Intelligent delineation method for radiotherapy target volume and organs at risk of cervical cancer

Abstract

The application discloses an intelligent delineation method for a radiotherapy target region and a critical organ of cervical cancer, relates to the technical field of medical images, and comprises the following steps: acquiring CT and MRI image data sets of a cervical cancer patient; inputting the CT and MRI images containing metal artifacts into a pre-trained metal artifact removal model to obtain CT and MRI images from which metal artifacts are removed; inputting the CT and MRI images from which metal artifacts are removed into a double-encoder U-Net network containing a cross-modality attention fusion gating module to obtain final fusion features; processing the final fusion features through a U-Net decoder, constructing a small sample adaptive hybrid loss function to optimize network parameters, and outputting a segmentation mask; and performing contour processing on the segmentation mask to complete delineation of the target region and the critical organ. Through automatic delineation with high precision and high efficiency, the application can enable patients to obtain higher quality and higher efficiency of treatment, and meanwhile, the working intensity of doctors can be reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical images, in particular to an intelligent delineation method for a radiotherapy target area and an organ at risk of cervical cancer. BACKGROUND

[0002] Cervical cancer is the most common tumor disease in the female reproductive system, and is sensitive to radiotherapy in clinical practice, and most can be cured by radiotherapy. Radiotherapy for cervical cancer includes external beam radiotherapy and brachytherapy, and brachytherapy has entered the three-dimensional era from the two-dimensional era. In precise three-dimensional brachytherapy for tumors, standard delineation of target areas and organs at risk, removal of source metal artifacts, and dose prediction are crucial.

[0003] However, clinical practice mainly relies on manual delineation by doctors, which has the problems of low efficiency, strong subjectivity, and large differences between different doctors. Although there are automatic delineation methods based on deep learning, there are limitations, such as the use of applicators and other metal implants in cervical cancer brachytherapy, which produce serious stripe artifacts on CT images, seriously affecting the identification and delineation accuracy of target area and organ at risk boundaries. Although CT and MRI images have complementary advantages in soft tissue contrast, existing methods lack effective multi-modal information fusion mechanisms and cannot fully utilize the geometric accuracy of CT and the soft tissue resolution of MRI.

[0004] A method and system for automatic delineation of high-risk clinical target volume and organs at risk in cervical cancer brachytherapy based on deep learning are disclosed in Chinese Patent No. CN120198387A, which is applied to the field of data processing. The CT image information and MR image information of the cervical cancer patient are processed based on the positioning network to generate 3D position information and shape information of the applicator, as well as image feature sequence information. The CT image information of the cervical cancer patient is classified and processed based on the positioning network to generate prior knowledge labels. The prior knowledge labels are processed based on the segmentation network to generate applicator positioning target factors. The applicator positioning target factors, 3D position information and shape information of the applicator, and image feature sequence information are processed based on the segmentation network to generate high-risk clinical target volume and organ segmentation parameter information. The high-risk clinical target volume and organ segmentation parameter information are processed to generate automatic delineation results of the high-risk clinical target volume and organs at risk in cervical cancer brachytherapy. SUMMARY

[0005] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide an intelligent delineation method for a radiotherapy target area and an organ at risk of cervical cancer.

[0006] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0007] The intelligent delineation method for the target region and the organ at risk of cervical cancer radiotherapy comprises the following steps:

[0008] Step S1, acquiring CT and MRI image data sets of a cervical cancer patient;

[0009] Step S2, inputting the CT and MRI images containing metal artifacts into a pre-trained metal artifact removal model to obtain CT and MRI images with metal artifacts removed;

[0010] Step S3, inputting the CT and MRI images with metal artifacts removed into a dual-encoder U-Net network containing a cross-modal attention fusion gating module to obtain final fusion features;

[0011] Step S4, processing the final fusion features through a U-Net decoder, constructing a small sample adaptive hybrid loss function to optimize network parameters, and outputting a segmentation mask;

[0012] Step S5, performing contour processing on the segmentation mask to complete the delineation of the target region and the organ at risk.

[0013] Further, the step S1 specifically comprises:

[0014] Based on the expert consensus on the delineation standards of the organs at risk and the target region of three-dimensional brachytherapy for cervical cancer, the recommended delineation standards of RTOG are referred to, and a standard data set of 300 patients with advanced cervical cancer is established, which includes CT and MRI multi-modal images before and after metal artifact removal, and a standard data set manually delineated by a radiotherapy specialist.

[0015] Further, in the step S2, the pre-trained metal artifact removal model is based on the CycleGAN architecture, the generator adopts a U-Net structure with residual connection containing 8 layers of encoder and 8 layers of decoder, and the discriminator adopts a 5-layer convolution PatchGAN structure.

[0016] The CycleGAN architecture adopts a coupled dual CycleGAN architecture to process the metal artifact removal of CT and MRI modalities, wherein the generator and the discriminator parameters of the CT path and the MRI path are independent, and are jointly optimized through a dual-path consistency joint loss function.

[0017] Further, in the pre-trained metal artifact removal model, the dual-path consistency joint loss function includes an adversarial loss term, a cycle consistency loss term, and a semantic consistency loss term.

[0018] Further, in the step S3, the cross-modality attention fusion gating module is embedded after each down-sampling layer of the U-Net encoder, and is constructed as a U-Net variant with double encoders and a single decoder, wherein the double encoders include an encoder CT for processing the de-artifacted CT image and an encoder MRI for processing the de-artifacted MRI image.

[0019] Further, the cross-modality attention fusion gating module realizes adaptive fusion of CT and MRI features by calculating a cross-attention map and dynamic fusion weights, and specifically includes the following steps:

[0020] respectively calculating a cross-attention map from CT features to MRI features and a cross-attention map from MRI features to CT features;

[0021] modulating the original features using the cross-attention map to obtain modulated features;

[0022] splicing the modulated CT and MRI features, and generating dynamic fusion weights through a learnable weight network;

[0023] performing weighted summation on the CT and MRI features using the dynamic fusion weights to obtain final fusion features.

[0024] Further, the U-Net decoder uses a ResNet-50 network pre-trained on an ImageNet dataset as a basic architecture, removes the original fully connected layer and average pooling layer of the ResNet-50, and redesigns the four convolution stages of the ResNet-50, i.e., conv2_x to conv5_x, as the up-sampling path of the decoder, with 2 times up-sampling performed in each stage in turn, and finally restores to the original image resolution.

[0025] Further, in the step S4, the small sample adaptive hybrid loss function is obtained by weighted combination of a generalized Dice loss function, a weighted cross-entropy loss function and a Focal loss function, wherein the generalized Dice loss function is used to optimize the overlap of the segmentation region and handle the class imbalance problem, the weighted cross-entropy loss function is used to improve the pixel-level classification accuracy, and the Focal loss function is used to focus on the training of the classification samples, and the weight coefficients of the three loss functions are set to 0.6, 0.3 and 0.1 respectively.

[0026] Further, in the step S5, the contour processing of the segmentation mask uses a contour tracking algorithm, and specifically includes the following steps:

[0027] convert the segmentation probability map of each class into a binary mask, and set a fixed threshold to distinguish the target region from the background;

[0028] An edge tracking algorithm is used to sequentially find the boundary pixel points of each connected region to generate an initial contour line;

[0029] The contour point sequence of the initial contour line is simplified to remove redundant points and retain final contour points;

[0030] The final contour point coordinates are converted from the image pixel coordinate system to the patient three-dimensional physical coordinate system;

[0031] The target region and the delineation structure file of the organ at risk are generated according to the medical standard format.

[0032] Further, the contour tracking algorithm further includes a topological structure checking and correcting step, specifically including:

[0033] Detect whether the generated polygon contour has self-intersection;

[0034] Re-triangulate the contour with self-intersection;

[0035] Distinguish the outer contour and the inner hole contour, and ensure that the outer contour is sorted in counterclockwise direction and the inner hole contour is sorted in clockwise direction;

[0036] Establish a hierarchical relationship tree of the contour to accurately represent the inclusion relationship between organ structures;

[0037] Smooth the corrected contour, and use a 3rd order B-spline curve interpolation to obtain a smooth boundary conforming to the anatomical structure.

[0038] Compared with the prior art, the present application has the following advantages:

[0039] 1、The double-path metal artifact removal model based on the CycleGAN architecture of the present application combines a double-path consistency joint loss function, which not only effectively removes metal artifacts in CT and MRI images, but also ensures that the multi-modal images after artifact removal remain consistent in anatomical structure.

[0040] 2、The cross-modal attention fusion gating mechanism is adopted in the present application, which dynamically calculates the fusion weights of CT and MRI features at multiple scales of the U-Net encoder, and can adaptively adjust the contribution of the two modalities according to the local image content, such as enhancing the CT feature weight in the bone region and enhancing the MRI feature weight in the soft tissue region, to take advantage of the complementary advantages of multi-modal images.

[0041] 3、The present application designs a double-encoder path to process CT and MRI images respectively, so that the network can fully learn and retain the electron density information, geometric accuracy of CT images, and superior soft tissue contrast and tumor display capability of MRI images, avoiding the problem of feature confusion or loss of modality-specific information caused by early fusion.

[0042] 4. The fully automated sketching process of this invention reduces the sketching time for a single patient from several hours in the traditional manual method to minutes, while eliminating sketching differences between different doctors. It provides reliable technical support for standardized and regulated cervical cancer radiotherapy and has important clinical application value. Attached Figure Description

[0043] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0044] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the metal artifact removal model structure according to an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the dual encoder U-Net structure according to an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of the drawing result in an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] like Figure 1 As shown, the intelligent delineation method for cervical cancer radiotherapy target areas and organs at risk includes the following steps:

[0050] Step S1: Obtain the CT and MRI image dataset of cervical cancer patients;

[0051] Step S2: Input the CT and MRI images containing metal artifacts into the pre-trained metal artifact removal model to obtain CT and MRI images with metal artifacts removed.

[0052] Step S3: Input the CT and MRI images with metal artifacts removed into a dual-encoder U-Net network containing a cross-modal attention fusion gating module to obtain the final fused features;

[0053] Step S4: Process the final fused features through the U-Net decoder, construct a few-sample adaptive mixing loss function to optimize network parameters, and output a segmentation mask;

[0054] Step S5: Perform contour processing on the segmentation mask to complete the delineation of the target area and organs at risk.

[0055] The CT and MRI image data set in the step S1 specifically includes: CT and MRI images containing metal artifacts, CT and MRI multi-modal images after removing metal artifacts, and standard data sets manually outlined by oncology radiotherapy specialists.

[0056] The training data is specifically an expert consensus based on a three-dimensional brachytherapy organ at risk and target area delineation standard for cervical cancer, and the RTOG recommended delineation standard is referred to, a standardized training data set containing 300 cases of locally advanced cervical cancer patients is constructed, and the data set covers the following data:

[0057] 1. original CT+MRI image data containing metal artifacts;

[0058] 2. CT+MRI multi-modal images after metal artifact removal;

[0059] 3. manual delineation results completed by experienced oncology radiotherapy specialists in accordance with the delineation standard;

[0060] Based on the constructed standard training data set, the following strategies are adopted in the training process:

[0061] The data set is divided into a training set, a validation set and a test set in a ratio of 7:2:1;

[0062] The model is trained using the same training parameters and optimization strategies.

[0063] The pre-trained metal artifact removal model in the step S2 is based on the CycleGAN architecture, the generator adopts the U-Net structure containing 8 layers of encoder and 8 layers of decoder, and the discriminator adopts the PatchGAN structure with 5 layers of convolution;

[0064] The CycleGAN architecture adopts a coupled double CycleGAN architecture to process metal artifact removal of CT and MRI modalities, wherein the generator and discriminator parameters of the CT path and the MRI path are independent, and are jointly optimized through a double-path consistency loss function.

[0065] The generator architecture specifically includes:

[0066] The U-Net input and output are 512x512x1;

[0067] The encoder has 8 convolutional layers with channel numbers [64, 128, 256, 512, 512, 512, 512, 512];

[0068] The decoder has 8 deconvolutional layers with channel numbers [512, 512, 512, 512, 256, 128, 64, 1];

[0069] The encoder outputs of each layer are spliced with the decoder input of the corresponding layer using a skip connection;

[0070] The encoder activation function uses LeakyReLU(0.2), the decoder activation function uses ReLU, and the output activation function uses TanH.

[0071] The discriminator architecture specifically includes:

[0072] The PatchGAN includes 5 convolutional layers;

[0073] The number of channels is [64, 128, 256, 512, 1];

[0074] The convolution kernel is set to 4x4, and the step is 2;

[0075] The activation function is LeakyReLU(0.2);

[0076] The output is a 30x30 discriminant matrix.

[0077] In the pre-trained metal artifact reduction model, the specific formula of the dual-path consistency joint loss function is:

[0078]

[0079] wherein, represents the dual-path consistency joint loss function, represents the adversarial loss term, represents the cycle consistency loss term, represents the semantic consistency loss term, , and respectively represent the corresponding loss weight coefficients.

[0080] wherein, the specific formula of the adversarial loss is:

[0081]

[0082] wherein, represents the discrimination result of the discriminator on the real image, that is, the probability value output by the discriminator network on the image sampled from the real data distribution without metal artifacts, represents the discrimination result of the discriminator on the generated image, that is, the probability value output by the discriminator network on the de-artifact image generated by the generator, represents the expected value operator;

[0083] wherein, the discriminator network outputs a 30x30 matrix, and each element represents the probability that a local region in the image is real;

[0084] The specific formula of the cycle consistency loss is:

[0085]

[0086] wherein, denotes the generator from CT domain to MRI domain, i.e. converting CT images with metal artifacts to artifact-free MRI images, denotes the generator from MRI domain to CT domain, i.e. converting MRI images with metal artifacts to artifact-free CT images, denotes the input of CT images with metal artifacts, denotes the input of MRI images with metal artifacts, denotes the L1 norm, i.e. the sum of absolute values, which is less sensitive to outliers than the L2 norm, can produce clearer images, and the cycle consistency loss ensures that the images can be reconstructed back to the original images after double conversion of CT→MRI→CT and MRI→CT→MRI, preventing the generator from producing arbitrary unrelated outputs and maintaining the semantic consistency of image content;

[0087] The calculation formula of the semantic consistency loss is:

[0088]

[0089] wherein, denotes the number of samples in the batch, denotes the i-th sample in the batch, denotes the CT image after metal artifact reduction of the i-th sample, denotes the MRI image after metal artifact reduction of the i-th sample, denotes the pre-trained CT image feature extraction network, denotes the pre-trained MRI image feature extraction network, denotes the square of the L2 norm, denotes the structural loss weight coefficient, denotes the structural similarity index, denotes the learnable grayscale transformation on the MRI image, denotes the learned linear transformation matrix, which is used to roughly align the grayscale space of the MRI image to the grayscale space of the CT image.

[0090] The pre-trained feature extraction network and The VGG-19 network pre-trained on the ImageNet dataset is adopted, the fully connected layers are removed, and the output of the conv4_3 layer is used as the feature representation; the value range of the structural loss weight coefficient is 0.1-1.0; the learned linear transformation matrix The size of the 2x2 is optimized together with other weights during the training process as network parameters.

[0091] In the step S3, the cross-modal attention fusion gating module is embedded after each down-sampling layer of the U-Net encoder, and is constructed as a U-Net variant with double encoders and single decoder, the double encoders including: an encoder CT for processing the de-artifacted CT image; and an encoder MRI for processing the de-artifacted MRI image.

[0092] The cross-modal attention fusion gating module realizes adaptive fusion of CT and MRI features by calculating cross-attention maps and dynamic fusion weights, and specifically includes the following steps:

[0093] The cross-attention map from the CT feature to the MRI feature and the cross-attention map from the MRI feature to the CT feature are calculated respectively;

[0094] Taking the cross-attention map from the MRI feature to the CT feature as an example, the specific calculation formula is:

[0095]

[0096] wherein, represents the cross-attention map from the MRI feature to the CT feature, , , represents a query vector matrix obtained by linear transformation from the CT feature, represents a feature map from the CT encoder, represents a learnable query transformation weight matrix, represents a key vector matrix obtained by linear transformation from the MRI feature, represents a feature map from the MRI encoder, represents a learnable key transformation weight matrix, represents a feature dimension of the query and key vectors;

[0097] represents the cross-attention map from the MRI feature to the CT feature, the CT feature as the query and the MRI feature as the key, through similarity calculation and normalization, information of which MRI positions should be paid attention to by each CT position is obtained; for modulating the CT feature, so that the CT feature is enhanced in the important information area provided by the MRI.

[0098] Similarly, the cross-attention map from the CT feature to the MRI feature is calculated simultaneously, represents the cross-attention map from the CT feature to the MRI feature, for modulating the MRI feature.

[0099] The original features are modulated by using a cross-attention map to obtain modulated features, and the specific formula is:

[0100]

[0101]

[0102] wherein, represents the modulated CT feature, represents the modulated MRI feature, and represents the CT and MRI features from the same level, represents the cross-attention map from the CT feature to the MRI feature, represents the cross-attention map from the MRI feature to the CT feature, represents element-wise multiplication, i.e. Hadamard product;

[0103] The modulated CT and MRI features are spliced, and a dynamic fusion weight is generated through a learnable weight network, and the calculation formula is:

[0104]

[0105] wherein, represents the dynamically generated CT feature importance weight, represents a Sigmoid activation function, and represents a learnable weight matrix and a bias vector, represents the modulated CT feature, represents the modulated MRI feature, represents a splicing operation in the channel dimension;

[0106] The CT and MRI features are weighted and summed using the dynamic fusion weight to obtain the final fusion feature, and the calculation formula is:

[0107]

[0108] wherein, represents the final fusion feature, and represents the CT and MRI features from the same level.

[0109] The U-Net decoder uses a ResNet-50 network pre-trained on an ImageNet dataset as a basic architecture, removes the original fully connected layer and average pooling layer of ResNet-50, and redesigns the four convolution stages of ResNet-50, i.e., conv2_x to conv5_x, as the up-sampling path of the decoder, each stage being sequentially 2 times up-sampled, and finally recovering to the original image resolution.

[0110] The U-Net decoder uses a ResNet-50 network pre-trained on an ImageNet dataset as a basic architecture, and makes the following improvements:

[0111] 1. Network structure adjustment:

[0112] The original fully connected layer and average pooling layer of ResNet-50 are removed, and the four convolution stages of ResNet-50 (conv2_x to conv5_x) are redesigned as the up-sampling path of the decoder, each stage being sequentially 2 times up-sampled, and finally recovering to the original image resolution.

[0113] 2. Up-sampling module design:

[0114] Each up-sampling stage contains bilinear interpolation up-sampling -> 3x3 convolution -> batch normalization -> ReLU activation, the up-sampling rate sequence is [2, 2, 2, 2], and the corresponding feature map size changes: 16x16 -> 32x32 -> 64x64 -> 128x128 -> 256x256, and the channel number sequence is [1024, 512, 256, 128, 64].

[0115] 3. Integration of skip connection:

[0116] The corresponding level features from the encoder are first adjusted in channel number by 1x1 convolution, the adjusted encoder features are spliced with the up-sampled features in channel dimension, and the spliced features are fused and information integrated by 3x3 convolution.

[0117] In the step S4, the small sample adaptive hybrid loss function is obtained by weighted combination of a generalized Dice loss function, a weighted cross-entropy loss function and a Focal loss function, wherein the generalized Dice loss function is used to optimize the overlap of the segmentation region and handle the class imbalance problem, the weighted cross-entropy loss function is used to improve the pixel-level classification accuracy, and the Focal loss function is used to focus on the training of the classification samples, and the weight coefficients of the three losses are set to 0.6, 0.3 and 0.1 respectively. This combination fully utilizes the complementary advantages of each loss function, effectively improves the generalization ability of the model under small sample conditions while ensuring the segmentation accuracy.

[0118] In the field of deep learning-based medical image segmentation, in order to deal with common problems such as class imbalance and edge blur, there are various classic loss functions, such as generalized Dice loss, which is mainly used to deal with sample imbalance, especially when the target region (such as tumor) and the background region have significant size difference; weighted cross-entropy loss deals with class imbalance by assigning weights to different classes; focal loss is used to solve the problem of difficult and easy sample imbalance.

[0119] In the step S5, the contour processing on the segmentation mask uses a contour tracking algorithm, specifically including the following steps:

[0120] Convert the segmentation probability map of each class to a binary mask, and set a fixed threshold to distinguish the target region from the background;

[0121] Use a boundary tracking algorithm to find the boundary pixel points of each connected region in turn, and generate an initial contour line;

[0122] Polygon simplification is performed on the contour point sequence of the initial contour line, redundant points are removed, and final contour points are retained;

[0123] Convert the final contour point coordinates from the image pixel coordinate system to the patient three-dimensional physical coordinate system;

[0124] Generate a delineation structure file of the target region and the organ at risk according to the medical standard format.

[0125] The contour tracking algorithm further includes a topology checking and correction step, specifically including:

[0126] Detect whether the generated polygon contour has self-intersection;

[0127] Re-triangulate the contour with self-intersection;

[0128] Distinguish between outer contour and inner hole contour, and ensure that the outer contour is sorted in counterclockwise direction and the inner hole contour is sorted in clockwise direction;

[0129] Establish a hierarchical relationship tree of the contour to accurately represent the inclusion relationship between organ structures;

[0130] Smooth the corrected contour, and use a 3rd order B-spline curve interpolation to obtain a smooth boundary that conforms to the anatomical structure.

[0131] As shown in Figure 2 The metal artifact removal model is based on a coupled double CycleGAN architecture, which includes two independent image conversion paths:

[0132] CT→MRI path: artifacted CT image→G_AB generator→artifacted MRI image; MRI→CT path: artifacted MRI image→G_BA generator→artifacted CT image;

[0133] The loss function includes: adversarial loss, which drives the generator to generate output that is indistinguishable from the real artifact-free image; cycle consistency loss, which ensures that the image can be reconstructed back to the original image after CT→MRI→CT or MRI→CT→MRI conversion; dual-path consistency joint loss: ensures that the CT and MRI images after artifact removal remain consistent in anatomical structure;

[0134] The generator and discriminator parameters of the CT path and the MRI path are independent;

[0135] Joint optimization is performed by the dual-path consistency joint loss function, and the specific settings of the optimizer are: Adam(lr=1e-4,β1=0.5,β2=0.999), batch size 4, and training rounds 200 epochs.

[0136] As shown in Figure 3 , a double-encoder-single-decoder U-Net variant is used, which is specially designed for multi-modal medical image segmentation, specifically including:

[0137] The CT encoder and the MRI encoder are both based on a pre-trained ResNet34;

[0138] 5-level feature extraction, with output sizes of [256, 128, 64, 32, 16]

[0139] The cross-modal fusion module is embedded after each encoder downsampling layer, with a total of 5 fusion points, which realizes pixel-level adaptive fusion through cross-attention calculation and dynamic weight generation;

[0140] The decoder is based on a ResNet-50 pre-trained on ImageNet, with the original fully connected layer and average pooling layer removed, and the conv2_x to conv5_x four convolution stages redesigned as up-sampling paths;

[0141] The up-sampling design is divided into 4 up-sampling stages, each of which performs 2 times up-sampling, and finally restores to the original image resolution, with the channel number sequence being [1024, 512, 256, 128, 64];

[0142] The fused encoder features are spliced at the corresponding level;

[0143] The 1×1 convolution and Softmax output segmentation probability map, and the post-processing is converted into the DICOM RT StructureSet format of the delineated contour.

[0144] The delineation structure file generation specifically comprises:

[0145] 1. For each target region or organ at risk category to be delineated, the corresponding segmentation probability map is converted into a binary mask; a probability threshold is set, and pixels with a probability value greater than or equal to the threshold are determined as target regions and assigned a value of 1, and pixels with a probability value less than the threshold are determined as background and assigned a value of 0;

[0146] 2. The obtained binary mask is subjected to connected component analysis, and all connected pixel regions are identified; according to a preset minimum area threshold, noise regions with an area that is too small are filtered out, and candidate connected regions that meet the conditions are retained;

[0147] 3. For each retained connected region, a Moore neighborhood tracking algorithm is used to track from the outermost boundary; the algorithm sequentially finds and connects adjacent boundary pixel points in the preset search direction order until a complete and closed boundary contour line is formed;

[0148] 4. The original contour point sequence obtained by tracking is subjected to polygon simplification, and the Douglas-Peucker algorithm is used to remove redundant collinear points, thereby reducing the number of contour points while ensuring shape accuracy; subsequently, spline curve interpolation is used to smooth the simplified contour, so as to obtain a smooth boundary that is more consistent with the anatomical structure;

[0149] 5. It is checked whether the generated contour has self-intersection or unreasonable topological structure; for a complex contour, an outer contour and an inner hole contour are distinguished, and it is ensured that the directions conform to the right-hand rule, so as to correctly represent the inside and outside relationship of the region;

[0150] 6. The final contour point coordinates are converted from the image pixel coordinate system to the patient three-dimensional physical coordinate system; according to the DICOM RT Structure Set standard format, organ identifiers, geometric descriptors and other metadata are added to each contour line, and a delineation structure file that can be directly used by a clinical radiotherapy planning system is generated.

[0151] As shown in Figure 4 , the delineation structure file generated finally according to the above delineation structure file generation specific steps and in accordance with the DICOM RT Structure Set standard format.

[0152] The examples described in the present application are only used to describe the preferred embodiments of the present application, and do not limit the concept and scope of the present application; without departing from the design idea of the present application, various modifications and improvements of the technical solutions of the present application made by the engineering and technical personnel in the field shall fall within the protection scope of the present application.

Claims

1. An intelligent delineation method for cervical cancer radiotherapy target areas and organs at risk, characterized in that, Includes the following steps: Step S1: Obtain the CT and MRI image dataset of cervical cancer patients; Step S2: Input the CT and MRI images containing metal artifacts into the pre-trained metal artifact removal model to obtain CT and MRI images with metal artifacts removed. Step S3: Input the CT and MRI images with metal artifacts removed into a dual-encoder U-Net network containing a cross-modal attention fusion gating module to obtain the final fused features; Step S4: Process the final fused features through the U-Net decoder, construct a few-sample adaptive mixing loss function to optimize network parameters, and output a segmentation mask; Step S5: Perform contour processing on the segmentation mask to complete the delineation of the target area and organs at risk; In step S2, the pre-trained metal artifact removal model is based on the CycleGAN architecture. The generator adopts a U-Net structure with residual connections, which includes an 8-layer encoder and an 8-layer decoder. The discriminator adopts a PatchGAN structure with 5 convolutional layers. The CycleGAN architecture adopts a coupled dual-CycleGAN architecture to handle metal artifact removal in CT and MRI modalities respectively. The generator and discriminator parameters of the CT path and MRI path are independent and are jointly optimized through a dual-path consistency joint loss function. In the pre-trained metal artifact removal model, the dual-path consistency joint loss function includes an adversarial loss term, a cyclic consistency loss term, and a semantic consistency loss term. The specific formula for the dual-path consistency joint loss function is as follows: in, This represents the joint loss function for bipath consistency. Indicates the counter-loss term. Represents the cycle consistency loss term. Represents the semantic consistency loss term. , and These represent the weighting coefficients of the corresponding loss terms; The specific formula for the adversarial loss term is as follows: in, This represents the discriminator's judgment result on the real image, that is, the probability value output by the discriminator network for an image sampled from the real data distribution that is free of metal artifacts. This represents the discriminator's judgment result on the generated image, that is, the probability value output by the discriminator network for the artifact-free image generated by the generator. This represents the expected value operator; The discriminator network outputs a 30×30 matrix, where each element represents the probability that a local region in the image is true. The specific formula for the cycle consistency loss term is: in, This represents a generator that converts images from the CT domain to the MRI domain, specifically converting CT images containing metal artifacts into artifact-free MRI images. This represents a generator that converts MRI images to CT images, specifically transforming MRI images containing metal artifacts into artifact-free CT images. This represents a CT image input containing metal artifacts. This represents an MRI image input containing metal artifacts. Represents the L1 norm; The formula for calculating the semantic consistency loss term is as follows: in, Indicates the number of samples in the batch. This represents the i-th sample in the batch. This represents the CT image of the i-th sample after removing metal artifacts. This represents the MRI image of the i-th sample after metal artifact removal. This represents a pre-trained CT image feature extraction network. This represents a pre-trained MRI image feature extraction network. Denotes the square of the L2 norm. This represents the structural loss weighting coefficient. Represents the structural similarity index. This indicates that a learnable grayscale transformation can be performed on MRI images. The linear transformation matrix represents the learning process, used to roughly align the grayscale space of the MRI image to the grayscale space of the CT image. The cross-modal attention fusion gating module achieves adaptive fusion of CT and MRI features by calculating cross-attention maps and dynamic fusion weights, specifically including the following steps: Calculate the cross-attention map from CT features to MRI features and from MRI features to CT features, respectively; The original features are modulated using a cross-attention map to obtain modulated features. The specific formula is as follows: in, Indicates the modulated CT features, This indicates the modulated MRI features. and Indicates CT and MRI features from the same level. A cross-attention map representing CT features to MRI features. A cross-attention map representing the transition from MRI features to CT features. This represents element-wise multiplication, i.e., the Hadamard product; The modulated CT and MRI features are stitched together, and dynamic fusion weights are generated through a learnable weight network. The calculation formula is as follows: in, This represents the importance weight of dynamically generated CT features. This represents the Sigmoid activation function. and This represents the learnable weight matrix and bias vector. Indicates the modulated CT features, This indicates the modulated MRI features. This indicates a splicing operation at the channel level; The CT and MRI features are weighted and summed using the dynamic fusion weights to obtain the final fused features. The calculation formula is as follows: in, Indicates the final fusion characteristics, and This indicates CT and MRI features from the same level.

2. The method according to claim 1, characterized in that, In step S1, the CT and MRI image dataset specifically includes: CT and MRI images containing metal artifacts, CT and MRI multimodal images after removing metal artifacts, and a standard dataset manually drawn by a radiation oncology specialist.

3. The method according to claim 2, characterized in that, In step S3, the cross-modal attention fusion gating module is embedded after each downsampling layer of the U-Net encoder, forming a dual-encoder-single-decoder U-Net variant. The dual encoders include: encoder CT for processing artifact-free CT images and encoder MRI for processing artifact-free MRI images.

4. The method according to claim 3, characterized in that, The U-Net decoder uses a ResNet-50 network pre-trained on the ImageNet dataset as its basic architecture. The original fully connected layers and average pooling layers of ResNet-50 are removed, and the four convolutional stages of ResNet-50, namely conv2_x to conv5_x, are redesigned as the upsampling path of the decoder. Each stage is upsampled by 2 times in sequence, and finally the original image resolution is restored.

5. The method according to claim 4, characterized in that, In step S4, the small-sample adaptive hybrid loss function is obtained by weighted combination of the generalized Dice loss function, the weighted cross-entropy loss function, and the Focal loss function. The generalized Dice loss function is used to optimize the overlap of segmented regions and handle class imbalance problems, the weighted cross-entropy loss function is used to improve pixel-level classification accuracy, and the Focal loss function is used to focus on the training of classified samples. The weight coefficients of the three loss functions are set to 0.6, 0.3, and 0.1, respectively.

6. The method according to claim 5, characterized in that, In step S5, the contour processing of the segmentation mask employs a contour tracking algorithm, specifically including the following steps: The segmentation probability map of each category is converted into a binary mask, and a fixed threshold is set to distinguish the target region from the background; The boundary tracing algorithm is used to find the boundary pixels of each connected region in turn to generate the initial contour line; The initial contour point sequence is simplified into polygons, redundant points are removed, and the final contour points are retained. The final contour point coordinates are transformed from the image pixel coordinate system to the patient's three-dimensional physical coordinate system; Generate target area and organ at risk delineation structure files according to medical standard format.

7. The method according to claim 6, characterized in that, The contour tracking algorithm also includes a topology check and correction step, specifically including: Detect whether the generated polygonal contours have self-intersections; Retriangulate the contours that have self-intersections; Distinguish between the outer contour and the inner hole contour, ensuring that the outer contour is arranged counterclockwise and the inner hole contour is arranged clockwise; Establish a hierarchical tree of the outline to accurately represent the inclusion relationships between organ structures; The corrected contour is smoothed, and a smooth boundary conforming to the anatomical structure is obtained by using cubic B-spline interpolation.

Images