Radiation pneumonia risk prediction method and system based on federal transfer learning

Abstract

The application discloses a method and system for predicting the risk of radiation pneumonitis based on federal transfer learning, which comprises the following steps: acquiring the planning CT images, treatment information and outcome labels of patients in the source domain who have not received immunotherapy and patients in the target domain who have received immunotherapy; after standardizing the planning CT images, obtaining a lung mask through a three-dimensional segmentation model and calculating a segmentation reliability coefficient; extracting radiomics features and combining the segmentation reliability coefficient to obtain a weighted feature vector; constructing a treatment background vector based on the treatment information, combining the outcome labels and clinical scores to generate soft labels; training a multilayer perceptron model through federal learning to obtain a baseline model, and finally outputting the probability of radiation pneumonitis risk based on the weighted feature vector of the patient to be tested. The method guarantees data privacy while improving the accuracy and transferability of the prediction.

Description

Technical Field

[0001] This application relates to the field of transfer learning technology, and more specifically, to a method and system for predicting the risk of radiation pneumonitis based on federated transfer learning. Background Technology

[0002] Radiation pneumonitis is a common and serious complication in patients undergoing thoracic radiotherapy, especially in clinical settings where thoracic radiotherapy is combined with immunotherapy. The risk of developing radiation pneumonitis is significantly increased, severely impacting treatment outcomes and quality of life. Therefore, accurate prediction of radiation pneumonitis risk in these patients is of significant clinical importance. Currently, radiomics-based machine learning models are widely used in clinical practice for radiation pneumonitis risk prediction, but these models still have many technical limitations in practical applications.

[0003] Existing predictive models are mostly trained on data from patients who have not received immunotherapy. When directly applied to immunotherapy recipients, the differences in the biological mechanisms of lung injury between the two groups can easily lead to distortion of the model's decision boundaries and a significant decrease in predictive performance. Furthermore, multi-center clinical data exhibit differences in scanning equipment and parameters, and automatic lung tissue segmentation is prone to errors, resulting in insufficient stability of radiomics features and further impacting model reliability. In addition, there is attribution overlap between immunotherapy-related pneumonia and radiation-induced pneumonia. When training models using traditional hard labels, uncertain cases are easily included as certain knowledge during training, amplifying the model's transfer distortion problem.

[0004] On the other hand, medical data is subject to privacy protection and cross-institutional sharing restrictions, making it difficult to train multi-center data in a centralized manner. Simple federated learning methods cannot effectively handle the domain shift problem caused by differences in treatment backgrounds, and the small sample size of the immunotherapy population further exacerbates the difficulty of model training. Summary of the Invention

[0005] This application provides a method and system for predicting the risk of radiation pneumonitis based on federated transfer learning, in order to at least solve the technical problems existing in the related technologies described above.

[0006] According to a first aspect of the embodiments of this application, a method for predicting the risk of radiation pneumonitis based on federated transfer learning is provided, including: Acquire planned CT images, treatment information, and radiation pneumonitis outcome labels for patients who did not receive immunotherapy in the source domain nodes and patients who received immunotherapy in the target domain nodes; The standardized CT volume data obtained by standardizing the planned CT images is input into the pre-trained three-dimensional segmentation model to obtain lung masks. Based on the consistency between multiple lung masks obtained by multiple inferences on the same standardized CT volume data, the segmentation confidence coefficient is calculated. Using the lung mask as the region of interest, radiomics features are extracted from the standardized CT volume data, and the radiomics features are weighted based on the segmentation confidence coefficient to obtain a weighted feature vector. A treatment background vector is constructed based on the treatment information, and soft labels are generated based on the radiation pneumonia outcome labels and clinical scoring indicators. Using the weighted feature vector as input and the radioactive pneumonia outcome label as supervision signal, a multilayer perceptron model is trained on the source domain node through federated learning to obtain a baseline model containing a shared basic representation layer and a shared main classifier head; Based on the weighted feature vector of the patient to be tested, the probability of radiation pneumonitis risk is output through the baseline model.

[0007] As an optional approach, the construction of a target domain adaptation model is also included: The baseline model is migrated to the target domain node, and a treatment background conditional encoding layer and a treatment-specific residual head are added to obtain an initial adapted model. The treatment background conditional encoding layer takes the treatment background vector as input, and the input of the treatment-specific residual head is the concatenation result of the output of the shared basic representation layer and the output of the treatment background conditional encoding layer. Using the weighted feature vector and the treatment background vector as inputs, and the soft label as a supervision signal, the initial adaptation model is fine-tuned and trained on the target domain node to obtain the target domain adaptation model; Based on the weighted feature vector of the patient to be tested, the probability of radiation pneumonitis risk is output through the target domain adaptation model.

[0008] As an optional approach, the calculation of the segmentation confidence coefficient based on the consistency among multiple lung masks obtained from multiple inferences on the same standardized CT volume data includes: The original standardized CT volume data, the standardized CT volume data after geometric transformation, and the standardized CT volume data after pixel transformation are respectively input into the three-dimensional segmentation model to obtain three corresponding lung masks. Calculate the ratio of the number of consistent voxels among the three lung masks to the total number of voxels, and use this ratio as the segmentation confidence coefficient.

[0009] As an optional approach, the weighting of the radiomics features based on the segmentation confidence coefficient to obtain a weighted feature vector includes: The sensitivity value of each radiomics feature to changes in the lung mask is calculated to obtain the feature sensitivity vector; For each radiomics feature, the confidence weight of that feature is calculated based on the segmentation confidence coefficient and the sensitivity value of the corresponding dimension in the feature sensitivity vector. The standardized radiomics feature values ​​of each dimension are multiplied by the corresponding confidence weights to obtain the weighted feature vector.

[0010] As an optional approach, the calculation of the sensitivity value of each radiomics feature to changes in the lung mask includes: Perform dilation and erosion operations on the lung mask of the source domain training samples to obtain the dilated mask and the eroded mask respectively; Using the expanded mask and the etched mask as regions of interest, radiomics features are extracted again to obtain the expanded feature set and the etched feature set. Based on the expanded feature set, the eroded feature set, and the original feature set extracted from the original lung mask, the average relative change rate of each feature under the expanded and eroded states relative to the original state is calculated, and this average relative change rate is used as the sensitivity value.

[0011] As an optional approach, generating soft labels based on the radiation-induced pneumonia outcome labels and clinical scoring indicators includes: The target domain samples are scored according to preset clinical scoring indicators, including the onset time window, image distribution location, whether infection is ruled out, and physician attribution consistency. Based on the preset range in which the total score falls, the radiation pneumonia outcome label is converted into a continuous value between 0 and 1, which is used as the soft label.

[0012] As an optional approach, fine-tuning the initial adaptation model on the target domain nodes includes: Construct a total loss function that includes soft label cross-entropy loss and boundary emphasis loss, wherein the boundary emphasis loss is a weighted loss after assigning higher loss weights to samples in the soft labels that indicate attribution uncertainty; The initial adaptation model is trained with the goal of minimizing the total loss function.

[0013] As an optional approach, fine-tuning the initial adaptation model on the target domain nodes includes: In the first stage, while keeping the parameters of the shared basic representation layer unchanged, the shared master classifier, the treatment background condition encoding layer, and the treatment-specific residual head are trained. In the second stage, after the training in the first stage is completed, the parameters of some network layers in the shared basic representation layer are opened for training.

[0014] As an optional approach, if the patient to be tested is a patient who has not received immunotherapy, the weighted feature vector is input into the baseline model, and the baseline risk value output by the shared master classifier in the baseline model is mapped by an activation function as the probability of radiation pneumonitis risk. If the patient to be tested is a patient receiving immunotherapy, then the weighted feature vector and the treatment background vector are input into the target domain adaptation model, and the sum of the baseline risk value output by the shared master classifier and the output of the treatment-specific residual head is mapped by the activation function as the probability of radiation pneumonitis risk.

[0015] According to a second aspect of the embodiments of this application, a radiation pneumonitis risk prediction system based on federated transfer learning is also provided, comprising: The data acquisition module is used to acquire planned CT images, treatment information, and radiation pneumonitis outcome labels of patients who have not received immunotherapy in the source domain nodes and patients who have received immunotherapy in the target domain nodes. The image processing and credibility assessment module is used to input the standardized CT volume data obtained by standardizing the planned CT image into the pre-trained three-dimensional segmentation model to obtain the lung mask, and calculate the segmentation credibility coefficient based on the consistency between multiple lung masks obtained by multiple inferences on the same standardized CT volume data. The feature extraction and weighting module is used to extract radiomics features from the standardized CT volume data with the lung mask as the region of interest, and to weight the radiomics features based on the segmentation confidence coefficient to obtain a weighted feature vector. The tag generation module is used to construct a treatment background vector based on the treatment information and generate soft tags based on the radiation pneumonia outcome tags and clinical scoring indicators. The federated learning training module is used to train a multilayer perceptron model on the source domain node by federated learning, using the weighted feature vector as input and the radioactive pneumonia outcome label as supervision signal, to obtain a baseline model containing a shared basic representation layer and a shared main classifier head. The risk prediction module is used to output the probability of radiation pneumonitis risk based on the weighted feature vector of the patient to be tested, through the baseline model.

[0016] This application effectively addresses the problems of insufficient feature stability, model transfer distortion, and training errors caused by overlapping attributions in existing technologies through an innovative design that incorporates segmentation confidence weighting, treatment background condition encoding, and hierarchical training of uncertain samples with soft labels. By combining multi-center federated learning and transfer learning, joint training is achieved without sharing original patient data, ensuring data privacy and overcoming sample size limitations. The weighting mechanism of segmentation confidence and feature sensitivity enhances the stability of radiomics features; the treatment background vector and dedicated network layers adapt to the mechanistic differences among different treatment populations; and soft labels and hierarchical loss functions optimize the training effect for uncertain samples. This significantly improves the accuracy, stability, and transferability of radiation pneumonitis risk prediction in the context of combined chest radiotherapy and immunotherapy.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Furthermore, no embodiment in this disclosure is required to achieve all the effects described above. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0019] Figure 1 This is a schematic diagram of a radiation pneumonitis risk prediction method based on federated transfer learning, provided as an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the segmentation confidence coefficient calculation process provided in the embodiments of this disclosure.

[0021] Figure 3 A schematic diagram of the baseline model construction process provided in the embodiments of this disclosure.

[0022] Figure 4 This diagram illustrates the construction and training of the target domain adaptation model provided in the embodiments of this disclosure.

[0023] Figure 5 This is a schematic diagram of the structure of a radiation pneumonitis risk prediction system based on federated transfer learning provided in an embodiment of this application. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0025] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, collected data, used data, generated data, processed data, etc.) and the data (including but not limited to data used for analysis, stored data, displayed data, collected information, used information, generated information, processed information, etc.) are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0026] This method is applicable to the prediction of radiation pneumonitis risk in patients undergoing thoracic radiotherapy, and is particularly suitable for the scenario of radiation pneumonitis risk prediction in patients receiving thoracic radiotherapy combined with immunotherapy. It is executed collaboratively by a central server and multiple local nodes in medical institutions. Each local node pre-stores digital imaging and communication medical format planned computed tomography images, clinical treatment information, and radiation pneumonitis follow-up outcome labels of its patients. The central server establishes a stable federated communication connection with each local node, and sets the node of patients who have not received immunotherapy as the source domain node and the node of patients who have received thoracic radiotherapy combined with immunotherapy as the target domain node.

[0027] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0028] Figure 1 This is a flowchart of a radiation pneumonitis risk prediction method based on federated transfer learning according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes steps S1-S6: In step S1, the planned CT images, treatment information, and radiation pneumonitis outcome labels of patients who did not receive immunotherapy in the source domain node and patients who received immunotherapy in the target domain node are obtained.

[0029] Each source domain node and target domain node extracts the corresponding patient's planned computed tomography (CT) images, clinical treatment information, and radiation pneumonitis outcome labels from their local radiotherapy information system, image archiving system, and electronic medical record system, respectively. The planned CT images are localization scans of the patient before receiving chest radiotherapy. The clinical treatment information includes structured treatment data such as the patient's radiotherapy plan, chemotherapy plan, and immunotherapy plan. The radiation pneumonitis outcome label is an identifier indicating the occurrence of radiation pneumonitis based on the patient's post-radiotherapy follow-up results.

[0030] For example, source domain nodes extract data from patients with non-small cell lung cancer (NSCLC) or small cell lung cancer who have not received immunotherapy, while target domain nodes extract data from patients with NSCLC or small cell lung cancer who have received thoracic radiotherapy combined with immunotherapy. During the extraction process, preliminary quality screening is performed on the data to exclude patients with substandard image quality, missing treatment information, or insufficient follow-up time, ensuring the effectiveness of the data for subsequent model training. Each node then structures the extracted patient data by individual, forming a one-to-one corresponding patient dataset.

[0031] In step S2, the standardized CT volume data obtained by standardizing the planned CT image is input into the pre-trained three-dimensional segmentation model to obtain lung masks, and the segmentation confidence coefficient is calculated based on the consistency between multiple lung masks obtained by multiple inferences on the same standardized CT volume data.

[0032] Figure 2 A schematic diagram illustrating the segmentation confidence coefficient calculation process provided in an embodiment of this disclosure is shown. Figure 2 As shown, in step S201, each node performs a unified preprocessing operation on the acquired planning computed tomography (CT) images. This unified preprocessing eliminates image differences caused by different scanning devices and parameters, resulting in standardized CT volumetric data. Specifically, each node first reads the planning CT images in digital imaging and communication medicine format, stacks and reconstructs the two-dimensional tomographic images according to the patient dimension, constructing three-dimensional volumetric data that can fully characterize the spatial structural information of the patient's chest.

[0033] Subsequently, voxel resampling is performed on the 3D volume data using linear interpolation to unify the voxel size to a preset, uniform specification, exemplarily configurable as 1×1×1 cubic millimeters, ensuring consistent voxel resolution across different patients. After voxel resampling, the voxel intensity of the 3D volume data is cropped, limiting the voxel intensity values ​​to a preset HU range, exemplarily configurable as -1000 to 400 HU. This range effectively highlights the texture features of lung tissue and suppresses interference from other tissues. Finally, linear normalization is performed on the cropped voxel intensity 3D volume data, mapping the voxel intensity values ​​to the 0-1 range, further eliminating intensity differences in the images.

[0034] In step S202, multiple lung masks are obtained through multiple inferences based on the 3D segmentation model. In this embodiment, a 3D convolutional neural network model is used as the 3D segmentation model. This model is a classic encoder-decoder network structure. The encoder side sequentially includes a convolution module, a normalization module, and a downsampling module. The convolution module uses convolution kernels for feature extraction, the normalization module performs batch normalization on the extracted features, and the downsampling module uses pooling to reduce the dimensionality of the feature map and expand the receptive field. The decoder side sequentially includes an upsampling module, a stitching module, and a convolution module. The upsampling module uses deconvolution to increase the dimensionality of the feature map, the stitching module stitches and fuses the feature maps of the corresponding layers at the encoder and decoder, and the convolution module performs convolution processing on the fused features to achieve refined feature extraction. The last layer of the 3D segmentation model uses a Sigmoid activation function to map the model output to the range of 0 to 1, outputting a bi-lung probability map with the same voxel size as the input standardized CT volume data. The value of each voxel in the bi-lung probability map represents the probability that the voxel belongs to lung tissue.

[0035] Optionally, source domain training samples with manually labeled lung masks are selected as the training set. For example, hundreds of samples can be selected to form the training set. Standardized CT volume data is used as input, and manually labeled lung masks are used as supervision signals to pre-train the constructed 3D convolutional neural network model. During the pre-training process, the binary cross-entropy loss function is used to calculate the loss between the model's predicted values ​​and the manually labeled values. An adaptive moment estimation optimizer is used to iteratively update the model parameters. Hyperparameters such as learning rate, batch size, and number of training epochs can be adjusted according to the model's training effect until the model's segmentation accuracy on the validation set reaches the preset requirements. The pre-trained 3D segmentation model is then obtained and deployed to each source domain node and target domain node for automatic lung tissue segmentation inference.

[0036] In practical implementation, for any patient's standardized CT volume data, each node performs three segmentation inference operations using a pre-trained 3D segmentation model to obtain multiple lung masks. The first inference directly inputs the original standardized CT volume data into the 3D segmentation model to obtain the first bilateral lung probability map. The second inference performs a geometric transformation on the standardized CT volume data; in this embodiment, a left-right flip geometric transformation is used. The transformed standardized CT volume data is then input into the 3D segmentation model to obtain the second bilateral lung probability map. The third inference performs a pixel transformation on the standardized CT volume data; in this embodiment, a pixel transformation with mild Gaussian noise enhancement is used. For example, the mean of the Gaussian noise can be configured to 0, and the variance can be configured to 0.01. This adds subtle differences to the image while ensuring that lung tissue features are not destroyed. The transformed standardized CT volume data is then input into the 3D segmentation model to obtain the third bilateral lung probability map.

[0037] The three bi-lung probability maps obtained from the three inferences are binarized to obtain three corresponding lung masks. The threshold for binarization can be configured to 0.5, meaning that when the probability value of a voxel in the bi-lung probability map is greater than 0.5, the voxel is identified as lung tissue and marked as 1; when the probability value of a voxel is less than or equal to 0.5, the voxel is identified as non-lung tissue and marked as 0. After binarization, each bi-lung probability map is converted into a binary lung mask composed of 0s and 1s, which can clearly represent the spatial location of lung tissue in the standardized CT volume data.

[0038] In step S203, the segmentation confidence coefficient is obtained by calculating the ratio of the number of consistent voxels among the three lung masks to the total number of voxels. Specifically, the segmentation confidence coefficient is a quantitative index characterizing the consistency among the three lung masks, with a value ranging from 0 to 1. This index is obtained by calculating the ratio of the number of consistent voxels among the three lung masks to the total number of voxels. Specifically, the three binary lung masks are compared point by point according to voxel positions, and the number of voxels with consistent marking results at the same voxel positions is counted and recorded as the number of consistent voxels; the total number of face voxels for each lung mask is counted. Since the voxel sizes of the three lung masks are consistent, the total number of voxels is a fixed value. The ratio of the number of consistent voxels to the total number of voxels is defined as the segmentation confidence coefficient, and the calculation formula is:

[0039] In the formula, For the segmentation confidence coefficient, The number of voxels is consistent across the three lung masks. The total number of elements in a single lung mask.

[0040] Among them, the magnitude of the segmentation confidence coefficient is positively correlated with the consistency of the segmentation results. The closer the value is to 1, the higher the consistency of the results of the three segmentation inferences, and the more reliable the segmentation results of the segmentation model for this sample. The closer the value is to 0, the lower the consistency of the three segmentation inference results and the weaker the reliability of the segmentation results. After the calculation is completed, each patient's standardized CT volume data corresponds to a lung mask and a segmentation confidence coefficient, which together serve as inputs for subsequent radiomics feature extraction and weighted processing.

[0041] In step S3, taking the lung mask as the region of interest, radiomics features are extracted from the standardized CT volume data, and the radiomics features are weighted based on the segmentation confidence coefficient to obtain a weighted feature vector.

[0042] Specifically, each node uses the lung mask it acquires as the region of interest (ROI) to extract radiomics features from standardized CT volume data. The ROI limits the scope of feature extraction, ensuring that all extracted features are characteristic information of lung tissue. In this embodiment, a radiomics feature extraction tool is used to extract pre-defined dimensions of handcrafted radiomics features from standardized CT volume data. For example, it can be configured to 107 dimensions. The extracted features include first-order statistical features, shape features, and texture features. The texture features cover various types such as gray-level co-occurrence matrix, gray-level run length matrix, gray-level size region matrix, neighborhood gray-level difference matrix, and gray-level dependency matrix. The extraction of all types of features follows the standard process of radiomics feature extraction, which can characterize the imaging features of lung tissue from different perspectives.

[0043] Furthermore, the feature sensitivity vector is a vector that characterizes the sensitivity of each radiomics feature to changes in the lung mask. Its dimension is consistent with the extracted radiomics feature dimension, and the value of each dimension is the sensitivity value of the corresponding radiomics feature. The sensitivity value is obtained by calculating the average relative change rate of the feature under lung mask expansion, corrosion and original state.

[0044] Specifically, an expansion operation and an erosion operation are performed on the lung mask of each sample in the source domain training set beforehand. The expansion operation expands the boundary of the lung mask outward by 1 voxel, and the erosion operation shrinks the boundary of the lung mask inward by 1 voxel, resulting in an expanded mask and an eroded mask.

[0045] Using the expanded and eroded masks as regions of interest, radiomics features consistent with the original feature dimensions were re-extracted from the corresponding standardized CT volume data, resulting in expanded and eroded feature sets. For each radiomics feature dimension, the relative rate of change of that feature dimension with respect to the original state in the expanded and eroded states was calculated. The absolute values ​​of the two relative rates were then taken, and the average value was calculated. This average value is the sensitivity value of that feature dimension, calculated using the following formula:

[0046] In the formula, For the first Sensitivity values ​​for radiomics characteristics For the first Extracted values ​​of dimensional features under the original lung mask. For the first Extracted values ​​of dimensional features under the dilated mask. For the first Extracted values ​​of dimensional features under the etched mask.

[0047] The sensitivity value calculation operation described above is performed on all dimensions of radiomics features to obtain a feature sensitivity vector. This feature sensitivity vector is then normalized. A minimum-maximum normalization method can be used to map the sensitivity values ​​of each dimension to the interval between 0 and 1. The normalized feature sensitivity vector is denoted as... The magnitude of each dimension's value is positively correlated with the sensitivity of the corresponding feature to changes in the lung mask.

[0048] Furthermore, to eliminate the dimensional differences in radiomics features among different patients, the extracted raw radiomics features are standardized. In this embodiment, Z-score standardization is used, and the standardization calculation only uses the feature data of the source domain training samples to avoid interference from the target domain samples on the standardization parameters, thus ensuring the rationality of the standardization.

[0049] Specifically, firstly, the mean and standard deviation of each dimension of radiomics features in the source domain training samples are calculated. For the th... 3D features, whose mean is denoted as The standard deviation is denoted as Then, standardize the radiomics features of each dimension for all samples at all nodes according to the following formula:

[0050] In the formula, For the sample number Raw values ​​of radiomics characteristics. For the sample number The standardized values ​​of the dimensional features For the source domain training sample, the first The mean of the dimensional features, For the source domain training sample, the first The standard deviation of the dimensional features. After Z-score standardization, the radiomics features of all samples follow a standard normal distribution with a mean of 0 and a variance of 1.

[0051] In one embodiment, to achieve weighted processing of radiomics features based on segmentation confidence, a confidence weight is defined for each dimension of radiomics features. The confidence weight is a quantitative index obtained by combining the segmentation confidence coefficient and the feature sensitivity value, and its value ranges from 0 to 1. The calculation formula is as follows:

[0052] In the formula, For the first Confidential weights of radiomics features For the segmentation confidence coefficient, For the first Sensitivity value after normalization of dimensional features.

[0053] It can be seen from the formula that when the segmentation confidence coefficient The larger the value, the stronger the reliability of the lung segmentation result, corresponding to the higher the credibility weight of the feature. The larger the feature sensitivity value, the better. The larger the value, the more sensitive the feature is to small changes in the lung mask, corresponding to a higher confidence weight for the feature. The smaller the value, the better the calculation logic can suppress the amplitude of features that are sensitive to segmentation fluctuations and have low segmentation confidence, while retaining higher weights for robust features that are not sensitive to segmentation and have high segmentation confidence. The confidence weight ranges from 0 to 1, and this range is determined by the ranges of the segmentation confidence coefficient and the feature sensitivity value, both of which are values ​​between 0 and 1. Therefore, the calculated confidence weight also falls within the range of 0 to 1, requiring no additional normalization processing.

[0054] After obtaining the confidence weights for each dimension of radiomics features, the standardized radiomics feature values ​​are multiplied dimension by dimension by their corresponding confidence weights to obtain the weighted feature vector. The calculation formula is as follows:

[0055] In the formula, The first eigenvector in the weighted eigenvector The numerical value of the dimension. For the first The standardized values ​​of the dimensional features For the first The dimensionality of the weighted feature vector is consistent with that of the original radiomics features. It integrates the reliability information of lung segmentation and the sensitivity information of the features, which can effectively improve the stability and effectiveness of the features.

[0056] In step S4, a treatment background vector is constructed based on the treatment information, and soft labels are generated based on the radiation pneumonia outcome label and clinical scoring indicators.

[0057] Specifically, each node reads the structured clinical treatment information of each patient from the local treatment planning system and electronic medical record system, and constructs a treatment background vector based on the treatment information. This vector is a one-dimensional numerical vector that can structurally represent the characteristics of the patient's treatment plan. The feature dimensions it includes at least include whether the patient received immunotherapy, the timing of immunotherapy administration, whether induction chemotherapy was performed, whether concurrent chemotherapy was performed, the planned target volume, lung V5, lung V20, and average lung dose. Lung V5 is the percentage of lung tissue volume with an irradiation dose of ≥5 Gy to the total lung volume, lung V20 is the percentage of lung tissue volume with an irradiation dose of ≥20 Gy to the total lung volume, and average lung dose is the average irradiation dose of the entire lung tissue.

[0058] For binary features in the treatment background vector, a 0-1 encoding method is used for numerical processing. For example, in the feature of whether immunotherapy was received, 0 indicates that immunotherapy was not received, and 1 indicates that immunotherapy was received. The encoding method for the features of whether chemotherapy was induced and whether chemotherapy was concurrent is similar. For continuous features, for example, the planned target volume, lung V5, lung V20, and mean lung dose are directly used as the actual values ​​obtained from clinical measurements. If there are a few missing values, they can be filled by the sample mean of the source domain or target domain, or by linear interpolation. For temporal features, for example, the immunotherapy administration sequence is encoded according to the order of immunotherapy administration and chest radiotherapy. For example, administration before radiotherapy is recorded as 1, administration concurrently with radiotherapy is recorded as 2, and administration after radiotherapy is recorded as 3.

[0059] The dimensions of the treatment background vector can be flexibly expanded according to the actual situation of clinical treatment. For example, features such as chemotherapy drug type, total radiotherapy dose, and number of radiotherapy fractions can be added, but this application does not limit this. The constructed treatment background vector can effectively quantify the differences in treatment regimens between people who have not received immunotherapy and those who have received immunotherapy.

[0060] For samples in the source domain that have not received immunotherapy, hard labels are used for labeling. Hard labels are binary labels. For example, if grade ≥2 radiation pneumonitis is defined as a positive event under the Adverse Event General Terminology Standard v4.0, samples with grade ≥2 radiation pneumonitis are labeled as 1, and samples without grade ≥2 radiation pneumonitis are labeled as 0. Hard labels serve as supervision signals for federated pre-training of the source domain model.

[0061] Furthermore, for immunotherapy-receiving samples from the target domain node, due to the overlap in attribution between immunotherapy-related pneumonia and radiation-induced pneumonia, soft labels are used for marking. Soft labels are continuous values ​​between 0 and 1, generated by scoring the target domain samples using preset clinical scoring indicators. The clinical scoring indicators at least include the onset time window, image distribution location, whether infection is ruled out, and physician attribution consistency. In this embodiment, a significant field distribution indicator may also be added, forming five scoring indicators. Each indicator is an independent scoring item. Scoring items that meet the criteria for supporting radiation-induced pneumonia receive 1 point, while those that do not receive 0 points. The total score for the five scoring items ranges from 0 to 5 points.

[0062] Based on the preset interval of the total score, the radiation pneumonitis outcome label is converted into a corresponding soft label value. The preset interval is set according to the actual attribution certainty in clinical practice. For example, when the total score is in the interval of 4 to 5 points, it indicates that the sample is a definite radiation pneumonitis sample, and the soft label is configured as 0.9; when the total score is in the interval of 2 to 3 points, it indicates that the sample is an attribution uncertain sample, with overlap in attribution between immunotherapy-related pneumonia and radiation pneumonitis, and the soft label is configured as 0.6; when the total score is in the interval of 0 to 1 points, it indicates that the sample is a definite non-radiation pneumonitis sample, and the soft label is configured as 0.1. The value of the soft label can characterize the attribution certainty of the sample. The closer the value is to 0 or 1, the higher the attribution certainty. The closer the value is to 0.5, the higher the attribution uncertainty. The generated soft labels serve as supervision signals for fine-tuning the training of the target domain model.

[0063] In step S5, using the weighted feature vector as input and the radioactive pneumonia outcome label as supervision signal, a multilayer perceptron model is trained on the source domain node through federated learning to obtain a baseline model containing a shared basic representation layer and a shared master classifier head.

[0064] Each node takes a weighted feature vector as input and a hard label of radiation pneumonitis outcome as supervision signal. A multilayer perceptron model is trained on the source domain nodes through federated learning to obtain a baseline model containing a shared basic representation layer and a shared master classifier head. This process enables joint training of multi-center source domain data, and only the model parameter update is transmitted during training, without transmitting the patient's original image data and feature data, effectively protecting the patient's data privacy. It is suitable for application scenarios where multi-center data cannot be centrally shared.

[0065] Specifically, Figure 3 A schematic diagram of the baseline model construction process provided in an embodiment of this disclosure is shown. For example... Figure 3As shown, in step S301, a multilayer perceptron model is constructed. The central server constructs a multilayer perceptron model, which consists of three parts connected in series: a shared basic representation layer, a shared master classifier head, and a treatment-specific residual head. During the federated learning training phase, only the shared basic representation layer and the shared master classifier head are trained, while the treatment-specific residual head remains in its parameter initialization state and does not participate in the training.

[0066] The shared basic representation layer is the feature extraction and representation learning module of the model. Its input is a weighted feature vector. For example, if the weighted feature vector is 107-dimensional, then the input dimension of the shared basic representation layer is 107-dimensional. This layer contains two fully connected layers. The output dimension of the first fully connected layer is configured with a preset value, for example, 64-dimensional. The activation function is a linear rectified function to achieve non-linear transformation and dimensionality reduction of the features. The input of the second fully connected layer is the output of the first fully connected layer, and its output dimension is configured with another preset value, for example, 32-dimensional. The activation function is also a linear rectified function. The output of the second fully connected layer is defined as the shared representation vector, which is the result of high-level representation learning of the weighted feature vector and serves as the input of the shared main classifier head.

[0067] The shared master classifier is the risk prediction module of the model. Its input is the shared representation vector output by the shared basic representation layer. This layer contains a fully connected layer and outputs a baseline risk value with a dimension of 1. After passing through an activation function such as the Sigmoid function, the baseline risk value is mapped to the interval between 0 and 1 and defined as the baseline risk probability, which represents the baseline risk of patients developing grade ≥2 radiation pneumonitis, thus achieving binary classification prediction of radiation pneumonitis risk.

[0068] In step S302, the multilayer perceptron model undergoes parameter initialization and is distributed to each source domain node. The central server initializes the parameters of the constructed multilayer perceptron model to obtain the initial global parameters of the model, and then distributes the initial global parameters to each source domain node, initiating the federated learning training loop. After receiving the global parameters of the model, each source domain node uses its local weighted feature vector as input and the hard label of the radiation pneumonia outcome as the supervision signal to train the model locally. The number of local training epochs is configured as a preset number of epochs, for example, 5 epochs. During training, the binary cross-entropy loss function is used to calculate the loss between the model's predicted value and the hard label, and an adaptive moment estimation optimizer is used to iteratively update the model parameters. The learning rate of the optimizer is configured as a preset value, for example, 1×10^-3, and the batch size is configured as a preset value, for example, 16.

[0069] In step S303, the parameter update amount and sample size are uploaded to the central server. After each source domain node completes local training, it calculates the update amount of the model parameters and simultaneously counts the number of samples participating in the local training. Only the parameter update amount and sample size are uploaded to the central server; no original patient data is transmitted. After receiving the parameter update amount and sample size uploaded by each source domain node, the central server performs a weighted aggregation of the parameter update amount according to the federated average rule. The weight is the ratio of the sample size of each source domain node to the total sample size of all source domain nodes. After aggregation, a new round of global model parameters is obtained.

[0070] In step S304, the central server distributes the new round of global parameters to each source domain node. Each source domain node updates its local model parameters using the new global parameters, and the training continues until the preset number of federated communication rounds is reached. Each source domain node updates its local model parameters using the new global parameters and then begins the next round of local training, repeating the above cycle of local training, parameter uploading, global aggregation, and parameter distribution until the preset number of federated communication rounds is reached. For example, this can be configured to 50 rounds, or the global validation loss decreases for a consecutive preset number of rounds if the decrease is less than a preset threshold. For example, if the decrease is less than 0.001 for 5 consecutive rounds, federated learning training stops. After training stops, the central server outputs the source domain federated baseline model, which includes the trained shared basic representation layer parameters and shared main classifier head parameters.

[0071] In some embodiments, the method also includes the construction and training of a target domain adaptation model. Specifically, the baseline model is transferred to the target domain nodes, and an initial adaptation model is obtained by adding network layers. Then, the initial adaptation model is fine-tuned and trained using weighted feature vectors and treatment background vectors as inputs and soft labels as supervision signals to obtain the target domain adaptation model. This process enables the model to transfer and adapt from the source domain to the target domain to the target domain to the target domain to the target domain to the target domain to the target domain to the target domain to the target domain. This solves the technical problem that federated learning alone is insufficient to handle differences in treatment backgrounds, and enables the model to better adapt to the radiation pneumonitis risk prediction needs of the target domain to the target domain to the target domain.

[0072] Figure 4 This diagram illustrates the construction and training of the target domain adaptation model provided in an embodiment of this disclosure. Figure 4 As shown, in step S401, an initial adaptation model is constructed based on the parameters of the baseline model. The central server distributes all the trained baseline model parameters to each target domain node. Based on the parameters of the baseline model, an initial adaptation model is constructed. The initial adaptation model adds a treatment context condition encoding layer to the baseline model. The whole model consists of a shared basic representation layer, a shared main classifier head, a treatment context condition encoding layer, and a treatment-specific residual head.

[0073] The input to the treatment background conditional encoding layer is the constructed treatment background vector. This layer contains a fully connected layer, and the output dimension of the fully connected layer is configured to a preset value, for example, it can be configured to 8 dimensions. The activation function is a linear rectified function. The output of this layer is defined as a conditional vector, which realizes feature encoding and dimensionality reduction of the treatment background vector and captures the key feature information of the treatment background.

[0074] The input to the treatment-specific residual head is the concatenation of the shared representation vector output from the shared basic representation layer and the conditional vector output from the treatment background conditional encoding layer. The dimension of the concatenated vector is the sum of the dimensions of the shared representation vector and the conditional vector. For example, if the shared representation vector is 32-dimensional and the conditional vector is 8-dimensional, the concatenated vector will be 40-dimensional. This layer contains a fully connected layer with a 1-dimensional output dimension and no activation function. Its output is defined as the residual risk value, representing the bias in the risk of radiation pneumonitis caused by differences in the treatment background.

[0075] The final risk output of the initial adaptation model is jointly determined by the output of the shared master classifier and the output of the treatment-specific residual classifier. The baseline risk value output by the shared master classifier is added to the residual risk value output by the treatment-specific residual classifier, and the sum is then input into a sigmoid activation function to map the output value to the interval between 0 and 1, yielding the final radiation pneumonitis risk probability. After the initial adaptation model is constructed, a parameter freezing strategy is adopted, freezing all parameters of the shared basic representation layer and only allowing training of the parameters of the shared master classifier, the treatment context conditional encoding layer, and the treatment-specific residual classifier. This strategy enables the model to quickly adapt to the treatment context features of the target domain while preserving the general feature representations learned by the source domain model.

[0076] In step S402, each target domain node uses a combination of weighted feature vectors and treatment background vectors as input, and soft labels as supervision signals to fine-tune the initial adaptation model. During training, a total loss function is constructed, including soft label cross-entropy loss and boundary emphasis loss. The model is trained with the goal of minimizing the total loss function, and a phased training strategy is employed to avoid overfitting in target domain scenarios with few samples.

[0077] The total loss function consists of two parts: soft-label cross-entropy loss and boundary emphasis loss, calculated as follows:

[0078] In the formula, For the total loss function, For soft-label cross-entropy loss, Emphasizing losses at the boundaries.

[0079] Soft-label cross-entropy loss is used to make the model's final risk output as close as possible to the soft label. It is suitable for supervised training with continuous soft labels, and its calculation formula is:

[0080] In the formula, This represents the number of samples used in the batch training. For the first The soft label value of each sample. For the first The final risk output value of the model for each sample.

[0081] Boundary Emphasis Loss is used to give extra attention and learning to samples with uncertain attribution. Different loss weights are assigned to samples with different attribution certainty. Samples with a soft label of 0.6 are defined as samples with uncertain attribution, and the loss weight for this type of sample is set to a preset multiple, for example, 1.5 times that of definite samples. definite samples of radioactive pneumonia with a soft label of 0.9 and definite samples of non-radioactive pneumonia with a soft label of 0.1 are defined as definite samples, and the loss weight for this type of sample is 1. Boundary Emphasis Loss is a weighted sum of the soft label cross-entropy loss under the sample weights, which can amplify the loss for samples with uncertain attribution, making the model pay more attention to the fitting effect of this type of sample during training.

[0082] Specifically, the fine-tuning training adopts a phased training strategy, dividing the training process into a first phase of basic fine-tuning training and a second phase of finite fine-tuning training. The training optimizer adopts an adaptive moment estimation optimizer, and the learning rate is configured to a preset value, for example, 1×10^-4. This learning rate is lower than the learning rate of source domain federated learning, which can stabilize the model adaptation process. The batch size is configured to a preset value, for example, 16.

[0083] In the first stage of basic fine-tuning training, the number of epochs is configured to a preset number of epochs, for example, 10 epochs. In this stage, the parameters of the shared basic representation layer are frozen, and only the parameters of the shared main classifier head, the treatment background conditional encoding layer, and the treatment-specific residual head are trained, so that the model can quickly capture the treatment background features of the target domain and initially construct the treatment background conditional decision boundary.

[0084] After completing the first stage of training, the second stage of limited fine-tuning training begins. In this stage, the model's loss on the validation set within the target domain is continuously monitored. When the validation loss no longer decreases, the parameters of the second fully connected layer in the shared basic representation layer are opened for training, while the first fully connected layer of the shared basic representation layer remains frozen. This allows for limited fine-tuning of the feature representation layer, enabling the model to capture the specific features of the target domain's immunotherapy population without destroying the general feature representation of the source domain.

[0085] Throughout the fine-tuning training process, an early stopping strategy is employed to prevent model overfitting. If the area under the curve of the model on the validation set does not improve for a preset number of consecutive epochs, for example, if there is no improvement for 3 consecutive epochs, training is stopped immediately, and the best model parameters during the training process are saved to obtain the target domain-fitted model.

[0086] In step S6, based on the weighted feature vector of the patient to be tested, the risk probability of radiation pneumonitis is output through the baseline model. In some embodiments, the acquisition of the weighted feature vector of the patient to be tested is performed in accordance with the aforementioned steps, and will not be repeated in this embodiment. Based on the planned computed tomography images and clinical treatment information of the patient to be tested, the same steps as described above are followed to obtain the corresponding weighted feature vector. If the patient to be tested is a recipient of immunotherapy, the corresponding treatment background vector also needs to be obtained. The corresponding model is selected for inference based on the treatment background of the patient to be tested, and the risk probability of radiation pneumonitis is output. At the same time, the reliability of the prediction result is rated based on the segmentation reliability coefficient.

[0087] Specifically, based on the treatment background of the patient being tested, a corresponding model is selected for inferring the probability of radiation pneumonitis risk. If the patient being tested is from a population that has not received immunotherapy, the processed weighted feature vector is input into the baseline model. The shared basic representation layer of the baseline model performs feature representation on the weighted feature vector. The shared master classifier outputs the baseline risk value based on the shared representation vector, which is then mapped by an activation function to obtain the baseline risk probability. This baseline risk probability is used as the probability of the patient developing grade ≥2 radiation pneumonitis.

[0088] If the patient to be tested is a person receiving immunotherapy, the weighted feature vector and the treatment background vector obtained from the processing are jointly input into the target domain adaptation model. The shared basic representation layer of the model outputs a shared representation vector, and the treatment background conditional encoding layer outputs a conditional vector. The two are concatenated and input into the treatment-specific residual head to obtain the residual risk value. The baseline risk value output by the shared main classification head is also obtained. The baseline risk value and the residual risk value are added together and then mapped through the Sigmoid activation function to obtain the final risk probability. This final risk probability is used as the risk probability of the patient developing grade ≥2 radiation pneumonitis.

[0089] Based on the segmentation confidence coefficient of the patient, the output radiation pneumonitis risk probability is rated for confidence. Three threshold ranges for confidence rating are preset. These threshold ranges are set according to the distribution characteristics of the segmentation confidence coefficient and actual clinical needs. For example, a prediction result with a segmentation confidence coefficient greater than or equal to 0.90 is recorded as high confidence, indicating that the patient's lung tissue segmentation results are highly consistent, the feature weighting processing is effective, and the corresponding risk prediction result is highly reliable. A prediction result with a segmentation confidence coefficient greater than or equal to 0.85 and less than 0.90 is recorded as medium confidence, indicating that the patient's lung tissue segmentation results are relatively consistent, and the risk prediction result is moderately reliable. A prediction result with a segmentation confidence coefficient less than 0.85 is recorded as low confidence, indicating that the patient's lung tissue segmentation results are less consistent, the risk prediction result is less reliable, and a comprehensive evaluation combined with manual interpretation by a clinician is required.

[0090] Optionally, the final inference output is a structured report, which may include the patient number, the probability of radiation pneumonitis risk of grade ≥ 2, the risk classification result, and the prediction confidence level. The risk classification result can be determined according to a preset risk threshold. For example, the risk probability threshold is configured to 0.5. Patients with a risk probability greater than or equal to 0.5 are identified as high-risk patients, and patients with a risk probability less than 0.5 are identified as low-risk patients. This report can provide clinicians with a quantitative reference for radiation pneumonitis risk in their treatment decisions.

[0091] Therefore, this method achieves accurate prediction of radiation pneumonitis risk in patients undergoing thoracic radiotherapy, especially those receiving thoracic radiotherapy combined with immunotherapy, through planned CT image preprocessing, lung mask acquisition and segmentation confidence coefficient calculation, radiomics feature extraction and weighting, treatment background vector construction and soft label generation, federated learning baseline model training, and target domain adaptation model construction and training. By introducing segmentation confidence weighting, treatment background condition encoding, and graded training of uncertain soft label samples, it effectively solves technical problems such as decreased feature stability due to segmentation errors, model transfer distortion due to treatment background differences, and hard label training distortion due to attribution overlap. Furthermore, by combining federated learning and transfer learning, it achieves joint training of multi-center data without sharing the original patient data, improving the model's prediction accuracy, stability, and transferability in immunotherapy recipients.

[0092] Please see Figure 5 , Figure 5 This is a schematic diagram of a radiation pneumonitis risk prediction system based on federated transfer learning provided in an embodiment of this application. As shown in the figure, the system includes: The data acquisition module 501 is used to acquire planned CT images, treatment information, and radiation pneumonitis outcome labels of patients who have not received immunotherapy in the source domain node and patients who have received immunotherapy in the target domain node. The image processing and credibility assessment module 502 is used to input the standardized CT volume data obtained by standardizing the planned CT image into the pre-trained three-dimensional segmentation model to obtain the lung mask, and calculate the segmentation credibility coefficient based on the consistency between multiple lung masks obtained by multiple inferences on the same standardized CT volume data. The feature extraction and weighting module 503 is used to extract radiomics features from the standardized CT volume data with the lung mask as the region of interest, and to perform weighting processing on the radiomics features based on the segmentation confidence coefficient to obtain a weighted feature vector. The tag generation module 504 is used to construct a treatment background vector based on the treatment information and generate soft tags based on the radiation pneumonia outcome tags and clinical scoring indicators. The federated learning training module 505 is used to train a multilayer perceptron model on the source domain node by federated learning, using the weighted feature vector as input and the radioactive pneumonia outcome label as supervision signal, to obtain a baseline model containing a shared basic representation layer and a shared main classifier head. The risk prediction module 506 is used to output the probability of radiation pneumonitis risk based on the weighted feature vector of the patient to be tested, through the baseline model.

[0093] Those skilled in the art will clearly understand that the technical solutions of the embodiments of this application can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware that can independently complete or cooperate with other components to complete a specific function, wherein the hardware may be, for example, a field-programmable gate array (FPGA), an integrated circuit (IC), etc.

[0094] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A method for predicting the risk of radiation pneumonitis based on federated transfer learning, characterized in that, include: Acquire planned CT images, treatment information, and radiation pneumonitis outcome labels for patients who did not receive immunotherapy in the source domain nodes and patients who received immunotherapy in the target domain nodes; The standardized CT volume data obtained by standardizing the planned CT images is input into the pre-trained three-dimensional segmentation model to obtain lung masks. Based on the consistency between multiple lung masks obtained by multiple inferences on the same standardized CT volume data, the segmentation confidence coefficient is calculated. Using the lung mask as the region of interest, radiomics features are extracted from the standardized CT volume data, and the radiomics features are weighted based on the segmentation confidence coefficient to obtain a weighted feature vector. A treatment background vector is constructed based on the treatment information, and soft labels are generated based on the radiation pneumonia outcome labels and clinical scoring indicators. Using the weighted feature vector as input and the radioactive pneumonia outcome label as supervision signal, a multilayer perceptron model is trained on the source domain node through federated learning to obtain a baseline model containing a shared basic representation layer and a shared main classifier head; The baseline model is migrated to the target domain node, and a treatment background conditional encoding layer and a treatment-specific residual head are added to obtain an initial adapted model. The treatment background conditional encoding layer takes the treatment background vector as input, and the input of the treatment-specific residual head is the concatenation result of the output of the shared basic representation layer and the output of the treatment background conditional encoding layer. Using the weighted feature vector and the treatment background vector as inputs, and the soft label as a supervision signal, the initial adaptation model is fine-tuned and trained on the target domain node to obtain the target domain adaptation model; based on the weighted feature vector of the patient to be tested, the radiation pneumonia risk probability is output through the target domain adaptation model.

2. The method according to claim 1, characterized in that: The calculation of segmentation confidence coefficients based on the consistency among multiple lung masks obtained from multiple inferences on the same standardized CT volume data includes: The original standardized CT volume data, the standardized CT volume data after geometric transformation, and the standardized CT volume data after pixel transformation are respectively input into the three-dimensional segmentation model to obtain three corresponding lung masks. Calculate the ratio of the number of consistent voxels among the three lung masks to the total number of voxels, and use this ratio as the segmentation confidence coefficient.

3. The method according to claim 1, characterized in that, The weighted processing of the radiomics features based on the segmentation confidence coefficient to obtain a weighted feature vector includes: The sensitivity value of each radiomics feature to changes in the lung mask is calculated to obtain the feature sensitivity vector; For each radiomics feature, the confidence weight of that feature is calculated based on the segmentation confidence coefficient and the sensitivity value of the corresponding dimension in the feature sensitivity vector. The standardized radiomics feature values ​​of each dimension are multiplied by the corresponding confidence weights to obtain the weighted feature vector.

4. The method according to claim 3, characterized in that, The calculation of the sensitivity value of each radiomics feature to changes in the lung mask includes: Perform dilation and erosion operations on the lung mask of the source domain training samples to obtain the dilated mask and the eroded mask respectively; Using the expanded mask and the etched mask as regions of interest, radiomics features are extracted again to obtain the expanded feature set and the etched feature set. Based on the expanded feature set, the eroded feature set, and the original feature set extracted from the original lung mask, the average relative change rate of each feature under the expanded and eroded states relative to the original state is calculated, and this average relative change rate is used as the sensitivity value.

5. The method according to claim 1, characterized in that, The generation of soft labels based on the radiation pneumonia outcome labels and clinical scoring indicators includes: The target domain samples are scored according to preset clinical scoring indicators, including the onset time window, image distribution location, whether infection is ruled out, and physician attribution consistency. Based on the preset range in which the total score falls, the radiation pneumonia outcome label is converted into a continuous value between 0 and 1, which is used as the soft label.

6. The method according to claim 1, characterized in that, Fine-tuning the initial adaptation model on the target domain nodes includes: Construct a total loss function that includes soft label cross-entropy loss and boundary emphasis loss, wherein the boundary emphasis loss is a weighted loss after assigning higher loss weights to samples in the soft labels that indicate attribution uncertainty; The initial adaptation model is trained with the goal of minimizing the total loss function.

7. The method according to claim 1, characterized in that, Fine-tuning the initial adaptation model on the target domain nodes includes: In the first stage, while keeping the parameters of the shared basic representation layer unchanged, the shared master classifier, the treatment background condition encoding layer, and the treatment-specific residual head are trained. In the second stage, after the training in the first stage is completed, the parameters of some network layers in the shared basic representation layer are opened for training.

8. The method according to claim 1, characterized in that, Also includes: If the patient to be tested is a patient who has not received immunotherapy, then the weighted feature vector is input into the baseline model, and the baseline risk value output by the shared master classifier in the baseline model is mapped by the activation function as the probability of radiation pneumonitis risk. If the patient to be tested is a patient receiving immunotherapy, then the weighted feature vector and the treatment background vector are input into the target domain adaptation model, and the sum of the baseline risk value output by the shared master classifier and the residual risk value output by the treatment-specific residual classifier, after being mapped by the activation function, is used as the probability of radiation pneumonitis risk.

9. A radiation pneumonitis risk prediction system based on federated transfer learning, characterized in that, include: The data acquisition module is used to acquire planned CT images, treatment information, and radiation pneumonitis outcome labels of patients who have not received immunotherapy in the source domain nodes and patients who have received immunotherapy in the target domain nodes. The image processing and credibility assessment module is used to input the standardized CT volume data obtained by standardizing the planned CT image into the pre-trained three-dimensional segmentation model to obtain the lung mask, and calculate the segmentation credibility coefficient based on the consistency between multiple lung masks obtained by multiple inferences on the same standardized CT volume data. The feature extraction and weighting module is used to extract radiomics features from the standardized CT volume data with the lung mask as the region of interest, and to weight the radiomics features based on the segmentation confidence coefficient to obtain a weighted feature vector. The tag generation module is used to construct a treatment background vector based on the treatment information and generate soft tags based on the radiation pneumonia outcome tags and clinical scoring indicators. The federated learning training module is used to train a multilayer perceptron model on the source domain node by federated learning, using the weighted feature vector as input and the radioactive pneumonia outcome label as supervision signal, to obtain a baseline model containing a shared basic representation layer and a shared main classifier head. The risk prediction module is used to migrate the baseline model to the target domain node and add a treatment background condition encoding layer and a treatment-specific residual head to obtain an initial adapted model. The treatment background condition encoding layer takes the treatment background vector as input, and the input of the treatment-specific residual head is the concatenation result of the output of the shared basic representation layer and the output of the treatment background condition encoding layer. Using the weighted feature vector and the treatment background vector as inputs, and the soft label as a supervision signal, the initial adaptation model is fine-tuned and trained on the target domain node to obtain the target domain adaptation model; based on the weighted feature vector of the patient to be tested, the radiation pneumonia risk probability is output through the target domain adaptation model.

Images