CN115170482A

CN115170482A - Automatic delineation method for cervical cancer target area and automatic generation method for adaptive radiotherapy plan

Info

Publication number: CN115170482A
Application number: CN202210701129.XA
Authority: CN
Inventors: 金献测; 谢聪颖; 艾遥; 肖承健
Original assignee: First Affiliated Hospital of Wenzhou Medical University
Current assignee: First Affiliated Hospital of Wenzhou Medical University
Priority date: 2022-06-20
Filing date: 2022-06-20
Publication date: 2022-10-11

Abstract

The invention discloses an automatic delineation method of a cervical cancer target area and an automatic generation method of a self-adaptive radiotherapy plan, which comprise the following steps: 1) Preprocessing an image; 2) Performing target segmentation on the input image based on a RefineSeplus 3D segmentation model, wherein the RefineSeplus 3D segmentation model comprises: the left encoder is formed by cascading K-level residual error networks and comprises a plurality of rolling blocks, wherein two adjacent rolling blocks realize the sequential reduction of the resolution ratio through downsampling; and the right decoder comprises K cascaded decoding convolutional layers, 3DRefine blocks are arranged between the decoding convolutional layers of the corresponding stage and the convolutional blocks, and the K +1 decoding convolutional layers are connected with the 3DRefine blocks corresponding to the K decoding convolutional layers. The automatic segmentation and dose prediction model provides a reasonable dose distribution reference while exhibiting a good automatic segmentation effect on the test data set.

Description

Automatic delineation method for cervical cancer target area and automatic generation method for adaptive radiotherapy plan

Technical Field

The invention particularly relates to an automatic delineation method of a cervical cancer target area and an automatic generation method of a self-adaptive radiotherapy plan.

Background

Gynecological tumors (Gynecology Oncology) are killers threatening the health of women, and the incidence rate is high and tends to increase year by year. Cervical Cancer (Cervical Cancer) is three common malignancies among gynecological tumors, and the age of onset is gradually becoming younger. For cervical cancer, while the incidence of cervical cancer has declined in developed countries, it remains the fourth leading cause of cancer deaths in women worldwide, with about 52 million and 26 million deaths worldwide annually accounting for 7.5% of all female cancer deaths, with the majority of patients already being locally advanced at the time of diagnosis (i.e., IB1 to IVA stages). The standard treatment regimen for locally advanced patients is concurrent chemoradiotherapy and brachytherapy, with an expected 5-year overall survival rate that varies from 80% (stage IB) to 15% (stage IVA-B), with approximately 30-40% of patients relapsing. According to national cancer statistics data reports of cancer centers in China, the incidence rate of cervical cancer is high, and the incidence rate of malignant tumors of female reproductive systems is 2 nd, and is second to breast cancer.

At present, synchronous radiotherapy and chemotherapy become the main treatment means for cervical cancer, however, in the actual precise radiotherapy process, the shape and position of the target region among the treatments can change along with the change of the position and shape of the treatment part and the position and the shape of the organs at risk, and the dose of the target region is easy to be insufficient and the incidence rate of complications is easy to be improved only by the plan made by CT before the treatment. Therefore, researchers have proposed that intensity modulated radiation therapy should be used with caution in the treatment of cervical cancer. Adaptive Radiotherapy (ART) is a new technology developed in intensity modulated radiotherapy, and aims to improve the accuracy of tumor radiotherapy, make the shapes of an irradiation field and a target region consistent on a three-dimensional level, realize high-dose irradiation on a tumor target region, and reduce the possibility that surrounding normal tissues are irradiated by high dose to the maximum extent, so that the advantages of intensity modulated radiotherapy can be exerted.

Disclosure of Invention

Aiming at the defects in the prior art, the invention aims to provide an automatic delineation method for a cervical cancer target area and an automatic generation method for a self-adaptive radiotherapy plan.

In order to realize the purpose, the invention provides the following technical scheme:

an automatic delineation method for a cervical cancer target area comprises the following steps:

1) Image preprocessing, namely uniformly inputting the size of an image through image size normalization operation;

2) Performing target region segmentation on the input image based on a RefineSeplus 3D segmentation model,

wherein the RefineNetplus3D segmentation model comprises:

the left encoder is used for performing down-sampling operation on an input image step by step to extract tumor features, is formed by cascading K-level residual error networks and comprises a plurality of rolling blocks, wherein two adjacent rolling blocks realize sequential reduction of resolution through down-sampling;

a right decoder for restoring the bottom layer characteristics to the output image of the resolution corresponding to the original image, which comprises K cascaded decoding convolution layers, wherein 3DRefine blocks are arranged between the decoding convolution layer of the corresponding stage and the convolution Block of the left encoder, and the K +1 decoding convolution layers are connected with the 3DRefine blocks corresponding to the K decoding convolution layers,

the 3DRefine Block comprises:

a residual convolution unit for generating a high-resolution predicted image by connecting and fusing a rough high-level semantic feature and a fine low-level feature through a long residual;

chain residual pooling for capturing context and background information of high resolution image regions;

merging, connecting the structures of the features processed by the residual convolution unit in two different paths, obtaining input from a residual network, integrating the input into a high-resolution feature map,

and a plurality of groups of RELU activation and batch normalization modules are arranged in the residual convolution unit and the chain type residual pooling and fusing.

The sampling rate of the first layer residual network is 1/2.

The volume block is a 3 x 3D volume block.

The chain residual pooling consists of a plurality of residual structure convolutions and pooling layers, arranged in a residual manner.

A self-adaptive radiotherapy plan automatic generation method based on the cervical cancer target area automatic delineation method comprises the following steps:

1) Building a 3DResUNet convolutional neural network for predicting the dose distribution;

2) Establishing a training database, wherein the training database is a cervical cancer clinical radiotherapy plan which comprises a case target area and case dose distribution;

3) Performing deep learning on the training database to train a 3DResUNet convolutional neural network;

4) Inputting the image to be predicted into a trained 3DRESUNet convolutional neural network to obtain the predicted three-dimensional dose distribution of the target area, wherein the image to be predicted comprises 1 image of CT image information, one or two images containing the contour information of the planned target area, and 5 images containing the body contour of the patient.

The 3DResUNet convolutional neural network includes:

the left encoder is used for extracting multilayer and multi-scale three-dimensional characteristics of three-dimensional matrixes such as CT images, organs at risk, target structures, RTDose and the like, and comprises 4 layers of residual networks which are sequentially cascaded, wherein each residual network comprises a downsampling convolution module which comprises two convolution layers with the length of 3 multiplied by 3 and the step length of 2 multiplied by 2;

a right decoder for regression fitting of three-dimensional features to dose distribution, comprising 4 layers of sequentially cascaded residual networks, each residual network comprising an upsampling convolution module,

the up-sampling convolution module is connected with the down-sampling convolution module of the upper layer in a jump connection mode, and after passing through a convolution layer with the size of 3 multiplied by 3 and a 3D convolution layer with the step size of 1 multiplied by 1, the up-sampling convolution module performs convolution operation with the down-sampling convolution module and performs convolution operation with the up-sampling convolution module of the current layer.

The CT image of the patient in the training database comprises a radiotherapy structure of a PTV contour and dose information of the patient, and meanwhile, the contour information of the PTV contour and the contour information of an organ at risk are extracted from an RTST file of cervical cancer data and used as the input of a cervical cancer dose prediction model, the PTV contour is sketched by an experienced clinician, and the number behind the PTV represents the prescription dose which a target area should receive.

The method comprises the steps of preparing labels for CT images to be preprocessed, wherein the labels are read from contours corresponding to names in RTStructure files in DICOM data of the CT images, extracting each radiotherapy target area and a critical organ contour which are sketched on CT images of external irradiation plans of cervical cancer patients in training data sets, assigning the area inside the contour line to be 1, assigning the area outside the contour line to be 0, obtaining a binary mask image of the external irradiation critical organ, storing obtained CT, PTV labels, OARs labels and the like into a PNG format, wherein the PTV contour is obtained by expanding a CTV by 5mm, and the PGTV is obtained by expanding a GTV by 2 mm.

The three-dimensional dose planarization is converted into a mask in a PNG format, and the maximum dose pixel point is selected for normalization, so that the case dose distribution information is obtained, and the resolution and the size are the same as those of a CT image.

The invention has the beneficial effects that: the automatic segmentation and dose prediction model can show a better automatic segmentation effect on a test data set and simultaneously provide reasonable dose distribution reference, thereby paving a way for further realizing automatic online dose optimization of the cervical cancer adaptive radiotherapy system.

Drawings

Fig. 1 is a logic block diagram of refnentplus 3D of the present invention.

FIG. 2 is a logic Block diagram of the 3DRefine Block of the present invention.

Fig. 3 is a diagram illustrating the effect of various segmentation methods. a-c) clinical target volume in transverse, sagittal and coronal planes; d-f) organ-at-risk contours in the transverse, sagittal, and coronal planes.

FIG. 4 is a logic block diagram of the 3D RESUNet convolutional neural network of the present invention.

FIG. 5 is a comparison of DVH curves for 3DResUNet (a, c) and 3DUNet (b, d) patients.

FIG. 6 is a comparison of 3DResUNet (a, c) and 3DUNet (b, d) 3600cGy and 6000cGy patient DVH curves.

Figure 7 is a comparison of dose differences between patients 3 dresinet and 3 DUNet.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

It should be noted that all directional indicators (such as up, down, left, right, front, back \8230;) in the embodiments of the present invention are only used to explain the relative positional relationship between the components, the motion situation, etc. in a specific posture (as shown in the attached drawings), and if the specific posture is changed, the directional indicator is changed accordingly.

In the present invention, unless otherwise explicitly stated or limited, the terms "connected", "fixed", and the like are to be understood broadly, for example, "fixed" may be fixedly connected, may be detachably connected, or may be integrated; the connection can be mechanical connection or connection; they may be directly connected or indirectly connected through intervening media, or they may be interconnected within two elements or in a relationship where two elements interact with each other unless otherwise specifically limited. The specific meanings of the above terms in the present invention can be understood according to specific situations by those of ordinary skill in the art.

image pre-processing is the cleaning of image data prior to automatic segmentation training or processing of the model. For the cervical cancer CT image, because the images of different CT images are different in size and need to be input in a unified mode, image size normalization operation is firstly carried out, and the operation is beneficial to input of a deep neural network, is beneficial to accelerating the training speed and reduces waste of system resources.

wherein the RefineNetplus3D segmentation model comprises:

and performing step-by-step down-sampling operation on the original image, extracting tumor features, and then normalizing the feature map to reduce the internal covariate offset. The resolution of the output feature map is also reduced by half from one residual layer to the next. This downward sampling has two effects: first, it increases the field of experience of the convolution at deeper levels, enabling the filter to capture more contextual information, which is essential for efficient pixel classification; second, the effectiveness and ease of model training is maintained because each layer contains a large number of filters, and therefore produces an output with a corresponding number of channels, but there is a tradeoff between the number of channels and the resolution of the feature map. Typically, the final feature map output is 32 times smaller in each spatial dimension than the original image, and such low resolution feature maps lose important semantic feature details captured by early shallow filters, which may lead to a rather coarse segmentation result.

3Drefine Block a fast connection for transferring low-level features from the encoder to the decoder and proposes an efficient and versatile way of fusing coarse high-level features with fine low-level features.

The right decoder part restores the underlying features to the output image of the corresponding resolution of the original, 3DRefine Block comprising Residual Convolution Unit (RCU), chain Residual Pooling (CRP) and Fusion (Fusion). In the decoder part, it relies on two modules: residual Convolution Units (RCU) and Chain Residual Pooling (CRP), the first being a concealment layer of the original residual block without batch normalization, the second being a sequence of multiple convolution and pooling layers, but arranged by means of residual concatenation. And (3) performing 1 × 1 convolution on each path in a fusion module, then fusing 2 output characteristic graphs of the coding layer and the decoding layer, entering the next layer for further training until the required resolution is generated, and finally outputting a prediction graph by the convolution layer.

The 3DRefine Block comprises:

a residual convolution unit for generating a high-resolution predicted image by connecting and fusing a rough high-level semantic feature and a fine low-level feature through a long residual; the first part of each reinenet block consists of an adaptive convolutional layer, which is mainly used to fine-tune the pre-trained ResNet weights. For this purpose, each input path sequentially passes through two Residual Convolution Units (RCUs), and the RCUs can fuse rough high-level semantic features and fine low-level features through long residual connection to generate a high-resolution predicted image, which is helpful for training convergence.

Chain residual pooling (Chained residual pooling) for capturing context and background information of high resolution image areas; CRP is composed of multiple residual structured convolution and pooling layers, arranged in a residual manner: the purpose is to capture context and background information for high resolution image areas. It can efficiently assemble multiple windows of different sizes and perform learnable weight adjustment operations. All modules use 1 x 1 convolution and 5 x 5 pooling operators and appropriate padding operations to ensure that the space size remains the same. One pooling layer takes the output of the previous pooling layer as input, so the current pooling layer can reuse the results of the previous pooling operation to obtain features from large target areas without using large pooling operators, and finally fuse the output feature maps of all pooling blocks with the input feature maps by residual concatenation summation. The chain residual pooling of this study uses the structure of this module to facilitate gradient propagation during the training process. In addition, the application also introduces a nonlinear activation layer (ReLU) in the chain residual pooling module. ReLU is important for the effectiveness of subsequent pooling operations, which can make the model less sensitive to changes in learning rate.

The fusion, namely Multi-resolution fusion module (Multi-resolution fusion), connects the structures of the features processed by the residual convolution unit in two different paths, obtains input from the residual network and integrates the input into the high-resolution feature map, the fusion module is a structure connecting the features processed by the RCU in two different paths, obtains input from the residual network and integrates the input into the high-resolution feature map, each path is composed of 1 × 1 convolution operation and an up-sampling structure, the input feature maps are all subjected to self-adaptive operation by using convolution layers, thereby generating the same feature dimension, and the results under all the channels are summed one by one. The adaptive part input in this module helps to properly recalibrate the eigenvalues along different paths, while also facilitating subsequent fusion operations.

And a plurality of groups of RELU activation and batch normalization modules are arranged in the residual convolution unit and the chain type residual pooling and Fusion, so that the problems of gradient disappearance and gradient explosion in RCU, CRP and Fusion are solved.

The ReLU (Rectified linear unit) activation function is also called a linear rectification unit and has the formula of f (x) = max (0, x), and the ReLU function is characterized in that the part of the ReLU function larger than zero is linear, the derivative is constant, the ReLU function can propagate in the reverse direction without the problem of gradient explosion, and the gradient can be linearly transferred in a multilayer network. When the input is a negative value, the output is 0, and when the input is a positive value, the output is the input. The ReLU function has high convergence efficiency, the calculated amount of the gradient is obviously lower than that of a Sigmod activation function and a tanh function no matter the forward propagation or the backward propagation, and in addition, the gradient saturation phenomenon of the function at the position of the origin does not occur like the Sigmod activation function.

The convolution layer adopts a 3D convolution layer, three-dimensional convolution applies a three-dimensional convolution kernel to the data set, and the convolution kernel moves towards 3 directions X, Y and Z to calculate low-level features. The output shape is a three-dimensional volume space, such as a cube or a cuboid, three-dimensional convolution training is beneficial to detection of target objects such as videos and three-dimensional medical images, the calculation principle of a 3D convolution layer is the same as that of a 2D convolution layer, and only one dimension is added to the input image and the convolution kernel. Taking an input image with the size of S W H C and a convolution kernel W with the size of F C K as an example, wherein S, W, H and C represent the number of slices, width, height and channels of the input image; f, F, F, C and K represent the slice number, width, height, channel number and number of the convolution kernel, the channel number of the input image should be consistent with the channel number of the convolution kernel, and the relationship of the number should satisfy that F is smaller than W, H and S.

And the initial sampling rate is 1/4, but this may result in loss of image details, so a downsampling layer with a sampling rate of 1/2 and corresponding 3 drefinieblock are added to the first layer of the RefineNetPlus 3D. The problem of feature loss in the network training process is solved by adopting a scheme that a convolution filtering operator is increased along with the increase of the number of convolution layers, and finally all features can be reserved to the maximum extent.

The trained model is put in a test set of cervical cancer, and the quality of the model is evaluated through the performance of the trained model on the test set. Commonly used evaluation indices in medical images are the dess coefficient Dice, the Jaccard coefficient and the Average Surface Symmetry Distance (ASSD). The specific index parameters of the model predictions are shown in the following table.

CTV, clinical target area; OARs are organs at risk; SI, small intestine, FR, right femur, FL, left femur, JSC, jaccard similarity coeffient; DSC, dice similarity coeffient; ASSD: average systematic surface distance.

It can be seen that our improved model of refinetplus 3D achieves the highest Dice values in the segmentation of cervical cancer CTV and OARs, while the segmentation takes the least time. The average Dice of RefinenPlus 3D in CTV is 0.82, JSC is 0.69, ASSD is 2.13mm, and the average segmentation time is 6.4s; average Dice of UNet3D was 0.80, jsc was 0.67, assd was 3.56mm, and average split time was 9.8s; resUNet3D has an average Dice of 0.81, JSC of 0.69, ASSD of 3.46mm, and an average split time of 11.4s. The results show that RefineNetPlus3D increased 2% and 1% over UNet3D and reset 3D, respectively, and that the time to segmentation for individual patients decreased by 3.4s and 5s.

For segmentation of cervical cancer organs-at-risk, the mean Dice for UNet3D, resUNet3D, and refinenplus 3D are: urinary bladder 0.96, 0.97; small intestine: 0.93, 0.95; the right femoral head; 0.97, 0.98; left femoral head: 0.97, 0.98; rectum: 0.88,0.91,0.91. The average JSCs were: bladder: 0.93, 0.94; small intestine: 0.88, 0.90; the right femoral head; 0.94, 0.96; left femoral head: 0.95, 0.96; rectum: 0.78,0.84,0.84. The average ASSDs are respectively: 0.59,0.48,0.30 of bladder; small intestine: 1.68,1.45,1.02; right femoral head: 0.34,0.23,0.16; left femoral head: 0.29,0.20,0.15; rectum: 1.37,0.92,0.61. The average segmentation times for the three models in the five organs at risk are: refineNetPlus3D:6.6s; UNet3D:10.3s; resUNet3D:11s, respectively.

FIG. 3 illustrates an example of the results of the segmentation of a test set during an experiment, a-c) clinical target volume in transverse, sagittal and coronal planes; d-f) organ-at-risk contours of the transverse, sagittal, and coronal planes; where yellow represents the manual delineation profile, purple represents reflonenetplus 3D, blue represents 3 dresinet, and green represents the 3DUNet profile. From this figure, it can be seen that the segmentation performance of refinenplus 3D is relatively satisfactory, and a small volume such as a rectal region can be segmented more accurately, and for the segmentation of a small intestine, the refinenplus 3D proposed by the present study also has a relatively ideal effect.

The segmentation result shows that the performance of the 3D model provided by the application has obvious advantages. Particularly for the segmentation of small intestine and rectum, the model can take the least time while achieving higher segmentation effect.

The invention also discloses a self-adaptive radiotherapy plan automatic generation method based on the cervical cancer target area automatic delineation method, which comprises the following steps:

the 3 dresinet convolutional neural network comprises:

The method has the advantages that 3DUNet keeps the resolution ratio of the high-level semantic feature graph restored to the original graph through the symmetrical part of the network, and meanwhile, residual error network ResidualNet is adopted to realize jump connection of the convolutional layer, so that gradient back propagation is prevented from disappearing when weight is updated and optimized; in addition, because the anatomical structures of the target area and the OARs, the CT image and the like are three-dimensional structures, the regression fitting error can be further reduced by theoretically adopting the three-dimensional neural network model, and the prediction precision of three-dimensional dose distribution is improved. The structure diagram of the model is shown in the figure below, the network is totally divided into 4 layers and consists of an encoder (left) and a decoder (right), the encoder extracts multi-level and multi-scale three-dimensional characteristics of three-dimensional matrixes such as CT images, organs at risk, target structures, RTDose and the like, and the decoder realizes regression fitting of the three-dimensional characteristics to dose distribution. The encoder contains 2 3D convolutional layers of 3 × 3 × 3 and 2 × 2 × 2 steps in each layer; after the data of the previous layer passes through the pooling layer with the size of 3 multiplied by 3 and the step length of 2 multiplied by 2, the size is halved, the number of channels is increased by 2, then the input of the next layer is made, and the following 3D convolution operation is continued. Each coding module has 2 convolution layers in total, the size of the three-dimensional convolution kernel is 3 multiplied by 3, the step length is 2 multiplied by 2, and the size is multiplied after the transposition matrix, and the number of channels is unchanged. In order to reduce the loss of high resolution features in the up-sampling path, data of the same size in the down-sampling path is connected to the data in the up-sampling path in a jump connection manner. The concatenated data will have the same size as the up-sampled path data, but the number of channels is the size of the number of channels of the down-sampled path. After passing through a convolution kernel of 3 × 3 × 3 size and a 3D convolution layer of 1 × 1 × 1 step size, the data is used as input data of the previous layer, and then the transposed convolution operation is performed again. To increase the robustness of the model, a batch normalization and activation function is added to each convolution layer. To automatically predict the three-dimensional dose, the last layer of the model performs nearest neighbor interpolation (nearest) on the data, restoring to a size consistent with the original input image.

the training database grouped 254 cervical cancer patients who received Volume Modulated Arc Therapy (VMAT) technology treatment in my hospital from 2018 to 2021, with the age ranging from 21-80 years and the median age of 55 years, with most patients diagnosed with squamous cell carcinoma. The detailed characteristics of the enrolled patients are shown in the table below. All cervical cancer patients were planned using the Monoca Treatment Planning System (TPS) and radiotherapy was applied using an Elekta linac with a reconstructed layer thickness spacing of 3mm for cervical cancer CT images. Each patient's clinical target volume CTV is outlined by a radiation oncologist, then outlined according to international guidelines RTOG18, and finally reviewed by an advanced radiation oncologist, evenly expanding the CTV5mm to form a Planning Target Volume (PTV) taking into account the uncertainty and organ motion effects of the patient when positioned. All cervical cancer patients in the training set received radiation therapy at a prescribed dose of 50Gy/25 or 45Gy/25, and all treatment plans were obtained by a physicist adjusting the objective function on the Monoca radiation therapy planning System, optimizing the parameters, using 6MV X-rays, and then clinically approved by a senior radiation oncologist with over 10 years of radiation therapy experience on a clinical routine basis.

The Mann-Whitney U test was used for continuous variable analysis; fisher's test is used for multi-class variable analysis

CT images of the patient, radiation Therapy Structures (RTST) including PTV contours, and dose information of the patient are derived in a format of. Dcm and read and operated using an open source software package pydicom [69]. Besides CT images, PTV contours and contour information of Organs At Risk (OARs) are extracted from RTST files of cervical cancer data and used as input of a cervical cancer dose prediction model, wherein the PTV5000 and the PTV4500 are respectively used as input of a cervical cancer dose prediction model, the OARs comprise bladders, rectum, small intestine, left thighbone heads and right thighbone heads, the PTV contours are sketched by experienced clinicians, and numbers behind the PTV represent prescription doses (unit: cGy) which a target area should receive.

Considering that the extracted dose grid, namely the resolution, is not matched with the resolution of the CT image and the PTV outline, the dose points on the dose grid are linearly interpolated to pixel points on the CT image, the CT image and the interpolated dose graph are normalized to 256 multiplied by 256 pixels, and then the normalized dose graph is input into a deep learning model for training. After the clinical data are collected, the clinical data are firstly processed, the processing method is that MATLAB is used for reading and converting basic data (such as DICOM-RT analysis, dose matrix generation, structure matrix generation and the like), a scientific computation package Numpy is used for processing image information, and PTV contour information is converted into a Numpy matrix [70] only containing 0 and 1. In summary, the input layers of the cervical cancer dose prediction model include 1 image of CT image information, one or two images containing planning target volume contour information, and 5 images containing patient body contours. The interpolated three-dimensional dose distribution information is used for Loss function (Loss) calculation during back propagation.

In deep learning model training, good data image quality plays a great role in model training, so that the model can be rapidly converged and the precision of the model can be improved by carrying out standardized preprocessing on training set data.

The preprocessing operation steps for the data in the application are as follows:

firstly, labels (Label) are required to be made, the labels are read from RTstructure files in DICOM data of CT images, the outlines corresponding to names are extracted, each radiotherapy target area and an organ-at-risk outline sketched on an external irradiation plan CT image of a cervical cancer patient in a training data set are extracted, areas inside the outlines are assigned to be 1, areas outside the outlines are assigned to be 0, a binary mask image of the external irradiation organ-at-risk is obtained, finally, the obtained CT, PTV labels, OARs labels and the like are stored in a PNG format, and subsequent model training is facilitated, wherein the PTV outline is obtained by expanding a CTV by 5mm, and the PGTV is obtained by expanding a GTV by 2 mm. On the other hand, for the dose distribution information of the radiotherapy plan, which is stored in the RTDose file of DICOM, the study converts the three-dimensional dose planarization into a PNG-format mask through a self-writing RTDose reading program, and selects the maximum dose pixel point for normalization, and the resolution and size of the dose matrix are the same as those of the CT image, and then the dose matrix is sent to a deep learning model for training.

4) Inputting a to-be-predicted image into a trained 3DRESUNet convolutional neural network to obtain the predicted three-dimensional dose distribution of the target area, wherein the to-be-predicted image comprises 1 image of CT image information, one or two images containing planning target area contour information and 5 images containing the body contour of a patient.

Dose matrix data obtained by predicting clinical planned doses and ResUNet3D and UNet3D models of the test set are read by an MATLAB dose conversion program and drawn into DVH curves, and 2 cervical cancer cases are randomly selected, as shown in FIG. 5. For PTV and PGTV, the DVH curves are essentially identical, but some cases prescribed at 4500cGy are higher in the high dose areas. The DVH curves for the small intestine and rectum vary greatly in different cases, with the DVH curves for the remaining OARs varying less in different test set cases; the DVH of PGTV and PTV of selected cases was less variable compared to the original clinical plan.

The following table summarizes the mean and standard deviation values for specific clinical indices of PTV and OAR for cervical cancer patients, with a mean dose (in Gy) of 0.59 + -0.44 for ResUNet3D PGTV, 0.53 + -0.45 for D2 (in Gy), and 1.30 + -0.88 for D98 (in Gy) in the test set samples. While the average dose (Gy) of 3DUNetPGTV was 0.97. + -. 0.57, D2 (Gy) was 0.83. + -. 0.45 and D98 (Gy) was 1.08. + -. 0.64.ResUNet3D predicts PTV more accurately than UNet3D with a mean dose (unit: gy) of 0.94 + -0.86, a D2 (unit: gy) of 1.04 + -1.10, and a D98 (unit: gy) of 1.27 + -1.67. On the other hand, for organ-at-risk dose prediction, resUNet3D is slightly larger than the prediction of UNet3D, but all within a clinically acceptable range. The average doses (unit: gy) of ResUNet3D bladder, rectum, small intestine and left and right femoral head are respectively 2.43 + -1.92, 1.98 + -1.83, 1.74 + -3.27, 2.70 + -2.36 and 1.87 + -1.91; d2 (Gy) of the bladder, rectum and left and right femoral heads is 1.91 +/-3.29, 0.98 +/-1.19, 2.14 +/-1.91 and 2.05 +/-1.92 respectively; d98 (Gy) of the bladder, rectum and left and right femoral heads is respectively 4.88 +/-3.99, 4.88 +/-4.13, 3.38 +/-3.09 and 2.59 +/-3.56; the maximum dose deviation (unit: gy) of the small intestine was 3.19. + -. 3.48.

Notes:Dx％:dose received by≥x％of the objective structure；SI:small intestine.FR:right

femoral head.FL:left femoral head.|δ _Dose |＝|D _prediction -D _{grond truth} |.

DVH prediction also benefits from all spatial information contained in the regression system, since the deep learning framework can account for different proximity to multiple PTVs and OARs simultaneously. As can be seen from the above table, the DVH curves for PTV and OARs have acceptable agreement between clinical and predicted outcomes for patients a and B. We found that these results are very comparable, except for a few indices highlighted in the table. Compared with the manual optimization scheme, the absolute value difference of the average value of each index in the Pred scheme is relatively small, and the feasibility of the method is proved. The DVH curves also met the clinical criteria for patients with prescribed doses of 3600cGy and 6000cGy, as shown in figure 6.

Figure 7 is a dose distribution difference (clinical planned dose minus predicted dose) between the clinically planned dose distribution at the transaxial level and the dose distribution predicted by two deep learning models for 2 randomly selected cases of the VMAT test set of cervical carcinomas. From the dose discrepancy profile, the dose discrepancy profile positions are about the same, but the dose discrepancy map shows that the dose predicted by the ResuNet3D model is closer to the clinically planned dose distribution than UNet 3D. On the other hand, the results show that UNet3D predicts a smoother shape of the dose distribution with the TPS plan at each dose level.

The invention realizes the VMAT planning dose distribution prediction of the cervical cancer by utilizing the 3DRESUNet model to automatically learn the multi-scale and multi-level three-dimensional characteristics in the CT image, the target area and the OARs structure of the cervical cancer. Because the dose distribution is three-dimensional and full-space and is closely related to three-dimensional anatomical structure characteristics, theoretically, the 3D ResUNet model of the research can automatically extract more characteristics than other two-dimensional CNN models, and the dose prediction of the cervical cancer VMAT plan is more accurate.

The examples should not be construed as limiting the present invention, but any modifications made based on the spirit of the present invention should be within the scope of protection of the present invention.

Claims

1. An automatic delineation method for a cervical cancer target area is characterized in that: which comprises the following steps:

2) Performing target area segmentation on the input image based on a RefinenPlus 3D segmentation model,

wherein the RefineNetplus3D segmentation model comprises:

the 3DRefine Block comprises:

chain residual pooling for capturing context and background information of the high resolution image area;

2. The automatic cervical cancer target zone delineation method of claim 1, wherein: the sampling rate of the first layer residual network is 1/2.

3. The cervical cancer target zone automatic delineation method of claim 1, characterized in that: the volume block is a 3 x 3D volume block.

4. The cervical cancer target zone automatic delineation method of claim 3, wherein: the chain residual pooling consists of a plurality of residual structure convolution and pooling layers, arranged in a residual manner.

5. An adaptive radiotherapy plan automatic generation method based on the cervical cancer target area automatic delineation method of any one of the above claims 1 to 4, characterized in that: which comprises the following steps:

6. The method of automatic generation of an adaptive radiotherapy plan of claim 5, wherein: the 3DResUNet convolutional neural network includes:

a right decoder for performing regression fitting of three-dimensional features to dose distributions, comprising 4 layers of sequentially cascaded residual networks, each residual network comprising an upsampling convolution module,

7. The method of automatic generation of an adaptive radiotherapy plan of claim 5, wherein: the CT image of the patient in the training database comprises a radiotherapy structure of a PTV contour and dose information of the patient, and meanwhile, the contour information of the PTV contour and the contour information of an organ at risk are extracted from an RTST file of cervical cancer data and used as the input of a cervical cancer dose prediction model, the PTV contour is sketched by an experienced clinician, and the number behind the PTV represents the prescription dose which a target area should receive.

8. The method of automatic generation of an adaptive radiotherapy plan according to claim 7, wherein: the method comprises the steps of preparing labels for CT images to be preprocessed, wherein the labels are read from contours corresponding to names in RTStructure files in DICOM data of the CT images, extracting each radiotherapy target area and a critical organ contour which are sketched on CT images of external irradiation plans of cervical cancer patients in training data sets, assigning the area inside the contour line to be 1, assigning the area outside the contour line to be 0, obtaining a binary mask image of the external irradiation critical organ, storing obtained CT, PTV labels, OARs labels and the like into a PNG format, wherein the PTV contour is obtained by expanding a CTV by 5mm, and the PGTV is obtained by expanding a GTV by 2 mm.

9. The method of automatic generation of an adaptive radiotherapy plan of claim 8, wherein: the three-dimensional dose planarization is converted into a mask in a PNG format, and the maximum dose pixel point is selected for normalization, so that the case dose distribution information is obtained, and the resolution and the size are the same as those of a CT image.