An intelligent segmentation method for lung cancer target area and organs at risk

By employing a joint segmentation network framework for lung cancer target areas and organs at risk, combined with conditional strategies and uncertainty-aware self-reinforcing models, the low segmentation efficiency of lung cancer target areas and organs at risk in existing technologies is addressed, achieving accurate and efficient segmentation that meets the clinical needs of radiologists.

CN113793304BActive Publication Date: 2024-11-01TIANJIN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110965670.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-08-23
Publication Date
2024-11-01
Estimated Expiration
2041-08-23

AI Technical Summary

Technical Problem

In existing technologies, the automatic segmentation technology for lung cancer target areas and organs at risk has a low degree of automation, unstable performance, and is easily affected by the diversity of databases, making it difficult to achieve accurate segmentation, especially for the esophagus and similar soft tissues in CT images with low contrast. Moreover, existing methods require training for individual models, resulting in low execution efficiency and data waste.

Method used

We employ a joint segmentation network framework for lung cancer target areas and organs at risk, using teacher and student models, combined with conditional policies and uncertainty-aware self-reinforcing models. We optimize the segmentation task through a joint loss function to achieve joint learning of target areas and organs at risk. We design conditional policies and personalized joint loss functions using multiple publicly available image datasets.

Benefits of technology

It achieves precise and efficient segmentation of the target area and organs at risk, improves the generalization ability and operational efficiency of the segmentation model, reduces computational costs, meets the clinical needs of radiologists, and improves the quality and efficiency of lung cancer treatment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113793304B_ABST
    Figure CN113793304B_ABST
Patent Text Reader

Abstract

The present invention discloses an intelligent target area and organ at risk segmentation method for lung cancer, which adopts a common segmentation network framework for target areas and organs at risk for lung cancer to perform dual-path segmentation; the common segmentation network framework includes a teacher model and a student model, both of which use nnU-Net as the backbone; for organs at risk and target areas, predefined auxiliary condition information is respectively given and embedded in the upsampling process, connected to the corresponding decoder in the decoding process, and the predicted mask is fused to obtain the final segmentation result. The present invention can accurately and efficiently segment the target area and organs at risk.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an intelligent segmentation method for lung cancer target areas and organs at risk. Background Art

[0002] Radiotherapy, as an important means of lung cancer treatment, helps to improve the clinical cure rate of patients with advanced non-small cell lung cancer, prolong the life cycle of patients, and improve the quality of life of patients. The main purpose of radiotherapy is to provide a curative dose to the tumor target area while protecting organs at risk from radiation to avoid complications. When formulating a treatment plan, the target area contour must be accurately outlined to determine the appropriate dose to ensure the inactivation of the target area. In addition, pathological organs at risk with low radiation resistance must be accurately outlined to ensure that they are not irradiated with additional doses, because the damage caused by radiation is irreversible. Therefore, the precise segmentation of the target area and organs at risk plays a vital role in the radiotherapy plan for lung cancer and is closely related to the success and quality of the operation.

[0003] In clinical applications, the process of outlining the target and organs at risk is performed manually by radiologists, and radiotherapy plans are formulated based on this process. However, outlining the target and organs at risk is an extremely tedious and rigorous task. Many unstable factors, such as distraction, fatigue, and limitations in professional level, can lead to misunderstandings of target segmentation. In particular, the unique spatial heterogeneity of lung tumors will greatly affect doctors' cognition of their boundary contours. Therefore, manual segmentation of target and organs at risk still has great limitations, which seriously affects the quality and efficiency of radiotherapy planning.

[0004] Automated target and organ-at-risk segmentation technology can overcome these challenges to a certain extent and provide relatively accurate segmentation results. It can not only help radiologists obtain the target anatomical structure in a shorter time, but also allow radiologists to have more time to further manually improve the segmentation results and develop more accurate radiotherapy plans. However, the automatic segmentation technology based on traditional methods has a low degree of automation, unstable performance, is easily affected by database diversity, and has poor generalization ability. In addition, it is difficult to achieve satisfactory segmentation performance for soft tissues and organs at risk that have low contrast in CT images such as the esophagus and target area and have similar imaging characteristics to the surrounding tissues.

[0005] With the development of deep learning technology, intelligent medical image segmentation technology has made significant progress. Intelligent segmentation algorithms based on deep learning technology use fully labeled data as a learning library. By building neural networks of different levels and dimensions, they can comprehensively learn the highly discriminative deep features of organs at risk and target areas in different patients, different cancer stages, and different body positions, realize the classification of the algorithm at the pixel level, and then provide accurate prediction results to achieve intelligent segmentation of medical images. Segmentation algorithms based on deep learning have stronger adaptability, robustness, and generalization capabilities, and mainly rely on the public availability of large multi-class labeled data sets. In addition, it is very expensive and laborious to create sufficient and professional medical image data sets with target and organ at risk labels. Existing public imaging data sets encountered in the medical field often include single or partially labeled organs. Therefore, almost all existing segmentation frameworks between target and organ at risk are performed separately with separate models. If you want to achieve accurate segmentation of target and organ at risk, you need to train two segmentation models for two independent labeled data sets to obtain the segmentation results of target and organ at risk. This process is cumbersome, inefficient, and a waste of computing power and labeled data. Summary of the invention

[0006] The purpose of the present invention is to provide an intelligent target area and organ at risk segmentation method for lung cancer in order to address the technical defects in the prior art and achieve fast, efficient and accurate target area and organ at risk segmentation.

[0007] The technical solution adopted to achieve the purpose of the present invention is:

[0008] An intelligent target and organ-at-risk segmentation method for lung cancer, which uses a common segmentation network framework for target and organ-at-risk segmentation for lung cancer to perform dual-path segmentation; the common segmentation network framework includes a teacher model and a student model, and both models use nnU-Net as the backbone;

[0009] For the organs at risk and target areas, predefined auxiliary condition information is respectively given and embedded in the upsampling process, connected to the corresponding decoder in the decoding process, and the predicted masks are fused to obtain the final segmentation result;

[0010] The target area labeled data and the organ at risk labeled data in the input data are used as labeled data and input into the student model for supervised target area segmentation, and the target area unlabeled data are input into the teacher model for semi-supervised organ at risk segmentation; the student model learns from the teacher model by minimizing the combined loss and consistency loss relative to the teacher model target; the weight of the teacher model is updated by the exponential moving average (EMA) of the student model weight to improve the performance quality in different training processes;

[0011] The teacher model estimates the uncertainty feature map of the OAR while generating the prediction mask to guide the student model to learn more reliable targets from the teacher model;

[0012] The student model is optimized by combining the fully supervised loss and the unsupervised consistency loss. In the supervised target segmentation, the CT image data and masks used for training are randomly sampled from the target dataset and input into the network training, and the target segmentation model is optimized only by the fully supervised loss.

[0013] The intelligent segmentation method for lung cancer target area and organs at risk of the present invention starts from actual clinical needs, integrates multiple publicly available image data sets, enriches data sources, uses no-new-Net (nnU-Net) as the segmentation framework, designs conditional strategies, and merges prior conditional auxiliary information as an additional input layer into the decoding process, effectively utilizing predefined auxiliary conditional information, breaking the barriers of different labeling targets between different independent data sets, and realizing the joint learning of target areas and organs at risk.

[0014] Secondly, an uncertainty-aware self-ensembling mode is used to assist in the segmentation of organs at risk. It can effectively utilize unlabeled data (images from target labeled datasets) by encouraging consistent predictions of the same input under different perturbations, thereby enhancing the segmentation accuracy of organs at risk.

[0015] In addition, different personalized joint loss functions are designed to optimize the different segmentation tasks of target area and organs at risk in the dual pathways.

[0016] The proposed method is evaluated on two publicly available datasets of spinal cord, left and right lungs, heart, esophagus, and target area, where satisfactory segmentation performance is achieved in terms of both region-based similarity coefficient (DSC) and boundary-based Hausdorff distance (HD).

[0017] The proposed semi-supervised conditional nnU-Net breaks the barrier between non-overlapping labeled datasets and further alleviates the “data hunger” and “data waste” problems in multi-class segmentation.

[0018] The method of the present invention can help radiologists formulate lung cancer radiotherapy plans in clinical practice, and can be integrated into the intelligent segmentation system as a clinical auxiliary tool for radiologists to accurately and efficiently segment the target area and organs at risk. It can greatly improve the work efficiency of doctors, reduce their burden, save labor costs, and improve the treatment quality and survival rate of lung cancer patients.

[0019] The intelligent segmentation method for lung cancer target and organs at risk proposed in the present invention is concise, advanced and original. The proposed conditional nnU-Net can realize joint learning between different labeled data sets, introduce semi-supervised learning into the joint segmentation task based on multiple independent data sets for the first time, and design different joint losses for different segmentation tasks to realize the joint segmentation of target and organs at risk. While ensuring the segmentation accuracy, the present invention can greatly reduce the computing cost, improve the utilization of computing power, and improve the operating efficiency and generalization ability of the segmentation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of the conditional nnU-Net network structure of an embodiment of the present invention;

[0021] Figure 2 This is a structural diagram of a semi-supervised nnU-Net based on an uncertainty-aware self-enhancement model according to an embodiment of the present invention.

[0022] Figure 3 Schematic diagram of five-fold cross-validation training loss;

[0023] Figure 4 This is an example of the prediction result of the semi-supervised nnU-Net. The filled area is the prediction result, and the solid line part is the true label provided by the public dataset. DETAILED DESCRIPTION

[0024] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0025] The present invention optimizes the nnU-Net basic framework network, introduces conditional strategies to break the labeling barriers between labeled data, and realizes joint learning between different data sets. In order to further utilize the images between the data sets, an uncertainty-aware self-enhancement model is introduced to assist in the segmentation of organs at risk. In order to better fit the learning performance of the segmentation model, two joint loss functions are designed to optimize the segmentation tasks of the target area and organs at risk, thereby improving the segmentation accuracy and generalization ability of the learning model.

[0026] The intelligent target and organ-at-risk segmentation method for lung cancer in an embodiment of the present invention has the following specific implementation steps: for 3DCT medical images, the problem of intelligent segmentation of target and organ-at-risk for lung cancer is solved, including five stages: first, CT image data acquisition; second, data preprocessing; third, introducing conditional strategies to break the barriers of different target labels between different data sets; fourth, introducing a semi-supervised learning mechanism to assist in the segmentation of organs-at-risk; and fifth, achieving joint learning of target and organ-at-risk.

[0027] To ensure the repeatability of the comparative experiment and the authority of the experimental data, the CT images used in the data acquisition phase are from the LCTSC dataset and the MSD LT dataset. The organ-at-risk labeling data is from the Lung CT Segmentation Challenge (LCTSC) dataset, and the target area labeling data is from the Medical Segmentation Decathlon (MSD LT) dataset. The labeled data are all labeled by authoritative radiologists.

[0028] Among them, the common segmentation network framework for lung cancer target areas and organs at risk is nnU-Net. After data preprocessing, the network will randomly read data from the LCTSC and MSD LT fused data set and send it to nnU-Net for learning. Among them, nn-UNet is a typical U-shaped symmetrical structure consisting of an encoding process and a decoding process, such as Figure 1 As shown in the figure, the encoding and decoding processes consist of 5 downsampling and 5 upsampling respectively, and each sampling consists of 2 convolution blocks. The output feature map size of the first downsampling is 48*192*192, the step size is (1,2,3), and the number of convolution kernels is 32. The upsampling process is the inverse operation of the downsampling process. Before each upsampling, the output of the downsampling at the same layer will be connected to the output of the upsampling through the Concat operation, and then the convolution operation will be performed.

[0029] In order to achieve joint learning between different labeled data sets, the embodiment of the present invention introduces a conditional strategy. Figure 1 As shown, for the organs at risk and the target area, the predefined auxiliary condition information C is given respectively. i . Auxiliary condition information C i The specific form of is a conditional tensor, which has the same size as the output of the downsampling at the same layer and the output of the previous upsampling. The initial value can be defined by yourself, but it is necessary to ensure that the initial values ​​of the conditional tensors corresponding to the organs at risk and the target area are not equal. For example, Figure 1 As shown in the figure, the feature map size output after the first upsampling is 2*320*6*12*12, and the output feature map of the same layer downsampling is also 2*320*6*12*12, where the batch size is 2. Assuming that the data in the first batch comes from the LCTSC dataset (organs at risk dataset) and the data in the second batch comes from the MSD LT dataset (target dataset), a conditional tensor C with a size of 2*1*6*12*12 can be defined. i, where the initial values ​​of all elements in the organ-at-risk condition tensor C0[0,:,:,:] corresponding to the first batch are any positive integers M, and the initial values ​​of all elements in the target area condition tensor C1[1,:,:,:] corresponding to the second batch are any positive integers N (note that M≠N). After the first upsampling is completed, the output of the same layer downsampling, the output of the upper layer upsampling and the condition tensor C are concatenated through the Concat operation. i Connect together and then perform convolution operation. In this way, according to the given auxiliary condition information C i ,During the decoding process, the target area and organs at risk can be ,defined a second time to enhance their data attributes so that the ,segmentation tasks of the target area and organs at risk do not ,interfere with each other, so as to achieve the joint learning of the target area and organs at risk.

[0030] The composition of conditional nnU-Net is as follows Figure 1 As shown. The conditional nnU-Net can not only learn multi-organ tissue segmentation from different datasets, but also support personalized organ reasoning. Because the conditional strategy is executed during the upsampling process, it does not interfere with the downsampling process. To some extent, the output of the encoder is a shareable feature map. Therefore, to choose to segment the organ at risk or the target area, you only need to connect the corresponding conditional tensor to the decoder and fuse the predicted mask to obtain the final segmentation result.

[0031] Conditional nnU-Net maximizes the value of discrete available data sets and improves segmentation efficiency in clinical applications. In order to explore the potential value between two independent data sets of target area labeling data sets and organ at risk labeling data sets, an embodiment of the present invention uses a semi-supervised mechanism based on an uncertainty-aware self-enhancement model to assist in organ at risk segmentation. For labeled data, the student model performs supervised learning. For unlabeled data, the segmentation map is predicted by the teacher model as the learning target of the student model, and the uncertainty of the learning target is evaluated at the same time.

[0032] Based on the uncertainty of learning objectives, a consistent loss function is used to improve the performance of the student model. Figure 2 As shown in the figure, during the training process, the target CT image data has two directions. On the one hand, as target labeled data, it carries the target label. With the help of conditional tensors, the student model conducts full supervised training and learning of the target. During the learning process, based on the real label of the target, the full supervised loss of the target is fed back to optimize the target segmentation model. On the other hand, because the target CT images also include imaging data of organs at risk such as arteries, heart and esophagus, it can participate in the segmentation process of organs at risk with the identity of unlabeled data and with the help of semi-supervised learning mechanism. These unlabeled data are used to predict the segmentation map through the teacher model as the learning target of the student model, and the uncertainty of each target prediction is estimated through Monte Carlo sampling at the same time.

[0033] Based on the uncertainty of the learning goal, the consistency loss function is used to improve the performance of the student model. There is only one direction for the OAR labeled data, that is, with the help of the OAR condition tensor, through the student model, under the joint feedback of the full supervision loss and the semi-supervised loss, the OAR segmentation model is optimized.

[0034] The teacher model and the student model have the same network structure, both of which are nnU-Net. In order to improve the prediction performance of the integrated network of the teacher model and the student model at different training stages, the teacher model adopts the exponential moving average (EMA) strategy to optimize the prediction results of the teacher model based on the update of the learning model parameters.

[0035] Due to the strong heterogeneity of the target area, the segmentation of the target area is more challenging than the segmentation of organs at risk. In order to maximize the optimization of the two main tasks of target area segmentation and organ at risk segmentation, the embodiment of the present invention personalizes two types of joint loss functions to optimize the segmentation network model. For the target area, the Dice loss function and the Focal loss function are jointly used to optimize the task. For organs at risk, the Dice loss function and the cross entropy loss function are used to optimize the segmentation model. The design of the loss function here, combined with the use environment of the conditional nnU-Net, can ensure that the dual-path segmentation tasks are optimized separately and do not interfere with each other.

[0036] The following is an explanation of the intelligent segmentation method for lung cancer target areas and organs at risk based on experimental data:

[0037] 1. Data acquisition:

[0038] The left and right lungs, heart, esophagus, spinal cord and target area were selected as segmentation targets, and the method of the present invention was evaluated on the LCTSC and MSD LT datasets. The LCTSC dataset is used to evaluate the performance of the automatic segmentation method of organs at risk for radiotherapy in chest CT images. It contains 60 sets of CT scan images from three different institutions with professional labels of the left and right lungs, heart, esophagus and spinal cord. The MSD LT dataset is designed to segment the target area in the CT image, which consists of 63 sets of CT scans and professional labels of the target area.

[0039] The method of the present invention is based on the open source platform PyTorch and is implemented using the NVIDIA TITAN Xp GPU. In order to capture as much contextual information as possible, the Batch size of the network is set to 2 to increase the size of the Patch size within the predefined GPU memory budget. As a key parameter, larger Spacing will lead to the loss of details, and smaller Spacing will lead to a lower ability to learn sufficient contextual information. Therefore, nnU-Net uses the median of the Spacing calculated independently for each axis as the final Spacing. In addition, five-fold cross-validation training is used, and 1000 Epochs are run, each Epoch is defined as 250 batches, and each batch has 2 Patches. The optimizer used by nnU-Net is stochastic gradient descent, and the initial learning rate is 0.01.

[0040] 2. Conditional Strategy

[0041] Compared with the classical approach of training separate models for OARs and TRs in LCTSC and MSD LT datasets, respectively, the key issue of “target ambiguity” needs to be addressed when using a joint partially labeled dataset. Consider datasets that label OARs and TRs separately: LCTSC dataset and MSD LT dataset, each of which contains a set of input images and associated labels.

[0042] It should be noted that there is no intersection in the labeled data between LCTSC and MSD LT. This means that the labeled targets are relatively independent in their own datasets. Therefore, if trained directly on the union of LCTSC and MSD LT, the loss function will not converge in the validation dataset because neither the labeled OARs nor the target regions can fully cover the union dataset. When the training data comes from MSD LT, only the target regions are labeled in this dataset, and the OARs will be identified as background. When the input data comes from LCTSC, only the OARs are labeled in this dataset, and the OARs will be treated as foreground. This "target ambiguity" problem may confuse segmentation due to the uncertain connection between the input image and the mask, resulting in poor generalization ability.

[0043] To solve this problem, the present invention designs a dual-path segmentation framework condition nnU-Net, such as Figure 1As shown in the figure, it can simultaneously learn deep features of OARs and target regions. Dual-path segmentation is performed by introducing a conditional strategy, where the conditional information is incorporated into each upsampling process as an additional input layer during upsampling. More specifically, if the input data is sampled from the LCTSC dataset, the predefined auxiliary tensor will be connected to the corresponding decoder to perform training for OAR segmentation. Similarly, when sampling from the MSD LT dataset, the auxiliary information becomes the conditional tensor of the target region. The predefined conditional tensor has the same size as the corresponding downsampled feature map and the output of the upper layer in the upsampling path. The conditionalization strategy allows different segmentation tasks to be performed in parallel and independently of each other.

[0044] 3. Uncertainty-aware self-enhancement model based on semi-supervision

[0045] By analyzing the dataset fingerprints of LCTSC and MSD LT, it is observed that the labeled OACs also appear in the MSD LT dataset without labeled OACs. Therefore, in order to further avoid the "data waste" in the segmentation task of partially labeled joint datasets, the present invention introduces semi-supervised learning into conditional nnU-Net to maximize the value of discrete available datasets. An uncertainty-aware self-enhancement model is used to assist in the segmentation of OACs, which can effectively utilize unlabeled data (images from MSD LT) by encouraging consistent predictions for the same input under different perturbations.

[0046] like Figure 2 As shown in Figure 1, the framework consists of a teacher model and a student model, both of which use nnU-Net as the backbone. The input data has two output paths, one is used as labeled data for target segmentation, and the other is used as unlabeled data for OAR segmentation. In semi-supervised OAR segmentation, the student model learns from the teacher model by minimizing the combined loss and consistency loss relative to the teacher model's target. In addition, the weights of the teacher model are updated by the exponential moving average (EMA) of the student model weights to improve the performance quality during different training processes.

[0047] Since there is no OAR mask in MSD LT, the prediction results of the teacher model may be unreliable and unsatisfactory. Therefore, an uncertainty-aware scheme is adopted to enhance the learning ability of the student model to solve this problem. The forward pass is performed on the teacher model based on unlabeled data and Gaussian noise. The teacher model also estimates the uncertainty feature map of the OAR while generating the prediction mask. The uncertainty estimation generated by the teacher model can effectively guide the student model to learn more reliable targets from the teacher model.

[0048] As for the student model, it is optimized by combining the fully supervised loss and the unsupervised consistency loss. In the supervised target segmentation, the CT image data and masks used for training are randomly sampled from the MSD LT dataset and input into the network training, and the target segmentation model is optimized only by the fully supervised loss.

[0049] 4. Joint loss function

[0050] Different from the loss function in multi-class segmentation, the present invention attempts to design two different loss function combinations for organs at risk and target areas respectively. Among them, for the target area, the Dice loss function and the Focal loss function are jointly used to optimize the task. Among them, for organs at risk, the Dice loss function and the cross entropy loss function are used to optimize the segmentation model. The loss function design here, combined with the use environment of the conditional nnU-Net, can ensure that the dual-path segmentation tasks are optimized separately and do not interfere with each other.

[0051] In the target segmentation path, the supervision loss is defined as:

[0052]

[0053] Among them, the main segmentation task uses Dice loss Auxiliary classification uses Focal loss Optimize. n represents the marked area of ​​the target area, p n Represents the predicted region of the segmentation model. α and γ are Focal loss coefficients, which are taken as 0.2 and 2 here. Since the target area appears in different locations in the lung parenchyma with different intensities and shapes, attached to the chest wall or mediastinum, it leads to poor voxel identification and unbalanced data distribution.

[0054] Therefore, Focal loss is used to participate in the classification task of target area segmentation, because Focal loss supports the network to better focus on learning poorly classified voxels.

[0055] In organ-at-risk segmentation, loss is defined as:

[0056]

[0057] The segmentation loss is based on Dice, the classification loss is replaced by the cross entropy loss function, Π represents the exponential function, p' υ 、p υ and υ denote the results of the teacher model, the student model, and the uncertainty prediction, respectively. H is the threshold for the selected target, which is taken as 0.8. In addition, the uncertainty-aware consistency loss is the voxel-wise mean squared error loss of the teacher and student models, and is only used for the most certain predictions.

[0058] 5. Five-fold crossover experiment verification

[0059] The training set, validation set, and test set used in the present invention are all taken from the LCTSC and MSD LT datasets. In order to enhance the authority and persuasiveness of the experimental data, the joint dataset of LCTSC and MSD LT is randomly divided into 5 subsets, and a five-fold cross-validation is used to conduct a validation test experiment on the intelligent diagnosis method for pulmonary nodules, and the DSC index based on regional evaluation and the HD based on contour evaluation are used as the measurement criteria for analysis and comparison. The training loss is as follows Figure 3 In the test set, the scores of the spinal cord, left and right lungs, heart, esophagus, and target area in Dice were 0.8529, 0.9813, 0.9707, 0.9498, 0.8215, and 0.7789, respectively, and the scores in HD were 1.5269, 1.5268, 2.8992, 0.7258, 1.0538, and 4.0796, respectively.

[0060] 6. Example Diagnostic Test

[0061] After the intelligent segmentation method for lung cancer target area and organs at risk was successfully established, an example verification test was carried out. Figure 4 The CT slices of the organs at risk and target areas presented are selected from the test data sets in the LCTSC and MSD LT data sets, respectively. The displayed DSC and HD are calculated on the current CT slice. The prediction results of the method of the present invention meet the clinical needs of radiologists, and can help radiologists quickly and accurately formulate lung cancer radiotherapy plans in clinical practice, improve the success rate of surgery, and improve the quality of life of patients.

[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An intelligent target and organ-at-risk segmentation method for lung cancer, characterized in that: A common segmentation network framework for lung cancer target areas and organs at risk is used for dual-path segmentation; the common segmentation network framework includes a teacher model and a student model, both of which use nnU-Net as the backbone; For the organs at risk and target areas, predefined auxiliary condition information is given and embedded in the upsampling process, connected to the corresponding decoder in the decoding process, and the predicted masks are fused to obtain the final segmentation result; The target area labeled data and the organ at risk labeled data in the input data are used as labeled data and input into the student model for supervised target area segmentation, and the target area unlabeled data are input into the teacher model for semi-supervised organ at risk segmentation; the student model learns from the teacher model by minimizing the combined loss and consistency loss relative to the teacher model target; the weight of the teacher model is updated by the exponential moving average (EMA) of the student model weight to improve the performance quality in different training processes; The teacher model estimates the uncertainty feature map of the OAR while generating the prediction mask to guide the student model to learn more reliable targets from the teacher model; The student model is optimized by combining the fully supervised loss and the unsupervised consistency loss. In the supervised target segmentation, the CT image data and masks used for training are randomly sampled from the target dataset and input into the network training, and the target segmentation model is optimized only by the fully supervised loss; By using predefined auxiliary condition information and introducing conditional strategies, we can break down the barriers between different labeled targets of different independent data sets, achieve joint learning between different labeled data sets, and introduce a semi-supervised learning mechanism to assist in the segmentation of organs at risk, so that the segmentation tasks of the target area and organs at risk do not interfere with each other, thus achieving joint learning of the target area and organs at risk. The input data is CT images, which come from the LCTSC dataset and the MSD LT dataset. The risk organ labeling data comes from the LCTSC dataset, and the target area labeling data comes from the MSD LT dataset. There is no intersection between the labeling data of LCTSC and MSD LT. In the target segmentation path, the supervision loss is defined as: Among them, the main segmentation task uses Dice loss Auxiliary classification uses Focal loss Optimize, y n represents the marked area of ​​the target area, p n represents the segmentation model prediction area, α and γ are Focal loss coefficients; In organ-at-risk segmentation, loss is defined as: The segmentation loss is based on Dice, and the classification loss is replaced by the cross entropy loss function. Π represents the exponential function, and p′ υ 、p υ and υ They represent the results of teacher model, student model and uncertainty prediction respectively, and H is the threshold of the selected target; The auxiliary condition information is expressed as a condition tensor, which has the same size as the output of the down-sampling at the same layer and the output of the previous up-sampling. The initial values ​​of the condition tensors corresponding to the organs at risk and the target area are not equal. Dual-path segmentation is performed by introducing a conditional strategy. During upsampling, auxiliary conditional information will be merged into each upsampling process as an additional input layer. The conditional strategy is performed during the upsampling process and does not interfere with the downsampling process. The output of the encoder is a shareable feature map.