System for automatically sketching tumor bed target region after breast conserving operation of breast cancer based on deep learning

Abstract

The invention discloses a breast cancer breast conserving postoperative tumor bed target area automatic sketching system based on deep learning, and belongs to the technical field of medical image processing. According to the method, the problems that a traditional manual sketching mode is time-consuming and labor-consuming and has obvious subjective difference, and accurate standardized sketching is difficult to realize in clinical practice due to the fact that an existing formulated sketching guide is used for standardizing the operation process are solved, and the tumor bed target area is automatically sketched by utilizing the deep learning model DynSegGAN, so that the accuracy of the target area sketching is improved. According to the method, the workload of manual sketching of doctors and errors of manual operation can be reduced, the sketching consistency and repeatability are improved, more accurate information is provided for clinicians, the accuracy of clinical decisions is improved, personalized radiotherapy plans can be better realized, and the pertinence and effect of treatment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to an automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning. Background Technique

[0002] Breast cancer has become an important threat to the health of women globally. According to the global cancer statistics in 2022, the number of newly diagnosed breast cancer cases reached 2.297 million, and its incidence rate ranked second among malignant tumors after lung cancer; the number of related death cases throughout the year was 666,000, ranking fourth among cancer-causing diseases. In the field of breast cancer radiotherapy, the precise positioning of the tumor bed (CTVtb) after breast-conserving surgery has always been a major challenge in clinical practice.

[0003] Since there is no significant density difference between CTVtb and the surrounding normal clinical target area tissues, the traditional manual delineation method is not only time-consuming and laborious, but also has significant subjective differences. Research shows that subtle deviations in the CTVtb contour can lead to significant changes in dosimetric parameters, directly affecting the treatment effect. Although the International Radiation Oncology Society has developed relevant delineation guidelines to standardize the operation process, it is still difficult to achieve precise and standardized delineation in clinical practice.

[0004] Therefore, it does not meet the existing requirements, and for this reason, we propose an automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning. Summary of the Invention

[0005] The purpose of the present invention is to provide an automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning. By using the deep learning model DynSegGAN to automatically delineate the tumor bed target area, it can reduce the workload of manual delineation by doctors and the errors of manual operations, improve the consistency and repeatability of delineation, provide more accurate information for clinical doctors, help improve the accuracy of clinical decision-making, better realize personalized radiotherapy plans, and improve the pertinence and effect of treatment, thus solving the problems raised in the above background technique.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] An automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning, the system includes: an image acquisition unit and an image processing unit, and the image processing unit includes a feature extraction module and an automatic delineation module;

[0008] The image acquisition unit is configured to acquire medical images after breast-conserving surgery for breast cancer and the position data of titanium clips implanted during surgery, and perform denoising preprocessing on the medical images;

[0009] The feature extraction module is configured to perform feature extraction from shallow to deep on historical patient samples through hierarchical 3D convolution and residual blocks in the deep learning model DynSegGAN, restore the spatial dimension of the medical image using the upsampling operation, perform authenticity discrimination on the input data using a hierarchical 3D convolution architecture, and output a single-channel 3D result to obtain the medical image after automatic delineation of the historical patient sample;

[0010] The automatic delineation module is configured to input the current patient's medical image and the position data of titanium clips implanted during surgery into the trained deep learning model DynSegGAN for prediction, and is used for automatic delineation of the tumor bed target area after breast-conserving surgery for breast cancer.

[0011] Furthermore, the image processing unit further includes: a model construction module and a model training module;

[0012] The model construction module is configured to construct a deep learning model DynSegGAN with a U-Net structure that is guided by a tumor bed marker mask and combines residual blocks and a channel attention mechanism; the deep learning model DynSegGAN includes: an input layer, a generator, an encoder, an intermediate layer, a decoder, a discriminator, a fully connected layer, and an output layer;

[0013] The model training module is configured to input the training set and the test set into the deep learning model DynSegGAN for training and testing, initialize through the normal distribution network weights, and optimize the hyperparameters of the deep learning model DynSegGAN using a hybrid loss function in combination with the generator.

[0014] Furthermore, initialize through the normal distribution network weights, and optimize the hyperparameters of the deep learning model DynSegGAN using a hybrid loss function in combination with the generator. Specifically:

[0015] Use the normal distribution to initialize the network parameters to avoid gradient explosion or disappearance;

[0016] Use a loss function mixed with adversarial loss, improved Dice coefficient, and edge constraint to optimize the hyperparameters. The calculation formula is as follows:

[0017] L total =λ adv ·L adv +λ recon ·L recon +λ edge ·L edge

[0018] Among them, L total represents the total loss function; L adv represents the adversarial loss function; λ advExpressed as the weight of the adversarial loss; L recon Expressed as the reconstruction loss function; λ recon Expressed as the weight of reconstruction loss; L edge Expressed as marginal loss function; λ edge is expressed as the weight of the marginal loss.

[0019] Furthermore, in the calculation formula of the hybrid loss function:

[0020] L adv =MSE(D(G(x)),1)

[0021] Where MSE is the mean square error; D(G(x)) is the evaluation result of the discriminator on the generator output;

[0022] L recon =BCEWithLogits(G(x),y)

[0023] Among them, BCEWithLogits represents the binary cross entropy loss with Logits; G(x) represents the output of the generator; y represents the true label;

[0024]

[0025] Among them, ⊙ represents the element-wise multiplication operation, which is used to calculate the intersection between the generator output and the true label; ε represents a small constant used to avoid division by zero; G(x) represents the output of the generator; y represents the true label.

[0026] Furthermore, the image processing unit further includes: a result evaluation module;

[0027] The result evaluation module is configured to evaluate whether the delineation results of the deep learning model DynSegGAN on the postoperative image are similar to the clinical CTVtb based on the region overlap index, boundary distance index and volume difference index.

[0028] Furthermore, the image delineation results are evaluated based on the regional overlap index, specifically:

[0029] The Dice similarity coefficient (DSC) was used to evaluate the overall overlap accuracy within the outlined area. The calculation formula is as follows:

[0030]

[0031] Among them, A represents the clinically approved CTVtb mask manually delineated; B represents the post-processed CTVtb mask predicted by the model; |A∩B| represents the common area; DSC represents the overlap degree between the predicted segmentation result and the true label, and the value range is [0,1]. The larger the value, the better the segmentation effect.

[0032] Furthermore, the image delineation result is evaluated based on the boundary distance index, specifically:

[0033] The 95th percentile Hausdorff distance HD95 is used to evaluate the maximum boundary deviation within the delineated area, and the calculation formula is as follows:

[0034] HD95 = max(95%-precentile(d(x,B)), 95%-precentile(d(y,A)))

[0035]

[0036] Among them, d(x,B) means that for each point x in set A, find the point in set B that is closest to x and calculate the distance between them d(y,A) means that for each point y in set B, find the point in set A that is closest to y and calculate the distance between them.

[0037] Furthermore, the image delineation result is evaluated based on the volume difference index, specifically:

[0038] The centroid distance CD is used to evaluate the distance deviation between the centroid of the delineated area and the centroid of the true area, and the calculation formula is as follows:

[0039]

[0040] Among them, and respectively represent the centroid coordinates of the predicted area and the true area; CD represents the Euclidean distance between the centroid of the predicted area and the centroid of the true area, which is used to measure the overall position deviation. The smaller the value of CD, the better.

[0041] Furthermore, a hierarchical three-dimensional convolutional architecture is adopted to distinguish the authenticity of the input data, specifically:

[0042] The input data enters the discriminator, and preliminary features are extracted through three-dimensional convolution and the LeakyReLU activation function;

[0043] Feature transformation is further carried out through three-dimensional convolution, and residual blocks are introduced. The feature propagation efficiency is enhanced through the skip connection mechanism;

[0044] After three-dimensional convolution, the SetAttention module is connected, and after activation by LeakyReLU, the residual block is used to further optimize the feature representation;

[0045] The feature dimension is regularized through global adaptive average pooling, and the feature is mapped to a single-channel output through a fully connected layer to complete the authenticity discrimination of the input data.

[0046] Further, the image processing unit further includes: a sample processing module and a sample division module;

[0047] The sample processing module is configured to set the CTV mask in the historical patient samples as the first label to constrain the generation area of the CTVtb mask outlined by the deep learning model DynSegGAN; and set the CTVtb region mask manually outlined by multiple clinical experts as the second label to verify whether the generation area of the CTVtb mask outlined by the deep learning model DynSegGAN is accurate;

[0048] The sample division module is configured to remove the historical patient samples containing unrecognized marker points, and use the remaining samples as model samples; adopt a five-fold cross-validation strategy to divide the model samples into a training set and a test set, and perform data augmentation processing using random flipping and random rotation.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] In the present invention, by using the deep learning model DynSegGAN to automatically outline the tumor bed target area, the workload of manual outlining by doctors and the error of manual operation can be reduced, the consistency and repeatability of outlining can be improved, more accurate information can be provided for clinicians, which helps to improve the accuracy of clinical decision-making, can better realize personalized radiotherapy plans, and improve the pertinence and effect of treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a composition diagram of the automatic tumor bed target area outlining system for breast cancer after breast-conserving surgery based on deep learning of the present invention;

[0052] Figure 2 It is a flowchart of automatic recognition of tumor bed markers for breast cancer after breast-conserving surgery based on deep learning of the present invention;

[0053] Figure 3 It is a diagram of the outlining results of the deep learning and non-deep learning methods of the present invention;

[0054] Figure 4 It is a comparison diagram of three outlining results based on automatic outlining of the present invention;

[0055] Figure 5Comparison diagram of three contouring results based on manual contouring by doctors for the present invention;

[0056] Figure 6 Comparison diagram of three contouring results based on automatic contouring and manual contouring by doctors for the present invention. Detailed implementation manners

[0057] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0058] To solve the technical problems in the prior art that the traditional manual contouring method is not only time-consuming and laborious, but also has significant subjective differences. Although the International Radiation Oncology Society has formulated relevant contouring guidelines to standardize the operation process, it is still difficult to achieve precise standardized contouring in clinical practice. Please refer to Figures 1-6 , the following technical solutions are provided in this embodiment:

[0059] An automatic contouring system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning. The system includes: an image acquisition unit and an image processing unit. The image processing unit includes a feature extraction module and an automatic contouring module;

[0060] The image acquisition unit is configured to obtain the medical images after breast-conserving surgery for breast cancer and the position data of titanium clips implanted during the surgery. The titanium clip position data is used to provide accurate position information of the tumor bed area. The medical images include: the medical images of the current patient and the historical patient samples. The historical patient samples are used to train and test the deep learning model DynSegGAN. Since the medical images may be affected by various noises during the acquisition process, such as equipment noise, motion artifacts, etc., it is necessary to perform denoising preprocessing on the medical images, including using filters to remove noise and improve the image quality, so as to provide clearer input data for subsequent feature extraction and model training.

[0061] The feature extraction module is configured to perform feature extraction on the historical patient samples from shallow to deep through hierarchical three-dimensional convolution and residual blocks in the deep learning model DynSegGAN, combine downsampling operations, use upsampling operations to restore the spatial dimension of the medical images, make up for the loss of detailed information during the downsampling process, and use a hierarchical three-dimensional convolution architecture to perform authenticity discrimination on the input data and output a single-channel three-dimensional result to obtain the automatically contoured medical image of the historical patient sample. Specifically:

[0062] The input data enters the discriminator, and preliminary features are extracted through three-dimensional convolution and the LeakyReLU activation function;

[0063] The feature transformation is further performed through three-dimensional convolution, and residual blocks are introduced to enhance the feature propagation efficiency through the skip connection mechanism and alleviate the gradient vanishing problem.

[0064] After the three-dimensional convolution, the SetAttention module is connected, and after activation by LeakyReLU, the residual block is used to further optimize the feature expression;

[0065] The feature dimension is regularized through global adaptive average pooling, and the feature map is compressed into a fixed-size output for subsequent fully connected layer processing;

[0066] The fully connected layer maps the features to a single-channel output to determine whether the input data belongs to a specific category. The output is a single-channel three-dimensional result, which is a segmentation map. This includes the tumor bed target area predicted by the model after breast-conserving surgery for breast cancer, thereby completing the authenticity judgment of the input data, realizing automatic outlining, and assisting doctors in making more accurate diagnoses and treatment plans.

[0067] The automatic outlining module is configured to input the current patient's medical images and the position data of the titanium clip implanted during surgery into the trained deep learning model DynSegGAN for prediction, align the titanium clip position data with the medical image data to ensure the spatial consistency of the two, and analyze the input medical images through the deep learning model DynSegGA to automatically identify and outline the tumor bed target area; and mark or highlight the tumor bed target area in the output three-dimensional segmentation map, clearly showing the tumor bed target area position predicted by the model, thereby realizing automatic outlining of the tumor bed target area after breast-conserving surgery for breast cancer.

[0068] The image processing unit also includes: a sample processing module, a sample division module, a model construction module and a model training module;

[0069] The sample processing module is configured to set the CTV mask in historical patient samples as the first label to constrain the deep learning model DynSegGAN to outline the generated area of the CTVtb mask; and set the CTVtb regional mask manually outlined by multiple clinical experts as the second label to verify whether the generated area of the CTVtb mask outlined by the deep learning model DynSegGAN is accurate.

[0070] The sample partitioning module is configured to remove unidentified markers from historical patient samples, leaving 67 samples as model samples; a five-fold cross-validation strategy is used to divide the model samples into training and test sets, with each fold containing 20% of the samples (13 samples) as the test set and the remaining 80% of the samples used for training; and random flipping and random rotation are used to enhance the data to improve the generalization of the deep learning model DynSegGAN.

[0071] A model construction module, configured to construct a deep learning model DynSegGAN with a U-Net structure that combines residual blocks and channel attention mechanism, guided by a tumor bed marker mask; the deep learning model DynSegGAN includes: an input layer, a generator, an encoder, an intermediate layer, a decoder, a discriminator, a fully connected layer, and an output layer.

[0072] A model training module, configured to input a training set and a test set into the deep learning model DynSegGAN for training and testing, initialize through a normal distribution network weight, and optimize the hyperparameters of the deep learning model DynSegGAN by combining a generator with a hybrid loss function. Specifically:

[0073] Initialize network parameters using a normal distribution to avoid gradient explosion or disappearance;

[0074] Use a loss function mixed with adversarial loss, improved Dice coefficient, and edge constraint to optimize the hyperparameters. The calculation formula is as follows:

[0075] L total =λ adv ·L adv +λ recon ·L recon +λ edge ·L edge

[0076] Among them, L total represents the total loss function; L adv represents the adversarial loss function; λ adv represents the weight of the adversarial loss; L recon represents the reconstruction loss function; λ recon represents the weight of the reconstruction loss; L edge represents the edge loss function; λ edge represents the weight of the edge loss.

[0077] Among them, in the calculation formula of the hybrid loss function:

[0078] L adv =MSE(D(G(x)),1)

[0079] Among them, MSE represents the mean square error; D(G(x)) represents the evaluation result of the discriminator on the output of the generator;

[0080] L recon =BCEWithLogits(G(x),y)

[0081] Among them, BCEWithLogits represents the binary cross entropy loss with Logits; G(x) represents the output of the generator; y represents the true label;

[0082]

[0083] Among them, ⊙ represents the element-wise multiplication operation, which is used to calculate the intersection between the generator output and the true label; ε represents a small constant used to avoid division by zero; G(x) represents the output of the generator; y represents the true label.

[0084] Hyperparameter optimization was performed using the AdamW optimizer, with an initial learning rate of 1e-4 and a weight decay of 1e-5. Since the input size can be arbitrary, a batch size of 1 was used for 400 epochs of training. Cosine annealing, learning rate scheduling, and early stopping were used, with a patience period of 20 epochs to prevent overfitting. The deep learning model, DynSegGAN, was developed based on the PyTorch 2.6 framework and implemented on an NVIDIA 3060Ti GPU platform using CUDA acceleration technology. During training, a mixed-precision strategy was used to reduce video memory usage and improve computational efficiency.

[0085] The image processing unit further includes: a result evaluation module;

[0086] A result evaluation module is configured to evaluate whether the delineation results of the deep learning model DynSegGAN on the postoperative image are similar to the clinical CTVtb based on the regional overlap index, the boundary distance index, and the volume difference index;

[0087] The image delineation results are evaluated based on the regional overlap index, specifically:

[0088] The Dice similarity coefficient (DSC) is used to evaluate the overall overlap accuracy within the outlined area. The Dice coefficient is an indicator that measures the similarity between two sets and is used to evaluate the degree of overlap between the predicted results and the true labels. The calculation formula is as follows:

[0089]

[0090] Among them, A represents the manually drawn clinically approved CTVtb mask; B represents the model-predicted CTVtb mask after post-processing; |A∩B| represents the shared area; DSC is used to measure the degree of overlap between the predicted segmentation result and the true label, with a value range of [0,1]. The larger the value, the better the segmentation effect.

[0091] The image delineation results are evaluated based on the boundary distance index, specifically:

[0092] The maximum boundary deviation within the contoured region is evaluated using the 95th percentile Hausdorff distance (HD95). The Hausdorff distance is a method for measuring the distance between two point sets, and the 95th percentile Hausdorff distance further reduces the influence of outliers, making the evaluation results more stable and reliable. The calculation formula is as follows:

[0093] HD95 = max(95%-precentile(d(x,B)), 95%-precentile(d(y,A)))

[0094]

[0095] Where d(x,B) means that for each point x in set A, find the point in set B that is closest to x and calculate the distance between them. d(y,A) means that for each point y in set B, find the point in set A that is closest to y and calculate the distance between them.

[0096] The image contouring results are evaluated based on the volume difference index, specifically:

[0097] The centroid distance (CD) is used to evaluate the distance deviation between the centroid of the contoured region and the centroid of the true region. The calculation formula is as follows:

[0098]

[0099] Where and represent the centroid coordinates of the predicted region and the true region respectively; CD represents the Euclidean distance between the centroid of the predicted region and the centroid of the true region, which is used to measure the overall position deviation. The smaller the value of CD, the better.

[0100] Through the comprehensive evaluation of the above three indicators, the performance of the deep learning model DynSegGAN in the automatic contouring task of the tumor bed target area after breast-conserving surgery for breast cancer can be comprehensively evaluated, ensuring the accuracy and reliability of the contouring results.

[0101] In one embodiment, assume Figure 3 , where A shows the situation of the CTVtb manually contoured by a doctor, B shows the contouring result of the automatically recognized markers input into DynSegGAN, C shows the contouring result of the automatically recognized markers input into DynSegUnet, D shows the contouring result of the automatically recognized markers input into Basic Unet, and E shows the contouring result of the automatically recognized markers using a non-deep learning method to simulate external expansion.

[0102] Table 1 and Table 2 summarize the comparison of DSC, HD95, and CD between DynSegGAN, DynSegUnet, and BasicUnet in the 5-fold cross-validation of manual and automatic delineation by doctors, as well as the mean and standard deviation among different cases.

[0103]

[0104] Table 1: Quantitative comparison of the segmentation results of DynSegUnet, DynSegGAN, and Basic Unet for five test folds with the mean among individual cases using the automatically delineated tumor bed markers as input

[0105]

[0106]

[0107] Table 2: Quantitative comparison of the segmentation results of DynSegUnet, DynSegGAN, and Basic Unet for five test folds with the mean among individual cases using the manually delineated tumor bed markers by doctors as input

[0108] Such as: Figure 4 In [reference], the quantitative comparison of the segmentation results of DynSegUnet, DynSegGAN, and Basic Unet for five test folds with the mean among individual cases using the automatically delineated tumor bed markers as input. Figure 5 In [reference], the quantitative comparison of the segmentation results of DynSegUnet, DynSegGAN, and Basic Unet for five test folds with the mean among individual cases using the manually delineated tumor bed markers by doctors as input. Figure 6 In [reference], the quantitative comparison of the segmentation results of DynSegUnet, DynSegGAN, and Basic Unet for five test folds with the mean among individual cases using both the automatically and manually delineated tumor bed markers as input.

[0109] In the DSC metric, the DSC of the automatically delineated and manually delineated DynSegGAN is approximately 0.1119 and 0.095 higher than that of the automatically delineated and manually delineated Basic Unet, respectively.

[0110] The DSC of the automatically delineated and manually delineated DynSegUnet is also approximately 0.1182 and 0.0948 higher than the corresponding values of Basic Unet, significantly optimizing the overlap consistency of the segmented regions.

[0111] In terms of the HD95 metric, the HD95 values of the automatic delineation by DynSegGAN and DynSegUnet were significantly lower than those of the automatic delineation by BasicUnet, with decreases of 9.6659 mm and 9.8783 mm, respectively. The deviation between the segmentation boundary and the true boundary was smaller.

[0112] In the CD metric, the CD values of the automatic delineation by DynSegGAN and DynSegUnet decreased by 7.304 mm and 6.8282 mm, respectively, compared with those of the automatic delineation by BasicUnet, highlighting their advantages in target center localization and significantly reducing the spatial position deviation.

[0113] It can be seen that DynSegGAN and DynSegUnet comprehensively demonstrated better segmentation performance than BasicUnet through the significant improvement of DSC, the substantial decrease of HD95 and CD. In addition, regarding DynSegGAN, in this study, it was found that increasing the adversarial weight of the adversarial network would lead to training instability and worse segmentation results; in the results of non - deep - learning methods, the average DSC was 0.7068 ± 0.0539, the mean value of HD95 was 4.3917 ± 2.4693 mm, and the mean value of CD was 2.4829 ± 2.3242 mm; thus, it can be seen that the non - deep - learning method simulating the external expansion of the clinical situation obtained smaller average HD95 and CD values, but had a larger standard deviation, with more obvious fluctuations, and the consistency of delineation was severely tested.

[0114] Beneficial effects achieved by the above content: By using the deep - learning model DynSegGAN to automatically delineate the tumor bed target area, the workload of manual delineation by doctors and the errors of manual operation can be reduced, the consistency and repeatability of delineation can be improved, more accurate information can be provided for clinicians, which helps to improve the accuracy of clinical decision - making, better realize personalized radiotherapy plans, and improve the pertinence and effectiveness of treatment.

[0115] Working principle: Obtain the medical images after breast - conserving surgery for breast cancer and the position data of titanium clips implanted during surgery. Use the deep - learning model DynSegGAN to extract the feature information of medical images, and combine the position data of titanium clips implanted during surgery for prediction to achieve the automatic delineation of the tumor bed target area after breast - conserving surgery for breast cancer; thereby improving the accuracy and efficiency of delineation, reducing the workload of doctors at the same time, and providing support for clinical treatment.

[0116] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "having" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus.

[0117] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based automatic delineation system for the target area of the tumor bed after breast-conserving surgery for breast cancer, characterized by: The system includes: an image acquisition unit and an image processing unit, and the image processing unit includes a feature extraction module and an automatic delineation module; The image acquisition unit is configured to acquire medical images after breast-conserving surgery for breast cancer and the position data of titanium clips implanted during the surgery, and perform denoising preprocessing on the medical images; The feature extraction module is configured to perform feature extraction from shallow to deep on historical patient samples through hierarchical three-dimensional convolution and residual blocks in the deep learning model DynSegGAN, restore the spatial dimension of the medical image by using upsampling operation, perform authenticity discrimination on the input data by adopting a hierarchical three-dimensional convolution architecture, and output a single-channel three-dimensional result to obtain the automatically delineated medical image of the historical patient sample; The automatic delineation module is configured to input the medical image of the current patient and the position data of the titanium clips implanted during the surgery into the trained deep learning model DynSegGAN for prediction, and is used for automatically delineating the tumor bed target area after breast-conserving surgery for breast cancer.

2. The deep learning-based automatic tumor bed target delineation system after breast-conserving surgery for breast cancer according to claim 1 is characterized by: The image processing unit further includes: a model construction module and a model training module; The model construction module is configured to construct a deep learning model DynSegGAN with a U-Net structure guided by a tumor bed marker mask and combined with a residual block and a channel attention mechanism; the deep learning model DynSegGAN includes: an input layer, a generator, an encoder, an intermediate layer, a decoder, a discriminator, a fully connected layer, and an output layer; The model training module is configured to input a training set and a test set into the deep learning model DynSegGAN for training and testing, initialize through normal distribution network weights, and optimize the hyperparameters of the deep learning model DynSegGAN by using a hybrid loss function in combination with the generator.

3. The automatic tumor bed target area delineation system for breast cancer after breast-conserving surgery based on deep learning according to claim 2, characterized in that: Initialize through normal distribution network weights, and optimize the hyperparameters of the deep learning model DynSegGAN by using a hybrid loss function in combination with the generator. Specifically: Use normal distribution to initialize network parameters to avoid gradient explosion or disappearance; Use a loss function mixed with adversarial loss, improved Dice coefficient, and edge constraint to optimize the hyperparameters. The calculation formula is as follows: L total =λ adv ·L adv +λ recon ·L recon +λ edge ·L edge Among them, L total represents the total loss function; L adv represents the adversarial loss function; λ adv represents the weight of the adversarial loss; L recon represents the reconstruction loss function; λ recon represents the weight of the reconstruction loss; L edge represents the edge loss function; λ edge represents the weight of the edge loss.

4. The automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning according to claim 3, wherein: In the calculation formula of the hybrid loss function: L adv = MSE(D(G(x)), 1) Among them, MSE represents mean squared error; D(G(x)) represents the evaluation result of the discriminator on the output of the generator; L recon = BCEWithLogits(G(x), y) Among them, BCEWithLogits represents binary cross-entropy loss with Logits; G(x) represents the output of the generator; y represents the true label; Among them, ⊙ represents an element-wise multiplication operation, which is used to calculate the intersection between the output of the generator and the true label; ε represents a small constant, which is used to avoid division by zero; G(x) represents the output of the generator; y represents the true label.

5. The automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning according to claim 1, wherein: The image processing unit further includes: a result evaluation module; The result evaluation module is configured to evaluate whether the delineation result of the deep learning model DynSegGAN on the postoperative image is similar to the clinical CTVtb based on region overlap metrics, boundary distance metrics, and volume difference metrics.

6. The automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning according to claim 5, characterized in that: Evaluate the image delineation result based on the region overlap metric. Specifically: The Dice similarity coefficient (DSC) was used to evaluate the overall overlap accuracy within the outlined area. The calculation formula is as follows: Among them, A represents the manually drawn clinically approved CTVtb mask; B represents the model-predicted CTVtb mask after post-processing; |A∩B| represents the shared area; DSC is used to measure the degree of overlap between the predicted segmentation result and the true label, with a value range of [0,1]. The larger the value, the better the segmentation effect.

7. The automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning according to claim 5, characterized in that: The image delineation results are evaluated based on the boundary distance index, specifically: The 95th percentile Hausdorff distance HD95 was used to evaluate the maximum boundary deviation within the delineated area, and the calculation formula is as follows: HD95=max(95%-precentile(d(x,B)),95%-precentile(d(y,A))) Among them, d(x,B) means that for each point x in set A, find the point in set B that is closest to x and calculate the distance between them. d(y,A) means that for each point y in set B, find the point in set A that is closest to y and calculate the distance between them.

8. The deep learning-based automatic tumor bed target delineation system after breast-conserving surgery for breast cancer according to claim 5 is characterized by: The image delineation results were evaluated based on volume difference indices, specifically: The centroid distance CD is used to evaluate the distance deviation between the centroid of the outlined area and the true area centroid. The calculation formula is as follows: Among them, and respectively represent the centroid coordinates of the predicted region and the ground truth region; CD represents the Euclidean distance between the centroid of the predicted region and the centroid of the ground truth region, which is used to measure the overall position deviation, and the smaller the value of CD, the better.

9. The automatic delineation system for the tumor bed target area after breast-conserving surgery for breast cancer based on deep learning according to claim 1, characterized in that: A layered 3D convolutional architecture is used to identify the authenticity of input data, specifically: The input data enters the discriminator and extracts preliminary features through three-dimensional convolution and LeakyReLU activation function; The feature transformation is further performed through three-dimensional convolution, and residual blocks are introduced to enhance the efficiency of feature propagation through the skip connection mechanism; After the three-dimensional convolution, the SetAttention module is connected, and after activation by LeakyReLU, the residual block is used to further optimize the feature expression; The feature dimension is regularized through global adaptive average pooling, and the features are mapped to single-channel output through the fully connected layer to complete the authenticity judgment of the input data.

10. The deep learning-based automatic tumor bed target delineation system after breast-conserving surgery for breast cancer according to claim 1, characterized in that: The image processing unit further includes: a sample processing module and a sample division module; The sample processing module is configured to set the CTV mask in the historical patient sample as the first label to constrain the deep learning model DynSegGAN to delineate the generated area of the CTVtb mask; and set the CTVtb regional mask manually delineated by multiple clinical experts as the second label to verify whether the generated area of the CTVtb mask delineated by the deep learning model DynSegGAN is accurate; The sample partitioning module is configured to remove unidentified markers from historical patient samples, and the remaining samples are used as model samples; a five-fold cross-validation strategy is used to divide the model samples into training and test sets, and random flipping and random rotation are used to enhance the data.

Images