choice of loss function aims to accurately reflect the deviation between the predicted result and the true value, and guide the network to update parameters in the direction of reducing this deviation.

[0160] Optimization algorithm selection and parameter update: Select a suitable optimization algorithm to update the parameters of the convolutional neural network. Common optimization algorithms include stochastic gradient descent (SGD) and its variants (such as Adagrad, Adadelta, Adam, etc.). Taking stochastic gradient descent as an example, the basic formula for updating parameters is: ;

[0161] in: Indicated in Model parameters at the iteration (such as convolution kernel weights and biases of each layer); Indicates The updated model parameters at the iteration; The learning rate controls the step size of each parameter update. The setting of the learning rate is crucial. If it is too large, the parameters may not converge or even diverge. If it is too small, the training process will be too slow. It needs to be reasonably adjusted according to the specific network structure, data characteristics and other factors. Represents the loss function About parameters The gradient indicates the direction of parameter update. By calculating the gradient information, the optimization algorithm dynamically adjusts the network parameters according to the feedback of the loss function, so that the model is continuously optimized during the training process, gradually reducing the error between the predicted dose distribution and the actual dose distribution.

[0162] During the training process, the prepared registered image dataset and its annotation information are input into the convolutional neural network in batches according to the set batch size (batchsize, i.e. the number of samples input into the network for training each time), and the predicted dose distribution is obtained through forward propagation calculation, and then the loss value is calculated through the loss function. Then, the optimization algorithm is used to calculate the gradient and update the network parameters based on the loss value through back propagation. This process is repeated continuously, and after multiple training cycles (epochs, i.e. the number of times the entire dataset is traversed), the performance indicators (such as the accuracy of dose distribution prediction, mean square error, etc.) of the model on the validation set (part of the data separated from the original dataset to evaluate the model performance and prevent overfitting) reach satisfactory results or no longer have significant improvements;

[0163] Model reasoning and dose distribution result output:

[0164] After completing the training and optimization of the model, the newly input registered image from the deformable registration module (i.e. the actual application stage) is directly input into the trained convolutional neural network. After the forward propagation calculation of the network, the corresponding internal and external irradiation dose distribution results are directly output. These results can be presented in the form of dose numerical matrix, dose cloud map, etc., which can intuitively show the internal and external irradiation dose distribution in different tissue areas and different spatial coordinate positions, and provide key data support for the subsequent dose superposition module and result analysis module.

[0165] In this embodiment, the dose superposition module performs weighted superposition on the internal and external irradiation doses output by the dose calculation module to generate a comprehensive dose volume histogram; specifically, the following steps are included:

[0166] Data acquisition and preparation:

[0167] Receiving dose data:

[0168] Obtain the calculated internal and external radiation dose distribution results from the dose calculation module; these results are usually presented in a specific data format, such as a three-dimensional dose value matrix (where each element corresponds to the dose value at a certain coordinate position in space, and the matrix dimension corresponds to the image space dimension, for example, a two-dimensional matrix for a two-dimensional image and a three-dimensional matrix for a three-dimensional image), and are accompanied by corresponding spatial coordinate information and tissue region identification (used to distinguish target areas, different normal tissues, etc.), while clarifying the corresponding data to which patient and other meta-information to ensure data traceability and accuracy;

[0169] Parameter preparation:

[0170] Based on clinical experience and the dosimetric characteristics of internal and external irradiation of cervical cancer, the weight coefficient for weighted superposition is determined; the internal irradiation weight is usually set to , the external radiation weight is ,and ; The determination of these weight coefficients may require reference to a large amount of past clinical case data, professional radiotherapy guidelines and related medical research results to ensure that the weight distribution can reasonably reflect the proportion of different irradiation methods in the comprehensive dose contribution. For example, for the treatment of cervical cancer in certain specific parts, internal irradiation contributes more to the local dose of the target area, and a higher weight value may be given to internal irradiation accordingly, while external irradiation mainly acts on a larger range of surrounding tissues, and the weight is set according to its actual degree of effect;

[0171] Dose-weighted superposition operation:

[0172] Space coordinate alignment (if necessary):

[0173] Ensure that the internal and external irradiation dose distribution data are consistent in spatial coordinates. If there are coordinate differences (due to different calculation processes or data storage formats, etc.), coordinate transformation operations are required to align them. For example, if the spatial resolution, origin coordinates or coordinate axis directions of the internal and external irradiation dose data are inconsistent, they are adjusted to a unified spatial coordinate system through translation, scaling, rotation and other coordinate transformation methods to ensure that the dose values ​​at the corresponding coordinate positions can be accurately superimposed in the subsequent operation;

[0174] Point-by-point dose overlay:

[0175] For each corresponding coordinate position in space (Take the three-dimensional case as an example, the two-dimensional case can be simplified accordingly ), the comprehensive dose value is calculated according to the set weighted superposition formula; assuming that the internal irradiation dose distribution function is The external radiation dose distribution function is , then the comprehensive dose distribution The calculation formula is: ;

[0176] Traverse all coordinate positions in the entire dose distribution space and calculate the comprehensive dose value point by point, so as to obtain the comprehensive dose distribution result that integrates the internal and external irradiation dose information. The result is also expressed in the form of a three-dimensional (or two-dimensional) dose numerical matrix, reflecting the comprehensive dose situation at different spatial positions, laying the foundation for the subsequent generation of dose volume histogram;

[0177] Generate a comprehensive dose volume histogram (DVH):

[0178] Tissue region division and dose statistics: Based on the preset tissue region identification (for example, the target area and different normal tissues such as bladder and rectum have been distinguished when the dose calculation module is output), the data corresponding to different tissue regions in the comprehensive dose distribution results are extracted separately; for each tissue region, the number of voxels (i.e., the smallest unit representing a certain volume in the image, whose dose value is determined by the dose value of the corresponding coordinate position in the comprehensive dose distribution matrix) falling into each dose interval in the region is counted;

[0179] Draw a comprehensive dose volume histogram (DVH):

[0180] Based on the voxel number data of different tissue regions in each dose interval obtained by the above statistics, the dose interval is used as the horizontal axis, and the corresponding voxel number (or the ratio of the voxel number to the total volume of the tissue, which can be obtained by further calculating the total volume of the voxels and converting the ratio) is used as the vertical axis to draw a comprehensive dose volume histogram (DVH) of each tissue region. Usually, professional drawing tools or programming libraries (such as matplotlib in Python, etc.) can be used to achieve visual drawing, and generate a graph that intuitively shows the dose distribution of each tissue region. The trend of the curve (or bar chart, etc., depending on the drawing method) intuitively reflects the trend of the change of the irradiated volume of the tissue with the increase of dose, helping doctors to quickly evaluate whether the target area has reached the expected treatment dose and the potential dose risk of normal tissue, etc., providing an important basis for clinical radiotherapy decision-making;

[0181] Through the above complete steps, the dose superposition module realizes the weighted superposition of the internal and external irradiation doses output by the dose calculation module, and generates a comprehensive dose volume histogram. It plays a key connection and analysis role in the cervical cancer internal and external irradiation dose superposition system based on deep learning, and provides an intuitive and important data display form for the subsequent result analysis module.

[0182] In this embodiment, the result analysis module uses statistical analysis methods and visualization tools to analyze the comprehensive dose volume histogram and dose data generated by the dose superposition module to evaluate the dose distribution of the target area and normal tissues, specifically including the following steps:

[0183] Data acquisition and organization:

[0184] Receiving data:

[0185] Obtain the generated integrated dose volume histogram (DVH) and related integrated dose distribution data (such as the dose value matrix at different spatial coordinate positions obtained after weighted superposition, etc.) from the dose superposition module; at the same time, clarify the patient identification and tissue area division information corresponding to these data (such as which areas belong to the target area and which are specific normal tissues such as the bladder, rectum, small intestine, etc.) to ensure the integrity and accuracy of the data and prepare for subsequent analysis;

[0186] Data integration and preprocessing (if necessary):

[0187] Data format unification: Check whether the received data format is consistent. If there is data in different formats (for example, DVH data may be stored in a specific data table structure, while dose distribution data is in the form of a multi-dimensional array, etc.), perform format conversion operations so that it can be easily processed in a unified manner in the subsequent analysis process;

[0188] Data cleaning: Check whether there are outliers, missing values, etc. in the data; for outliers (such as isolated data points that are obviously beyond the normal dose range, which may be caused by errors in the data collection or calculation process), decide whether to correct them (if a reasonable correction method can be determined, such as through adjacent data point interpolation) or mark them and deal with them appropriately in subsequent analysis (such as excluding the outliers in statistical analysis); for missing values, appropriate filling strategies (such as mean filling, adjacent data point filling, etc.) can be used to ensure that the data quality meets the analysis requirements;

[0189] Application of statistical analysis methods:

[0190] Mean dose calculation:

[0191] For the target area and normal tissue areas: for a given integrated dose distribution function (Take the three-dimensional case as an example, the coordinate dimension can be simplified accordingly in the two-dimensional case) ), the mean dose The calculation formula is: ;

[0192] In actual operation, since the dose distribution data is usually stored in the form of discrete spatial coordinate points (voxels), the above integral operation can be implemented by summation approximation calculation; the specific steps are as follows:

[0193] First, determine the area of ​​interest The set of voxels included (defined by tissue region division information);

[0194] For each voxel , get the corresponding dose value and voxel volume (The voxel volume is a fixed value when the image spatial resolution is determined. For example, for isotropic voxels, the volume is equal to the cube of the edge length);

[0195] Then follow the formula: ;

[0196] Perform calculations, i.e. accumulate areas The product of the dose value of all voxels in the region and the voxel volume is used as the numerator (representing the total volume integral of the dose in the region), and the cumulative region The volume of all voxels in the The total volume of the body), the average dose value is obtained by dividing the two;

[0197] Significance and application: Calculating the mean dose helps to understand the average radiation dose level of the target area and normal tissues during the entire radiotherapy process, and compare the expected treatment dose (for the target area) and the normal tissue tolerance dose (for normal tissue) to determine whether the current dose distribution is reasonable. For example, if the average dose of the target area is much lower than the expected treatment dose, it may indicate that the radiotherapy plan needs to be adjusted to improve the efficacy; if the average dose of a normal tissue is close to or exceeds its tolerance dose, attention should be paid to the risk of radiation damage to the tissue;

[0198] Calculation of maximum and minimum dose:

[0199] Traversal search: Given a comprehensive dose distribution function Within the covered space (covering the target area and the space where all normal tissues are located), by traversing all coordinate positions Corresponding dose values, find the maximum value and minimum value ,Right now: ; ;

[0200] in, Indicates the entire spatial range involved in the dose distribution. In actual operation, it is also based on discrete voxel data for traversal search. Starting from the dose value of the first voxel, it is compared with the dose values ​​of subsequent voxels in turn, and the maximum and minimum dose values ​​currently encountered are recorded until all voxels are traversed;

[0201] Analysis value: The maximum dose can help determine whether there is a "hot spot" area with excessive local dose in the entire radiotherapy area. Especially for normal tissues, excessive local dose may cause serious radiation damage. If the maximum dose is found in normal tissues far exceeding their tolerance dose, it is necessary to consider adjusting irradiation parameters to avoid damage. The minimum dose can reflect whether there is an area with insufficient dose. For the target area, a too low minimum dose may mean that some tumor cells cannot be effectively killed, affecting the treatment effect.

[0202] Dose coverage volume calculation:

[0203] Set threshold and judge: For a given dose threshold (This threshold can be set according to factors such as clinical treatment goals and normal tissue tolerance. For example, for the target area, it may be set as the minimum dose threshold required to achieve a certain treatment effect, and for normal tissue, it may be set as the critical dose threshold that may cause damage). Calculate the volume with a dose greater than or equal to the threshold. ; Specifically by judging the comprehensive dose distribution function Whether the coordinates are satisfied at each position To do this, then count the spatial volume formed by the coordinate positions that meet the conditions;

[0204] Calculation implementation (voxel-based): traverse all voxels, for each voxel , determine its dosage value Is it greater than or equal to , if the condition is met, the volume of the voxel This is included in the cumulative value of the coverage volume, and the total volume that meets the dose threshold requirement is finally obtained. , the calculation formula can be expressed as: ;

[0205] Clinical significance: By calculating the dose coverage volume, we can intuitively understand how much volume in the target area has reached the expected treatment dose, which is critical for evaluating the effectiveness of treatment. For example, clinically, we often focus on the target area. Minimum dose received by volume Indicators such as the ratio of the coverage volume to the total volume of the target area are measured by setting corresponding thresholds; for normal tissues, calculating the coverage volume that exceeds its tolerance dose is helpful to assess the possibility and scope of radiation damage.

[0206] Dose Volume Histogram (DVH) Indicator Analysis:

[0207] Extract key parameters: Extract more clinically meaningful parameters from the integrated dose-volume histogram (DVH), such as (express The dose received by the volume, e.g. It is often used to evaluate the target dose coverage. etc. can be used to observe the low-dose irradiation of normal tissues). (Accept The volume of the dose is a percentage of the total volume, e.g. Indicators such as irradiation dose-volume relationship can be used to analyze the normal tissue exposure dose-volume relationship). These parameters are usually obtained by finding the dose value at the corresponding proportional volume or the volume ratio at the corresponding dose value on the DVH curve. This can be achieved with the help of professional medical image analysis software or self-written algorithms.

[0208] Comprehensive evaluation: Combine these DVH parameters of different tissue regions (target area and normal tissues) for comprehensive analysis, compare clinical treatment guidelines, past successful treatment cases and other standards, evaluate the rationality of dose distribution in various tissues under the current radiotherapy plan, and determine whether the radiotherapy plan needs to be adjusted and optimized. For example, if the target area The dose is lower than the standard requirement, and the normal tissue If the dose is too high, it indicates that the irradiation angle may need to be replanned, the weight of internal and external irradiation doses may need to be adjusted, etc., so as to improve the target area dose coverage while reducing the irradiation dose to normal tissues;

[0209] Visualization tool application:

[0210] Dose cloud drawing:

[0211] Data mapping and rendering: Map the comprehensive dose distribution data (dose value matrix) into a three-dimensional (or two-dimensional, depending on display requirements) space, and assign different colors or grayscale levels according to the size of the dose value. For example, the lower dose area is represented by light blue or light gray, and as the dose value increases, the color gradually transitions to dark blue, purple or dark gray, etc. (the color mapping scheme can be adjusted according to clinical habits and visualization effects); render and draw through professional visualization software or programming libraries (such as VTK, ITK-SNAP and other tools commonly used in the field of medical image visualization, or Python matplotlib, Mayavi and other libraries combined with relevant medical image processing extensions) to generate a dose cloud map that intuitively displays the spatial distribution of dose in the entire radiotherapy area;

[0212] Add interactive functions (optional): Add interactive functions to the dose cloud map, such as zooming, rotating, and slice viewing (for 3D cloud maps, you can view 2D slice images of any plane), etc., to facilitate doctors to observe the relationship between dose distribution and human tissue anatomical structure from different angles and levels, and to more clearly judge the irradiation of the target area and normal tissues. For example, you can check whether there is dose unevenness in the target area at a certain level, and whether the dose gradient with the surrounding normal tissues is reasonable.

[0213] Dose curve drawing (optional, according to specific analysis requirements):

[0214] Select the drawing content: According to the analysis focus, you can draw the dose-volume curves of different tissue regions (based on DVH data), that is, use the dose value as the horizontal axis and the corresponding volume percentage as the vertical axis to draw the curve of each tissue region to show the change trend of the irradiated volume of each tissue as the dose increases; you can also draw the dose change curves in different spatial coordinate directions (for example, along a certain axis to view the distribution change of the dose on the straight line), etc., to further analyze the spatial characteristics of the dose distribution;

[0215] Curve display and interpretation: The drawn dose curve is displayed in the same coordinate system (different tissue regions or curves in different directions are distinguished by different colors or line types) for easy comparison and analysis; by observing the trend of the curve, the position of the intersection, the slope and other characteristics, the dose relationship between different tissues and the distribution law of the dose in space can be interpreted. For example, the intersection of two curves may indicate that the irradiated volumes of the two tissues are equivalent at this dose level, which can be used to analyze the rationality of the dose distribution between adjacent tissues, etc.;

[0216] Comprehensive visualization (integrating multiple visualization forms):

[0217] Integrate visualization elements such as dose cloud map, dose curve and integrated dose volume histogram (DVH) into a display interface (which can be a specially developed medical imaging analysis software interface, or realized through a web-based visualization platform). Doctors can view dose distribution display information from different angles and in different forms through switching, linkage and other operations, and evaluate the dose distribution of target area and normal tissue more comprehensively and intuitively, so as to make scientific and reasonable radiotherapy decisions based on these visualization analysis results, such as deciding whether to adjust the parameters of internal and external irradiation, modify the radiotherapy plan, etc., to achieve better treatment effect and reduce the risk of complications;

[0218] Through the above specific operation steps, the result analysis module can make full use of statistical analysis methods and visualization tools to conduct in-depth analysis of the data provided by the dose superposition module, provide strong support for dose assessment and scheme optimization during cervical cancer radiotherapy, and improve the accuracy and safety of radiotherapy.

[0219] 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. A deep learning-based cervical cancer internal and external irradiation dose superposition system, characterized by: include: Data input module; The data input module receives the internal and external irradiation CT image data, performs integrity and accuracy verification on the internal and external irradiation CT image data, and performs denoising and standardization preprocessing operations; Deformable registration module; the deformable registration module includes a deep learning-based deformable registration algorithm, which uses a V-Net architecture combined with a Transformer-based algorithm to spatially align the internal and external irradiation CT image data output by the data input module, thereby achieving superposition of images under different irradiation modes and obtaining a registered image; Dose calculation module; the dose calculation module uses a deep learning model to calculate the dose distribution, and the deep learning model is a convolutional neural network, which is used to process the registered image output by the variable registration module and output the distribution results of the internal and external irradiation doses; Dose superposition module: the dose superposition module performs weighted superposition on the internal and external irradiation doses output by the dose calculation module to generate a comprehensive dose volume histogram; Result analysis module; The result analysis module uses statistical analysis methods and visualization tools to analyze the comprehensive dose volume histogram and dose data generated by the dose superposition module to evaluate the dose distribution of the target area and normal tissues.

2. The deep learning-based cervical cancer internal and external irradiation dose superposition system according to claim 1 is characterized by: In the data input module, the algorithm used in the denoising operation is the Wiener filtering algorithm, and the calculation formula of the Wiener filtering algorithm is: ； in, Indicates the coordinates of the image after denoising The pixel value at is the point spread function that causes image blur, Its complex conjugate, and represent the noise power spectral density and the power spectral density of the original image signal respectively, is the original image with noise at coordinates The pixel value at , that is, the input image function to be denoised; The method used for standardization is Z-score standardization, and its calculation formula is: ;in, Represents the corresponding data value after standardization, is the original data value, is the mean of the original data, is the standard deviation of the original data.

3. The deep learning-based cervical cancer internal and external irradiation dose superposition system according to claim 2 is characterized by: The working mechanism of the deformation registration module is as follows: The V-Net architecture consists of multiple sequentially connected convolutional layers, downsampling layers, upsampling layers, and skip connection structures, which are used to extract and fuse image features and restore image spatial resolution; The convolutional layer assumes that the input feature map is , the output feature map is , the convolution kernel size is , the step length is , the convolution kernel weight is , biased to , a single convolutional layer is calculated according to the following formula: ； in, Output feature map Pixel coordinate position Represents the input feature map Corresponding position pixel value; In the downsampling layer, the pooling window size and step size are set to , the downsampling layer calculation formula is: ； in, The output feature map after downsampling is at coordinate The pixel value at , that is, from the input feature map Take the maximum value within the corresponding pooling window as output; In the upsampling layer, the transposed convolution kernel size is , the step length is , the upsampling layer calculation formula is: ； in, , are the weights and biases of the transposed convolution kernel respectively; The jump connection fuses the feature maps at different levels. The low-resolution feature map is , the high-resolution feature map is , the fused feature map The calculation formula is: ; In the Transformer-based algorithm, for the input image feature sequence , represents the sequence length, Represents the feature dimension, and generates query, key, and value through linear transformation respectively. , the calculation formula is as follows: ;in, , , is the learnable weight matrix, and its dimensions are ,in, ; The formula for calculating the attention score is: ； Among them, the softmax function is used to normalize the attention score to the probability distribution form. yes The transpose of The image features extracted by V-Net are input into the Transformer-based module. After the self-attention mechanism is used to capture the relationship between features, they are then sent back to the subsequent layers of V-Net to continue feature fusion and upsampling operations, and finally the aligned image is output.

4. The deep learning-based cervical cancer internal and external irradiation dose superposition system according to claim 3 is characterized by: The training mechanism of the deformable registration module is as follows: The network parameters were optimized and adjusted by training on a training data set containing a large number of cervical cancer cases with internal and external irradiation CT images and their standard registration results. The loss function used in the training process was the mean square error loss function, and the calculation formula was: ； in, is the number of training samples, Indicates The true label value of the sample, Represents the network prediction The output value of each sample; The optimization algorithm is the stochastic gradient descent algorithm, and its basic formula for updating parameters is: ； in, Indicated in The model parameters at the iteration, Indicates The updated model parameters at the iteration, is the learning rate, which controls the step size of each parameter update. Represents the loss function About parameters The gradient of , indicating the direction of parameter update.

5. The deep learning-based cervical cancer internal and external irradiation dose superposition system according to claim 1 is characterized by: In the dose calculation module, the convolutional neural network uses the dose distribution data of cervical cancer cases verified to be accurate by physical dose calculation as a reference standard for supervised learning during training. The network's hyperparameters include the convolution kernel size k, the step size s, and the number of layers l. The hyperparameters are adjusted and optimized based on the training results and the accuracy of the verification set.

6. The deep learning-based cervical cancer internal and external radiation dose superposition system according to claim 5 is characterized by: In the dose superposition module, the weighted coefficient of the dose superposition algorithm is set according to clinical experience and the dosimetric characteristics of internal and external irradiation of cervical cancer, and is used to allocate the weights of internal and external irradiation doses in the process of generating the comprehensive dose volume histogram. Assuming that the internal irradiation dose distribution function is , the external radiation dose distribution function is , set the internal illumination weight to , the external radiation weight is ,and The comprehensive dose distribution The calculation formula is: ； in, Represents the spatial coordinate position based on the comprehensive dose distribution To generate a comprehensive dose-volume histogram.

7. The deep learning-based cervical cancer internal and external radiation dose superposition system according to claim 6 is characterized by: In the result analysis module, the statistical analysis method at least includes statistical operations for calculating the dose mean, maximum, minimum and dose coverage volume. The visualization tool can generate dose cloud maps, dose curves and comprehensive dose volume histogram comparison maps that intuitively display the dose distribution. The specific calculation and display methods are as follows: Assume that the dose distribution function is In the area of ​​interest Mean dose The calculation formula is: ； in, represents a volume differential element, and the integral is over the entire region of interest The numerator represents the total volume integral of the dose in the area, and the denominator represents the area The total volume of the sample is divided by the average dose. Given a dose distribution function Within the covered space, by traversing all coordinate positions Corresponding dose values, find the maximum value and minimum value ,Right now: ； ； in, Indicates the entire spatial range involved in the dose distribution; For a given dose threshold , calculate the volume whose dose is greater than or equal to the threshold , by determining the dose distribution function Whether the coordinates are satisfied at each position ,Then count the space volume formed by the coordinate positions that meet the conditions, and the calculation formula is expressed as: ； Among them, the integral range is to meet the dose greater than or equal to the threshold The spatial region composed of all coordinate positions of .

Images