A helical tomotherapy post left breast cancer early subclinical cardiac injury prediction system and method

Abstract

The application discloses a kind of left breast cancer early subclinical heart injury prediction system and method after helical tomotherapy, data input and pre-processing module is responsible for obtaining and pre-processing the CT image of patient, radiotherapy dose map, 12 lead electrocardiogram time series signal and clinical variable;DEM module is responsible for converting CT image and radiotherapy dose distribution map into shallow layer joint dose embedding feature;Main network is responsible for converting shallow layer joint dose embedding feature into deep layer space comprehensive feature map by combining DM-RPB mechanism and CAG-AWA mechanism;MTCSA module is responsible for reconstructing the segmentation probability map and prediction mask of key heart substructure;Multimodal late cross-attention fusion head is responsible for converting deep layer space comprehensive feature map, electrocardiogram time series signal and clinical variable into fusion feature vector;Uncertainty prediction output head is responsible for outputting early subclinical heart injury occurrence risk probability and the uncertainty score of system prediction result according to fusion feature vector.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical fields of medical image processing technology, multimodal deep learning and radiotherapy toxicity prediction, and specifically relates to a system and method for predicting early subclinical cardiac damage in left breast cancer after helical tomotherapy. Background Technology

[0002] Breast cancer is one of the most common malignant tumors among women worldwide. Studies show that in recent years, the number of new breast cancer cases ranks second among all female cancers. Radical radiotherapy is one of the standard treatments for breast-conserving surgery or locally advanced breast cancer. Especially in adjuvant radiotherapy after left-sided mastectomy for breast cancer, precise dose distribution is crucial for reducing long-term cardiotoxicity.

[0003] Helical tomotherapy (TOMO / Radixact system), with its 360-degree helical irradiation, daily megavolt-level computed tomography image guidance, and high conformity, has been widely used in the industry for irradiating the chest wall and regional lymph nodes in left-sided breast cancer. However, the inherent "low-dose bath" effect of TOMO technology can lead to a significant increase in the low-dose volume (V5) of the heart and lungs, potentially increasing the risk of subclinical cardiac injury. The Chinese Society of Clinical Oncology's "Guidelines for the Prevention and Treatment of Cardiovascular Toxicity Related to Cancer Treatment" (2023 update) clearly states that radiotherapy-related cardiac injury is one of the most serious long-term complications after radiotherapy for left-sided breast cancer. Even with modern precision radiotherapy techniques that have significantly reduced the average cardiac dose, early subclinical cardiac injury can still occur, with an incidence rate of 20%–40%.

[0004] Traditional dosimetric prediction models suffer from insufficient sensitivity and specificity. Existing predictive metrics, such as mean heart dose (MHD) and mean dose in the left anterior descending coronary artery (LAD), only reflect overall cardiac radiation levels and fail to capture spatial dose heterogeneity in cardiac substructures (LAD, left ventricular basal segment, and pericardium), making individualized risk stratification difficult. The area under the receiver operating characteristic (AUC) of logistic regression models based on MHD and LAD dose is only about 0.69, indicating limited discriminative power.

[0005] Existing radiomics models rely on manual feature extraction, making it difficult to model the complex spatial patterns of the TOMO bath. Traditional radiomics methods require manually extracting first-order statistical features, shape features, and texture features (such as gray-level co-occurrence matrix, gray-level run-length matrix, and gray-level region size matrix) from planned computed tomography images and 3D dose distribution maps, and then combining them with machine learning classifiers (such as random forests) to build a predictive model. This type of method cannot effectively model the unique effects of the TOMO low-dose bath on cardiac substructures, and the manual feature extraction process is cumbersome and has poor reproducibility.

[0006] Existing deep learning models lack multimodal fusion of dose and electrophysiological information. Existing pure image deep learning models (such as 3D ResNet-18 and SwinUNETR) only take into account planned computed tomography scans and dose distribution maps as input, without integrating clinical variables and electrocardiogram information. At the same time, existing pure image deep learning models are not specifically designed for the TOMO low-dose bath, lack a highly interpretable spatial attention mechanism, and cannot output attention heatmaps to guide clinical intervention.

[0007] Current research lacks end-to-end predictive models for early subclinical cardiac injury specifically for TOMO. Existing multimodal deep learning models mostly focus on chemotherapy-related cardiotoxicity or general radiotherapy-induced cardiac injury. Pure end-to-end deep learning models specifically for TOMO, focusing on early subclinical cardiac injury 3 to 6 months after radiotherapy, and integrating electrocardiogram as a multimodal input, remain a research gap. Summary of the Invention

[0008] This invention aims to provide a high-discrimination, high-generalization-performance, interpretable, and high-reliability prediction system and method for early subclinical cardiac injury in left breast cancer after helical tomotherapy by constructing a dose-informed 3D Swin Transformer multi-task multimodal deep learning model (TomoCardioFormer). This overcomes the shortcomings of existing technologies, such as the inability to effectively model the complex three-dimensional dose heterogeneity spatial pattern unique to helical tomotherapy (TOMO), the lack of strong anatomical boundary constraints, and the inability to integrate one-dimensional electrophysiological signals with one-dimensional scalar clinical indicators.

[0009] To solve the above-mentioned technical problems and achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: A system for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy includes: The data input and preprocessing module is responsible for acquiring and preprocessing the patient's original three-dimensional planning CT images (X-ray). CT ), original three-dimensional helical tomotherapy dose map (X) Dose ), original baseline standard 12-lead electrocardiogram timing signal (X) ECG ) and clinical variables (Vclin ), outputting a 3D planning CT image tensor (X CT_norm ), Three-dimensional helical tomotherapy dose distribution tensor (X Dose_norm Baseline standard 12-lead ECG time tensor (X) ECG_norm ) and one-dimensional clinical variable vector (V clin_norm ); The dose embedding module (DEM) is responsible for planning the 3D CT image tensor (X-ray) using a physically prior mask weighted and a learnable projection matrix. CT_norm ) and three-dimensional helical tomotherapy dose distribution tensor (X Dose_norm Shallow spatial features are extracted and fused to output shallow joint dose embedding features (F). embed ); The dose-modulated 3D Swin Transformer backbone network is responsible for utilizing a four-stage 3D Swin Transformer that integrates a dose-modulated relative position bias (DM-RPB) mechanism and a cardiac substructure-guided adaptive window attention (CAG-AWA) mechanism to process shallow joint dose embedding features (F... embed Perform a four-stage deep feature extraction process and output a deep spatial comprehensive feature map; The Multi-Task Cardiac Substructure Attention Block (MTCSA), connected to the intermediate stage feature output of the 4-stage 3D Swin Transformer, is responsible for reconstructing the 3D segmentation probability map of key cardiac substructures (G) through a lightweight upsampling decoder. sub ) and key cardiac substructure three-dimensional prediction mask (M sub The three-dimensional segmentation probability map of the key cardiac substructures (G) sub During inference, the key cardiac substructure-guided adaptive window attention (CAG-AWA) mechanism is used as the basis for calculation. The three-dimensional prediction mask (M) of the key cardiac substructure serves as the basis for this mechanism. sub During training, it serves as an auxiliary segmentation task for joint training with strong supervision. The multimodal post-processing cross-attention fusion head is responsible for processing the one-dimensional clinical variable vector (V) separately. clin_norm ) and baseline standard 12-lead ECG time tensor (X ECG_norm Feature extraction is performed on the extracted one-dimensional clinical variable feature vector (F) using a multi-head cross-attention mechanism. clin ), one-dimensional electrophysiological feature vector (F ECG The spatial anatomy-dose total feature vector (F) obtained by three-dimensional global average pooling of the deep spatial comprehensive feature map. spatial Perform high-order post-fusion to output a fused feature vector (F). fusion ); The uncertainty prediction output head is responsible for fusing the feature vector (F) fusion The algorithm performs multiple forward propagations and employs a Monte Carlo Dropout mechanism to ensure the confidence level of the prediction results, ultimately outputting the probability of early subclinical cardiac injury (Y). prob ) and the uncertainty score of the system prediction results (Y uncert ).

[0010] Furthermore, the data input and preprocessing module includes: The multimodal heterogeneous data acquisition unit is responsible for acquiring the patient's original three-dimensional planning CT images (X-rays) in the radiotherapy planning system (TPS). CT ) and the corresponding original three-dimensional helical tomotherapy dose map (X Dose ); Read the patient's original baseline (within the cycle prior to radiotherapy) standard 12-lead electrocardiogram (ECG) timing signal (X-ray) ECG ); collected 24-dimensional clinical variables including patient age, BMI, history of hypertension, whether anthracycline / cyclophosphamide chemotherapy was used concurrently and the cumulative dose, and peak left ventricular outflow tract velocity (V). clin ); The 3D image axial rigid registration unit is responsible for using the SimpleITK toolbox to acquire the original 3D planning CT image (X). CT Using the baseline image (Fixed Image), the acquired original three-dimensional helical tomotherapy dose map (X-ray) was compared with the baseline image (Fixed Image). Dose Perform axial rigid registration of 3D images; The resampling unit is responsible for resampling the registered three-dimensional planning CT image and the three-dimensional spiral tomography dose map to a resolution of 256×256×128 voxels, with the physical size of the voxel space fixed at 1mm×1mm×2mm. The outlier window clipping unit is responsible for uniformly clipping the window width and level of the resampled 3D planning CT image to [-200, 300] HU (focusing on the mediastinum and soft tissue regions) and performing routine Z-score normalization; extracting the absolute dose value (in Gy) from the resampled 3D helical tomography dose map, dividing the absolute dose value by the patient's actual prescribed total dose scalar value, and linearly mapping it to the [0, 1] normalization interval to obtain the 3D planning CT image tensor (X). CT_norm ) and three-dimensional helical tomotherapy dose map tensor (X Dose_norm ); The ECG digital filtering unit is responsible for removing the original baseline (within the cycle before radiotherapy) standard 12-lead electrocardiogram (ECG) digital signal (X-ray) by passing it through a 0.5-100Hz bandpass filter and a 50Hz power frequency notch filter. ECGBaseline drift and electromyography artifacts were analyzed; the first 5000 time-domain sampling points were uniformly extracted, and the maximum and minimum values ​​were normalized to reconstruct a 12×5000 baseline standard 12-lead ECG time tensor (X). ECG_norm ); The clinical variable unique hot coding unit is responsible for encoding the patient's 24-dimensional clinical variables (V... clin The categorical variables in the data are coded using One-Hot encoding. The continuous scalar variables in the 24-dimensional clinical variables of the patient are standardized using mean and variance, and then combined into a one-dimensional clinical variable vector (V). clin_norm ).

[0011] Furthermore, the dose embedding module (DEM) includes: The CT feature extraction branch is responsible for planning the CT image tensor (X) in three dimensions using 3×3×3 convolution. CT_norm Feature extraction is performed to output the shallow layer features (F) of the 3D planned CT image. CT ); The initial dose feature extraction branch is responsible for performing 3×3×3 convolution on the three-dimensional helical tomography dose map tensor (X). Dose_norm Feature extraction is performed to output the shallow features of the three-dimensional helical tomography dose map (F). Conv_Dose ); The physical prior dose soft mask calculation unit is responsible for calculating the dose map tensor (X) based on the three-dimensional helical tomography radiotherapy tensor. Dose_norm Using the raw data, normalized physical dose threshold constant (T), and smoothing factor (τ), the adaptive physical guidance soft mask (M) is calculated. Dose ); The dose feature weighting unit is responsible for weighting the initially extracted dose distribution spatial pattern features (F... Conv_Dose The calculated soft mask is then subjected to element-wise Hadamard product to achieve physical prior weighting, outputting a dose feature representation (F) that has been filtered by physical prior and focuses on the high-dose region. Dose ); The cross-modal feature fusion unit is responsible for representing the dose features obtained after physical prior weighting (F... Dose The features are projected through a 1×1×1 three-dimensional learnable point convolution and then combined with the extracted shallow features (F) of the three-dimensional planning CT image. CT The summation of these two data points outputs a unified embedded representation that integrates anatomical information (from 3D planning CT images) and treatment planning information (from 3D helical tomography dose maps), namely, the superficial combined dose embedding feature (F...). embed ); where B represents the batch size.

[0012] Furthermore, the dose modulation 3D Swin Transformer backbone network includes: The 4-stage 3D Swin Transformer is responsible for feature evolution, specifically the evolution of features embedded from shallow combined dose (F). embed The backbone network features formed after assignment (F) backbone The deep feature extraction process is performed in four stages, gradually incorporating the input backbone network features (F... backbone The process transforms the input image into a more semantically informative feature representation. Specifically, in each of the four stages, a series of Transformer blocks and window attention mechanisms are used to progressively extract and transform features from the input image. From low-level edge and texture features to high-level semantic and structural features, the level of feature abstraction is gradually improved. Then, by changing the feature map size and adjusting the number of channels in different stages, the fusion of multi-scale features is achieved, and finally, a deep spatial comprehensive feature map is output. This allows the model to simultaneously capture global structural information and local detail information in cardiac images, improving the ability to recognize cardiac structures of different sizes and shapes. The dose-modulated relative position bias (DM-RPB) mechanism is responsible for optimizing the position bias in the self-attention calculation during the four-stage deep feature extraction process, thereby enhancing the model's attention to and discrimination ability of different dose regions in cardiac images. Specifically, it first optimizes the position bias based on the input three-dimensional helical tomography dose map tensor (X). Dose_norm Using the window coordinates, the mean dose, maximum dose, and standard deviation of dose within each window are calculated. Then, the ratio of the range to the standard deviation of the dose is calculated, scaled using the Tanh activation function, and a baseline of 1 and a learnable parameter are added. α The dimensional adaptive dose adjustment factor (φ) reflecting the spatial gradient of the low-dose bath within the window was calculated. Dose Finally, the standard 3D relative position is biased (Bias). standard Multiply by the calculated dimensional adaptive dose adjustment factor (φ) Dose Multiplicative weighting is performed to complete self-attention mapping and generate the relative position bias of dose modulation. DM This is used to calculate the self-attention matrix for dose-informed modulation; this allows the model to focus more on areas with large dose changes when calculating self-attention, thus improving its sensitivity to areas of cardiac lesions. The Cardiac Substructure Guided Adaptive Window Attention (CAG-AWA) mechanism is responsible for dynamically adjusting the size of the attention window during the four-stage deep feature extraction process. This allows the model to adaptively focus on substructure regions of different sizes and complexities in cardiac images, improving the ability to identify and segment fine cardiac structures. Specifically, in stages 3 and 4, the key cardiac substructure 3D segmentation probability map (G) predicted from the tail of stage 2 is first used. subThe average segmentation probability of cardiac substructures is calculated, and then the size of each dynamic window is calculated based on the calculated average segmentation probability of cardiac substructures and the discretization calculation logic. For regions with more complex or important cardiac substructures, the window size will be increased accordingly so that the model can capture the feature information of the region more comprehensively. For regions with relatively simple substructures, the window size will be kept smaller to improve computational efficiency. By combining the dose-modulated relative position bias (DM-RPB) mechanism with the cardiac substructure guided adaptive window attention (CAG-AWA) mechanism, the 4-stage 3D Swin Transformer model can better capture complex features and structural information in cardiac images.

[0013] Furthermore, the Multi-Task Cardiac Substructure Attention Block (MTCSA) includes a parallel auxiliary segmentation decoder derived at the deep feature map transition point of the second stage of the 4-stage 3D Swin Transformer. This auxiliary segmentation decoder consists of consecutive 2×2×2 3D Transposed Convolutions and residual convolution blocks, responsible for upsampling and restoring the deep feature map output from the second stage of the 4-stage 3D Swin Transformer to its original size H×W×D. The number of channels is set to K=4 (representing background, whole heart, left ventricular basal segment, and LAD vascular region), outputting 3D segmentation probability maps and 3D prediction masks for the whole heart, left ventricular basal segment, and LAD vascular region, and a 3D segmentation probability map (G) for key cardiac substructures. sub ) and key cardiac substructure three-dimensional prediction mask (M sub The auxiliary segmentation decoder participates in the global loss function penalty during the training period, forcing the underlying backbone of the 4-stage 3D Swin Transformer to learn high-risk anatomical boundaries.

[0014] Furthermore, the multimodal post-processing cross-attention fusion head includes: The one-dimensional clinical variable encoder employs a multilayer perceptron (MLP) consisting of three fully connected layers, InstanceNorm1D, and LeakyReLU, responsible for encoding the input one-dimensional clinical variable vector (V). clin_norm Feature extraction is performed to output a one-dimensional clinical variable feature vector (F). clin ); A one-dimensional electrocardiogram encoder includes a set of one-dimensional convolutional layers and a temporal network consisting of four standard Transformer Encoder layers; the one-dimensional convolutional layers are responsible for processing the input baseline standard 12-lead electrocardiogram temporal tensor (X). ECG_normThe temporal network performs time-domain downsampling; it is responsible for capturing long-range heart rhythm and waveform abnormalities in the time-domain downsampling results and outputting a one-dimensional electrophysiological feature vector (F). ECG ); The three-dimensional global average pooling unit is responsible for performing three-dimensional global average pooling on the deep spatial comprehensive feature map output by the dose modulation 3D Swin Transformer backbone network, and outputting a spatial anatomy-dose total feature vector (F). spatial ); The multi-head cross-attention post-fusion unit is responsible for utilizing the multi-head cross-attention mechanism to fuse the one-dimensional clinical variable feature vector (F... clin ), one-dimensional electrophysiological feature vector (F ECG ) and spatial anatomy-dose total eigenvector (F spatial Perform high-order post-fusion to output a fused feature vector (F). fusion ).

[0015] Furthermore, the uncertainty prediction output header includes: The classification fully connected layer, connected to the output of the multimodal late-stage cross-attention fusion head, is responsible for processing the fused feature vector (F). fusion Multiple independent random inactivation forward propagations are performed to infer a series of classification probability predictions. Then, by calculating this series of classification probability prediction values algebraic mean and discrete variance Ultimately, the probability of early subclinical cardiac injury (Y) was obtained. prob ) and the uncertainty score of the system prediction results (Y uncert ); The Monte Carlo random deactivation mechanism, deployed before the fully connected classification layer, is responsible for causing the fully connected classification layer to perform forward classification prediction with a random deactivation rate of p=0.3 during test inference, in order to ensure the confidence of the output results.

[0016] A method for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy based on the above system includes the following steps: Step 1: Acquisition and preprocessing of multimodal heterogeneous data: (1) Rigid registration of three-dimensional images: Acquiring the patient's original three-dimensional planning CT images (X-rays) in the radiotherapy planning system (TPS). CT ) and the corresponding original three-dimensional helical tomotherapy dose map (X Dose Using the SimpleITK toolbox, the original 3D-planned CT images (X-ray) were used. CTUsing this as the baseline image, the original three-dimensional helical tomography dose map (X-ray) was compared with the baseline image. Dose Perform axial rigid registration of 3D images; (2) Resampling and Outlier Window Clipping: The resolution of the registered 3D planning CT image and the 3D helical tomography dose map was resampled to 256×256×128 voxels, and the physical size of the voxel space was fixed at 1mm×1mm×2mm; the window width and window level of the resampled 3D planning CT image were uniformly clipped to [-200, 300] HU (focusing on the mediastinum and soft tissue regions), and routine Z-score normalization was performed; the absolute dose value (in Gy) of the resampled 3D helical tomography dose map was extracted, and the absolute dose value was divided by the patient's actual prescribed total dose scalar value, linearly mapped to the [0,1] normalization interval to obtain the 3D planning CT image tensor (X). CT_norm ) and three-dimensional helical tomotherapy dose map tensor (X Dose_norm ); (3) Digital ECG filtering: Read the baseline (within the cycle before radiotherapy) standard 12-lead electrocardiogram (ECG) digital signal (X-ray) ECG The baseline drift and electromyography artifacts were removed using a 0.5-100Hz bandpass filter and a 50Hz power frequency notch filter. The first 5000 time-domain sampling points were uniformly extracted, and the maximum and minimum values ​​were normalized to reshape them into a 12×5000 baseline standard 12-lead electrocardiogram time tensor (X). ECG_norm ); (4) Unique heat coding of clinical variables: 24-dimensional clinical variables (V) including patient age, BMI, history of hypertension, whether anthracycline / cyclophosphamide chemotherapy drugs were used in combination and the cumulative dose, and peak left ventricular outflow tract velocity were collected. clin The categorical variables are encoded using One-Hot encoding, and the continuous scalar variables are standardized using mean and variance, combining them into a one-dimensional feature vector (V) of clinical variables. clin_norm ); Step 2: Shallow combined dose embedding feature extraction: (1) CT feature extraction: The CT image tensor (X) is planned using 3×3×3 convolution. CT_norm Feature extraction is performed to obtain shallow features (F) of the three-dimensional planning CT image. CT ); (2) Preliminary extraction of dose characteristics: The three-dimensional helical tomography dose map tensor (X) was analyzed by 3×3×3 convolution. Dose_norm Feature extraction was performed to obtain the shallow features of the three-dimensional helical tomotherapy dose map (F). Conv_Dose ); (3) Calculate the physical prior dose soft mask: Based on the three-dimensional helical tomography dose map tensor (X Dose_norm Using the raw data, normalized physical dose threshold constant (T), and smoothing factor (τ), the adaptive physical guidance soft mask (M) is calculated. Dose This forces the model to explicitly perceive radiation-risk areas during forward computation; (4) Dose-characteristic weighting: The initially extracted dose distribution spatial pattern features (F) Conv_Dose ) and the calculated adaptive physical guidance soft mask (M Dose Element-wise Hadamard product is performed to complete physical prior weighting, resulting in a dose feature representation (F) that has been filtered by physical priors and focuses on the high-dose region. Dose ); (5) Cross-modal feature fusion: The dose characteristic representation obtained after dose characteristic weighting (F) Dose The features are projected through a 1×1×1 three-dimensional learnable point convolution and then combined with the extracted shallow features (F) of the three-dimensional planning CT image. CT The two components are added together and fused to obtain a unified embedding representation that integrates anatomical structure information (from 3D planning CT images) and treatment planning information (from 3D helical tomography dose maps), namely, the superficial combined dose embedding feature (F...). embed ); Step 3: Evolution of features of the dose-modulated 3D Swin Transformer backbone network: (1) Initialization: In the shallow combined dose embedding feature (F embed After being fed into the 4-stage dose modulation 3D Swin Transformer backbone network, the shallow joint dose embedding feature (F) is then incorporated. embed ) Assign values ​​to the backbone network features (F backbone This serves as the starting point for the input in the subsequent four stages of processing. (2) Four-stage cycle of the backbone network: The four-stage feature extraction of the dose-modulated 3D Swin Transformer backbone network is enabled, iterating from stage 1 to stage 4. In each stage, the backbone network features (F...) of the current stage are extracted. backbone The system is divided into non-overlapping 3D local windows, each with a size of W×W×W. Within each window, the relative position offset of dose modulation, dynamic window size adjustment, and self-attention calculation are performed independently and sequentially. Finally, the deep spatial integrated feature map is output by the fourth stage located at the top layer. (3) Calculate the dose modulation relative position bias (DM-RPB): A dose modulation relative position bias (DM-RPB) mechanism is introduced, within each self-attention calculation window (Ω), from the input three-dimensional helical tomography dose map tensor X. Dose_norm Slice the corresponding local dose block (D) Ω ), and use this to calculate the mean dose (D) for the current window. mean ), maximum dose (D) max ) and dose standard deviation (D std These three dose statistics are calculated by taking the ratio of the range to the standard deviation of the dose, scaling it using the Tanh activation function, and adding a baseline of 1 and a learnable parameter. α The dimensional adaptive dose adjustment factor (φ) reflecting the spatial gradient of the low-dose bath within the window was calculated. Dose ); Bias the standard 3D relative position. standard Multiply by the calculated dimensional adaptive dose adjustment factor (φ) Dose Multiplicative weighting is performed to complete self-attention mapping and generate the relative position bias of dose modulation. DM ), and use this to calculate the dose to inform the modulated self-attention matrix; (4) Predicting the three-dimensional segmentation probability map of key cardiac substructures (G sub ): The deep feature map output from stage 2 is upsampled and restored to its original size H×W×D through 3D transposed convolution using a lightweight upsampling decoder. The number of channels is set to K=4 (representing the background, whole heart, left ventricular basal segment, and LAD vascular region). The output is a 3D segmentation probability map of the whole heart, left ventricular basal segment, and LAD vascular region, i.e., a 3D segmentation probability map of key cardiac substructures (G). sub ); (5) Dynamic window size adjustment: A cardiac substructure-guided adaptive window attention (CAG-AWA) mechanism is introduced, based on the cardiac substructure 3D segmentation probability map (G) obtained from the tail prediction in stage 2. sub The model calculates the mean probability of cardiac substructure segmentation. In the forward feature transfer of the third and fourth stages, the model calculates the size of each dynamic window based on the calculated mean probability of cardiac substructure segmentation and the discretization calculation logic. This dynamically adjusts the size of the Swin local sliding attention window between 5×5×5, 7×7×7, 9×9×9, and 11×11×11. For regions with more complex or important cardiac substructures, the window size is increased accordingly so that the model can capture the feature information of the region more comprehensively. For regions with relatively simple substructures, the window size is kept smaller to improve computational efficiency. (6) Self-attention execution: The window data is re-divided or processed using dynamically calculated window sizes, and standard 3D Swin self-attention operations are performed in conjunction with dose-modulated relative positional biases; this means that the attention range changes dynamically with the importance of anatomical structures, and the attention weights are modulated by local dose distribution characteristics. Step 4: One-dimensional multimodal asynchronous feature encoding and multi-head cross-attention fusion: (1) Perform three-dimensional global average pooling on the deep spatial comprehensive feature map output by the dose modulation 3D Swin Transformer backbone network to reduce the dimensionality to obtain a one-dimensional spatial anatomy-dose total feature vector (F). spatial ); (2) A multilayer perceptron (MLP) consisting of three fully connected layers, InstanceNorm1D, and LeakyReLU is used to process the input one-dimensional clinical variable vector (V). clin_norm The projection is a one-dimensional feature vector of clinical variables (F). clin ); (3) First, a set of one-dimensional convolutional layers is used to process the input baseline standard 12-lead electrocardiogram time tensor (X). ECG_norm Temporal downsampling is performed, followed by the use of a temporal network consisting of four standard Transformer Encoders to capture long-term heart rhythm and waveform abnormalities, thereby encoding a one-dimensional electrophysiological feature vector (F). ECG ); (4) Define the learnable projection matrix W Q W K W V The spatial anatomy-dose total feature vector (F) spatial ) as a query, taking the one-dimensional clinical variable feature vector (F) clin ) and one-dimensional electrophysiological eigenvectors (F ECG The feature vector (F) obtained after concatenation context The data is mapped to keys and values ​​through different linear layers, input to a multi-head cross-attention layer, performs higher-order modality fusion, and outputs a fused feature vector (F). fusion ); Step 5: Monte Carlo random inactivation inference and high-confidence prediction output: Fuse feature vectors (F) fusion The input is a fully connected classification layer with a Monte Carlo random inactivation mechanism, where the random inactivation rate of the Monte Carlo random inactivation mechanism is p=0.3. This fully connected classification layer automatically triggers an iterative controller to process the input fused feature vector (F...). fusionThe forward classification prediction is performed continuously N times (N=10) independently and repeatedly. The Monte Carlo random deactivation mechanism remains on throughout the test inference phase to ensure that the output of the fully connected classification layer has confidence, thereby obtaining the N classification probability prediction values. Finally, calculate the predicted probability values ​​for these N classification iterations. algebraic mean As the probability of developing early subclinical cardiac injury (Y) prob ), and calculate the predicted classification probabilities for these N iterations. discrete variance The uncertainty score (Y) is the result of the system prediction. uncert ); ; ; Where σ(·) is the standard binary classification sigmoid activation function.

[0017] A computer device includes a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface communicate with each other via the communication bus. The memory stores at least one executable instruction, which causes the processor to perform an operation corresponding to the aforementioned method for predicting early subclinical cardiac damage in left breast cancer after helical tomotherapy.

[0018] A computer-readable storage medium storing at least one executable instruction that causes a processor to perform operations corresponding to the above-described method for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A world-first dose embedding plus multi-task Transformer framework for TOMO low-dose baths. This invention is the first to integrate the dose embedding module (DEM), dose modulation relative position bias (DM-RPB), and cardiac substructure guided adaptive window attention (CAG-AWA) into a 3D Swing Transformer backbone, specifically for modeling the complex spatial pattern of the TOMO spiral dose distribution, thus solving the technical bottleneck of traditional methods that cannot effectively capture the impact of TOMO low-dose baths on cardiac substructures.

[0020] 2. This invention is the first to integrate 12-lead electrocardiogram (ECG) as a fourth modality to achieve end-to-end multimodal fusion. Existing models do not incorporate ECG information into the prediction of early subclinical cardiac injury. This invention encodes ECG signals using a one-dimensional Transformer and then fuses them with three-dimensional image features using delayed fusion and cross-attention mechanisms. Baseline ECG abnormalities (such as T-wave inversion, ST-segment changes, and arrhythmias) are used as independent predictors, significantly improving prediction accuracy. Ablation experiments show that removing the ECG modality reduces the area under the receiver operating characteristic (AUC) curve from 0.852 to 0.809, confirming the significant contribution of the ECG modality to model performance.

[0021] 3. Pure end-to-end deep learning design, eliminating the need for manual feature extraction. This invention eliminates the need for traditional radiomics manual feature extraction or machine learning classifiers, fully capturing the spatial heterogeneity of cardiac substructures and electrocardiogram electrophysiological information. The area under the receiver operating characteristic (AUC) curve reached 0.852 (95% confidence interval: 0.812 to 0.934) in the internal test set and 0.784 (95% confidence interval: 0.730 to 0.838) in the external validation set, significantly outperforming traditional dosimetric models (AUC approximately 0.69), radiomics plus random forest models (AUC approximately 0.74), 3D ResNet-18 models (AUC approximately 0.77), and the SwinUNETR model (AUC approximately 0.80). The DeLong test showed statistically significant differences (P < 0.05).

[0022] 4. Strong interpretability of output, directly supporting individualized risk stratification in clinical practice. This invention visualizes the model's decision-making process using Grad-CAM++ and Transformer attention maps. Attention is primarily focused on high-dose regions of key cardiac substructures (especially the left anterior descending coronary artery and the basal segment of the left ventricle). Blinded evaluation by three radiation oncologists and one cardiologist yielded an average acceptability score of 4.2 ± 0.4 (out of 5) for the attention heatmap, demonstrating good clinical interpretability. Furthermore, Monte Carlo dropout uncertainty analysis showed an average confidence interval width of 0.082 for positive predictions, indicating higher confidence in high-risk cases. This provides a precise tool for early interventions recommended by the Chinese Society of Clinical Oncology guidelines (such as angiotensin-converting enzyme inhibitors / β-blockers and cardiac rehabilitation).

[0023] 5. Promoting a New Paradigm for Cardiac Protection in Precision Radiotherapy. The expected results of this invention will provide a new paradigm for cardiac protection in precision radiotherapy, promoting a shift from experience-driven to data- and biology-guided radiotherapy decisions, and have significant theoretical innovation and clinical translational value.

[0024] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the architecture of the left subclinical cardiac injury prediction system for left breast cancer after helical tomotherapy according to the present invention.

[0026] Figure 2 This is a schematic diagram of the internal architecture of the dose embedding module (DEM) in the prediction system for early subclinical cardiac injury of left breast cancer after helical tomotherapy according to the present invention.

[0027] Figure 3 This is a schematic diagram of the internal architecture of the dose-modulated 3D Swin Transformer backbone network (Tomo Cardio Swin Bcakbone) in the prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy according to the present invention.

[0028] Figure 4 This is a schematic diagram of the internal architecture of the Multi-Task Cardiac Substructure Attention Block (MTCSA) in the prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy according to the present invention. Detailed Implementation

[0029] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings to provide a clearer understanding of the invention's purpose, features, and advantages. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the invention's technical solution.

[0030] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known apparatuses, structures, and techniques associated with this application may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0031] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0032] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0033] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to mean “and / or” unless otherwise expressly stated herein.

[0034] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0035] See Figure 1 As shown, this invention provides a prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy, which mainly includes a data input and preprocessing module, a dose embedding module (DEM), a dose modulation 3D SwinTransformer backbone network (Tomo Cardio Swin Bcakbone), a multi-task cardiac substructure attention block (MTCSA), a multimodal late-stage cross-attention fusion head, and an uncertainty prediction output head.

[0036] 1. Data Input and Preprocessing Module: The data input and preprocessing module is responsible for acquiring and preprocessing the patient's original three-dimensional planning CT images. CT Original three-dimensional helical tomotherapy dose map X Dose Original baseline standard 12-lead electrocardiogram timing signal X ECG and clinical variable V clin Output the 3D planning CT image tensor X CT_norm 3D helical tomotherapy dose distribution tensor X Dose_norm Baseline standard 12-lead ECG temporal tensor X ECG_norm and the one-dimensional clinical variable vector V clin_norm .

[0037] As one embodiment, the data input and preprocessing module includes a multimodal heterogeneous data acquisition unit, a three-dimensional image axial rigid registration unit, a resampling unit, an outlier window clipping unit, an electrocardiogram digital filtering unit, and a clinical variable one-hot encoding unit.

[0038] The multimodal heterogeneous data acquisition unit is responsible for acquiring the patient's original three-dimensional planning CT images in the radiotherapy planning system (TPS). CT (DICOM format) and the corresponding original 3D helical tomotherapy dose map X Dose (RD Matrix); Read the patient's original baseline (within the cycle prior to radiotherapy) standard 12-lead electrocardiogram (ECG) timing signal X. ECG The clinical variable V, comprising 24 dimensions, was collected, including patient age, BMI, history of hypertension, whether anthracycline / cyclophosphamide chemotherapy was used concurrently and the cumulative dose, and peak left ventricular outflow tract velocity. clin .

[0039] The three-dimensional image axial rigid registration unit is responsible for using the SimpleITK toolbox to acquire the original three-dimensional planning CT image X. CT As a reference image (Fixed Image), the acquired original three-dimensional helical tomotherapy dose map X is compared. Dose Perform axial rigid registration of 3D images.

[0040] The resampling unit is responsible for resampling the registered three-dimensional planning CT image and the three-dimensional spiral tomography dose map to a resolution of 256×256×128 voxels, with the voxel spatial physical size fixed at 1mm×1mm×2mm.

[0041] The outlier window clipping unit is responsible for uniformly clipping the window width and level of the resampled 3D planning CT image to [-200, 300] HU (focusing on the mediastinum and soft tissue regions) and performing conventional Z-score normalization; extracting the absolute dose value (in Gy) from the resampled 3D helical tomography dose map, dividing the absolute dose value by the patient's actual prescribed total dose scalar value, and linearly mapping it to the [0, 1] normalization interval to obtain the 3D planning CT image tensor X. CT_norm and three-dimensional helical tomotherapy dose map tensor X Dose_norm .

[0042] The ECG digital filtering unit is responsible for removing the original baseline (within the previous radiotherapy cycle) standard 12-lead electrocardiogram (ECG) digital signal X from the standard ECG using a 0.5-100Hz bandpass filter and a 50Hz power frequency notch filter. ECGBaseline drift and electromyography artifacts were identified; the first 5000 time-domain sampling points were uniformly extracted, and maximum and minimum values ​​were normalized to reconstruct a 12×5000 baseline standard 12-lead ECG time tensor X. ECG_norm ; The clinical variable one-hot encoding unit is responsible for encoding the patient's 24-dimensional clinical variable V. clin The categorical variables in the data are coded using One-Hot encoding. The continuous scalar variables in the 24-dimensional clinical variables of the patients are standardized using mean and variance, and then combined into a one-dimensional clinical variable vector V. clin_norm .

[0043] 2. Dosage Embedding Module (DEM): The dose embedding module (DEM) is responsible for planning the 3D CT image tensor X using a physical prior mask weighting and a learnable projection matrix. CT_norm and the dose distribution tensor X of three-dimensional helical tomotherapy Dose_norm Shallow spatial features are extracted and fused to output shallow joint dose embedding features F. embed .

[0044] As one example, see Figure 2 As shown, the dose embedding module (DEM) includes a CT feature extraction branch, a preliminary dose feature extraction branch, a physical prior dose soft mask calculation unit, a dose feature weighting unit, and a cross-modal feature fusion unit.

[0045] The CT feature extraction branch is responsible for performing 3D planning of the CT image tensor X using 3×3×3 convolution. CT_norm Feature extraction is performed, and the shallow layer features F of the 3D planned CT image are output. CT .

[0046] The dose feature preliminary extraction branch is responsible for processing the three-dimensional helical tomography dose map tensor X using a 3×3×3 convolution. Dose_norm Feature extraction is performed, and the shallow feature F of the three-dimensional helical tomography dose map is output. Conv_Dose .

[0047] The physical prior dose soft mask calculation unit is responsible for calculating the dose map tensor X based on the three-dimensional helical tomotherapy dose map tensor. Dose_norm Using the raw data, the normalized physical dose threshold constant T, and the smoothing factor τ, the adaptive physical guidance soft mask M is calculated. Dose .

[0048] The dose feature weighting unit is responsible for weighting the initially extracted dose distribution spatial pattern features F Conv_Dose Element-wise Hadamard product is performed with the calculated soft mask to achieve physical prior weighting, outputting the dose feature representation F, which is physically prior filtered and focused on the high-dose region.Dose .

[0049] The cross-modal feature fusion unit is responsible for representing the dose feature F obtained after physical prior weighting. Dose After feature projection through a 1×1×1 3D learnable point convolution, it is then combined with the extracted shallow features F of the 3D planning CT image. CT The summation and fusion result in a unified embedding representation of the output tensor with a shape of B×64×128×128×64, which is the shallow joint dose embedding feature F. embed Where B represents the batch size.

[0050] One of the innovations of this invention is the use of a dose embedding module (DEM) combined with a physically prior-weighted bi-branch three-dimensional convolutional structure to achieve explicit modeling of the low-dose bath effect unique to helical tomotherapy (TOMO). By physically prior-weighting the high-dose region (greater than 10 Gy), the model can adaptively capture the unique spatial dose heterogeneity of the TOMO helical dose distribution on cardiac substructures, thus solving the problem that traditional methods cannot effectively model the impact of TOMO low-dose bath on cardiac injury.

[0051] 3. Dosage modulation 3D Swin Transformer backbone network: The dose-modulated 3D Swin Transformer backbone network is responsible for utilizing a four-stage 3D Swin Transformer that integrates a dose-modulated relative position bias (DM-RPB) mechanism and a cardiac substructure-guided adaptive window attention (CAG-AWA) mechanism to process shallow joint dose embedding features (F... embed The deep feature extraction is performed in four stages, and the deep spatial comprehensive feature map is output.

[0052] By combining the dose-modulated relative position bias (DM-RPB) mechanism with the cardiac substructure guided adaptive window attention (CAG-AWA) mechanism, the 4-stage 3D Swin Transformer model can better capture complex features and structural information in cardiac images.

[0053] As one example, see Figure 3 As shown, the dose modulation 3D Swin Transformer backbone network includes a 4-stage 3D Swin Transformer, a dose modulation relative position bias (DM-RPB) mechanism, and a cardiac substructure guided adaptive window attention (CAG-AWA) mechanism.

[0054] The four-stage 3D Swin Transformer is responsible for feature evolution, that is, for the shallow joint dose embedding feature F embed The backbone network feature F formed after assignment backbone A four-stage deep feature extraction process is performed to progressively integrate the input backbone network features F. backbone The model transforms the input image into a more semantically informative feature representation. Specifically, in four stages, each stage uses a series of Transformer blocks and window attention mechanisms to progressively extract and transform features from the input image. This process gradually increases the level of feature abstraction, moving from low-level edge and texture features to high-level semantic and structural features. Furthermore, by varying the feature map size and adjusting the number of channels at different stages, multi-scale feature fusion is achieved, ultimately outputting a deep spatial comprehensive feature map. This allows the model to simultaneously capture both global structural information and local detail information in cardiac images, improving its ability to recognize cardiac structures of different sizes and shapes.

[0055] The dose-modulated relative position bias (DM-RPB) mechanism is responsible for optimizing the position bias in the self-attention calculation during the four-stage deep feature extraction process, thereby enhancing the model's attention to and ability to distinguish features of different dose regions in cardiac images. Specifically, it first optimizes the position bias in the input three-dimensional helical tomography dose map tensor X. Dose_norm Given window coordinates, calculate three dose statistics for each window: mean dose, maximum dose, and standard deviation dose. Then, calculate the ratio of the range to the standard deviation of the dose, scale using the Tanh activation function, and add a baseline of 1 and a learnable parameter. α The dimensional adaptive dose adjustment factor φ, which reflects the spatial gradient of the low-dose bath within the window, was calculated. Dose Finally, the standard 3D relative position bias is applied. standard Multiply by the calculated dimensional adaptive dose adjustment factor φ Dose Multiplicative weighting is performed to complete self-attention mapping and generate the dose modulation relative position bias. DM This is used to calculate the autonomous force matrix for dose-informed modulation. This allows the model to focus more on regions with large dose changes when calculating self-attention, thus improving its sensitivity to areas of cardiac lesions.

[0056] The cardiac substructure-guided adaptive window attention (CAG-AWA) mechanism dynamically adjusts the size of the attention window during the four-stage deep feature extraction process. This allows the model to adaptively focus on substructure regions of different sizes and complexities in cardiac images, improving the ability to identify and segment fine cardiac structures. Specifically, in stages 3 and 4, the key cardiac substructure 3D segmentation probability map G, predicted from the tail of stage 2, is first used. subThe algorithm calculates the mean segmentation probability of cardiac substructures, and then, based on the calculated mean segmentation probability and discretization logic, calculates the size of each dynamic window. For regions with more complex or important cardiac substructures, the window size is increased accordingly so that the model can capture the feature information of that region more comprehensively; while for regions with relatively simple substructures, the window size remains smaller to improve computational efficiency.

[0057] One of the innovations of this invention is the fusion of dose information and cardiac anatomical priors within the 3D SwinTransformer backbone network through two core attention mechanisms: Dose-Modulated Relative Position Bias (DM-RPB) and Cardiac Substructure-Guided Adaptive Window Attention (CAG-AWA). DM-RPB directly replaces the standard relative position bias and is added after the transpose of the query and key and before the Softmax function. CAG-AWA dynamically adjusts the window size and offset during the window partitioning stage. Both are calculated in real-time from the original dose map and the planned computed tomography image, ensuring computational efficiency and physical / anatomical adaptability, and significantly improving the model's ability to learn the joint dose-anatomical representation of cardiac substructures.

[0058] 4. Multitasking Cardiac Substructural Attention Block (MTCSA): The Multi-Task Cardiac Substructure Attention Block (MTCSA) is connected to the intermediate stage feature output of the 4-stage 3D Swin Transformer, and is responsible for reconstructing the 3D segmentation probability map G of key cardiac substructures through a lightweight upsampling decoder. sub and key cardiac substructure 3D prediction mask M sub The three-dimensional segmentation probability map G of the key cardiac substructure sub The key cardiac substructure 3D prediction mask M serves as the computational basis for the cardiac substructure-guided adaptive window attention (CAG-AWA) mechanism during inference. sub During training, it is used as an auxiliary segmentation task to implement joint training with strong supervision.

[0059] As one example, see Figure 4As shown, the Multi-Task Cardiac Substructure Attention Block (MTCSA) includes a parallel auxiliary segmentation decoder derived at the deep feature map transition point of the second stage of the 4-stage 3D Swin Transformer. The auxiliary segmentation decoder consists of consecutive 2×2×2 3D Transposed Convolutions and residual convolution blocks. It is responsible for upsampling the deep feature map output from the second stage of the 4-stage 3D Swin Transformer back to its original size H×W×D, with the number of channels set to K=4 (representing background, whole heart, left ventricular basal segment, and LAD vascular region). It outputs 3D segmentation probability maps and 3D prediction masks for the whole heart, left ventricular basal segment, and LAD vascular region, i.e., the 3D segmentation probability map G of key cardiac substructures. sub and key cardiac substructure 3D prediction mask M sub The auxiliary segmentation decoder participates in the global loss function penalty during training, forcing the underlying backbone of the 4-stage 3D Swin Transformer to learn high-risk anatomical boundaries.

[0060] 5. Multimodal post-processing cross-attention fusion head: The multimodal post-processing cross-attention fusion head is responsible for processing the one-dimensional clinical variable vector V. clin_norm Baseline standard 12-lead ECG time tensor X ECG_norm Feature extraction is performed, and then the extracted one-dimensional clinical variable feature vector F is processed through a multi-head cross-attention mechanism. clin One-dimensional electrophysiological feature vector F ECG The spatial anatomy-dose total feature vector F obtained by three-dimensional global average pooling of the deep spatial comprehensive feature map. spatial Perform high-order post-fusion to output the fused feature vector F fusion .

[0061] As one embodiment, the multimodal late-stage cross-attention fusion head includes a one-dimensional clinical variable encoder, a one-dimensional electrocardiogram encoder, a three-dimensional global average pooling unit, and a multi-head cross-attention late-stage fusion unit.

[0062] The one-dimensional clinical variable encoder employs a multilayer perceptron (MLP) consisting of three fully connected layers, InstanceNorm1D, and LeakyReLU, and is responsible for encoding the input one-dimensional clinical variable vector V. clin_norm Feature extraction is performed, and a one-dimensional clinical variable feature vector F is output. clin ; The one-dimensional electrocardiogram encoder includes a set of one-dimensional convolutional layers and a temporal network consisting of four standard Transformer Encoder layers. The one-dimensional convolutional layers are responsible for processing the input baseline standard 12-lead electrocardiogram temporal tensor X. ECG_norm Temporal downsampling is performed. The temporal network is responsible for capturing long-range cardiac rhythm and waveform abnormalities in the temporal downsampling results and outputting a one-dimensional electrophysiological feature vector F. ECG .

[0063] The three-dimensional global average pooling unit is responsible for performing three-dimensional global average pooling on the deep spatial comprehensive feature map output by the dose modulation 3D Swin Transformer backbone network, and outputting the spatial anatomy-dose total feature vector F. spatial .

[0064] The multi-head cross-attention post-fusion unit is responsible for using the multi-head cross-attention mechanism to transform the one-dimensional clinical variable feature vector F clin One-dimensional electrophysiological feature vector F ECG With spatial anatomy-dose total eigenvector F spatial Perform high-order post-fusion to output the fused feature vector F fusion .

[0065] One of the innovations of this invention is that it is the first to use 12-lead electrocardiogram (ECG) signals as the fourth modality. After one-dimensional Transformer encoding, it employs late fusion and cross-attention mechanisms to perform end-to-end multimodal fusion with three-dimensional image features. Combined with Multi-Task Cardiac Substructure Attention (MTCSA), it achieves joint optimization of primary task injury risk prediction and auxiliary cardiac substructure segmentation. This enables accurate and individualized prediction of early subclinical cardiac injury 3 to 6 months after TOMO radiotherapy for left breast cancer. At the same time, it outputs attention heatmaps and uncertainty estimates, filling the international gap in prediction of early subclinical cardiac injury specifically for TOMO.

[0066] 6. Uncertainty prediction output head: The uncertainty prediction output head is responsible for processing the fused feature vector (F) output by the multimodal post-cross-attention fusion head. fusion The predictions undergo multiple forward propagations and employ a Monte Carlo Dropout mechanism to ensure confidence levels, ultimately outputting the probability Y of early subclinical cardiac injury. prob Uncertainty score Y of the system prediction resultsuncert .

[0067] As one embodiment, the uncertainty prediction output header includes a fully connected classification layer with an integrated Monte Carlo random deactivation mechanism.

[0068] The fully connected classification layer is connected to the output of the multimodal late-stage cross-attention fusion head, and is responsible for processing the fused feature vector F. fusion Multiple independent random inactivation forward propagations are performed to infer a series of classification probability predictions. Then, by calculating this series of classification probability prediction values algebraic mean and discrete variance Ultimately, the probability Y of early subclinical cardiac injury was obtained. prob Uncertainty score Y of the system prediction results uncert .

[0069] The Monte Carlo random deactivation mechanism is deployed before the classification fully connected layer. During test inference, it is responsible for causing the classification fully connected layer to perform forward classification prediction with a random deactivation rate of p=0.3 to ensure the confidence of the output results.

[0070] The training process of the early subclinical cardiac injury prediction system for left breast cancer after helical tomotherapy of the present invention is completed on a server equipped with a single NVIDIA A100 GPU.

[0071] The optimizer used is AdamW (Adaptive Moment Estimation with Decoupled Weight Decay), with a base learning rate lr0 = 1 × 10⁻⁶. -4 The weight decay coefficient is set to λ = 1 × 10 -5 .

[0072] The learning rate scheduling uses the Cosine Annealing LR algorithm. In the training cycle of a total of 200 epochs, the first 5 epochs implement a linear warm-up mechanism.

[0073] System-level construction of the loss function: ; in, The binary cross-entropy loss with Logits is used to supervise the damage main classification task. The three-dimensional multi-class generalized Dice loss function is used to assist in the joint strong supervision constraints of the segmentation task branches on three key heart structures: ; Where V is the total number of voxels, and η = 10 -5 The multi-task balancing hyperparameter is set to a fixed γ=0.3.

[0074] Based on the above system, this invention also provides a method for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy, see [link to relevant documentation]. Figure 1-4 As shown, the specific steps are as follows: Step 1: Acquisition and preprocessing of multimodal heterogeneous data: (1) Rigid registration of three-dimensional images: acquiring the patient's original three-dimensional planning CT image X in the radiotherapy planning system (TPS). CT (DICOM format) and the corresponding original 3D helical tomotherapy dose map X Dose (RD Matrix); Using the SimpleITK toolbox, plan the original 3D CT image X. CT As a baseline image (Fixed Image), the original three-dimensional helical tomotherapy dose map X... Dose Perform axial rigid registration of 3D images.

[0075] (2) Resampling and Outlier Window Clipping: The resolution of the registered 3D planning CT image and the 3D helical tomography dose map was resampled to 256×256×128 voxels, and the physical size of the voxel space was fixed at 1mm×1mm×2mm; the window width and window level of the resampled 3D planning CT image were uniformly clipped to [-200, 300] HU (focusing on the mediastinum and soft tissue regions), and routine Z-score normalization was performed; the absolute dose value (in Gy) of the resampled 3D helical tomography dose map was extracted, and the absolute dose value was divided by the patient's actual prescribed total dose scalar value and linearly mapped to the [0,1] normalization interval to obtain the 3D planning CT image tensor X. CT_norm and three-dimensional helical tomotherapy dose map tensor X Dose_norm .

[0076] (3) Digital ECG filtering: Read the baseline (within the cycle before radiotherapy) standard 12-lead electrocardiogram (ECG) digital signal X. ECG Baseline drift and electromyography artifacts were removed using a 0.5-100Hz bandpass filter and a 50Hz power frequency notch filter. The first 5000 time-domain sampling points were uniformly extracted, and the maximum and minimum values ​​were normalized to reconstruct a 12×5000 baseline standard 12-lead electrocardiogram time tensor X. ECG_norm .

[0077] (4) Unique heat coding of clinical variables: 24-dimensional clinical variables V were collected, including patient age, BMI, history of hypertension, whether anthracycline / cyclophosphamide chemotherapy drugs were used in combination and the cumulative dose, and peak left ventricular outflow tract velocity. clin The categorical variables were encoded using One-Hot encoding, and the continuous scalar variables were standardized using mean and variance, combining them into a one-dimensional feature vector V of clinical variables. clin_norm ∈R 214 .

[0078] Step 2: Shallow combined dose embedding feature extraction: (1) CT feature extraction: The CT image tensor X is planned using 3×3×3 convolution. CT_norm Feature extraction was performed to obtain the shallow layer features F of the 3D planning CT image. CT .

[0079] For example, let the input 3D planning CT image tensor X CT_norm ∈R 1×H×W×D Where 1, C, H, W, and D represent batch size, number of channels, height, width, and slice depth, respectively. After CT feature extraction, the shallow layer features F of the three-dimensional planned CT image are obtained. CT ∈R Cembed×H×W×D Among them, Conv3D 3×3×3 The kernel size is 3×3×3, the stride is S=1, and the padding is P=1.

[0080] (2) Preliminary extraction of dose characteristics: The tensor X of the three-dimensional helical tomography dose map was analyzed by 3×3×3 convolution. Dose_norm Feature extraction was performed to obtain the shallow features F of the three-dimensional helical tomotherapy dose map. Conv_Dose .

[0081] For example, suppose the input is a three-dimensional helical tomotherapy dose map tensor X. Dose_norm ∈R 1×H×W×D Where 1, C, H, W, and D represent batch size, number of channels, height, width, and slice depth, respectively. After preliminary extraction of dose features, the shallow features F of the three-dimensional helical tomotherapy dose map are obtained. Conv_Dose ∈R Cembed×H×W×D Among them, Conv3D 3×3×3 The kernel size is 3×3×3, the stride is S=1, and the padding is P=1.

[0082] (3) Calculate the physical prior dose soft mask: Based on the three-dimensional helical tomography dose map tensor X Dose_norm Using the raw data, the normalized physical dose threshold constant T, and the smoothing factor τ, the adaptive physical guidance soft mask M is calculated.Dose This forces the model to explicitly perceive radiation-risk areas during forward computation.

[0083] To force the model to explicitly perceive the radiotherapy risk area during forward computation, a physical dose threshold constant T (e.g., a normalized scalar value of the clinical low-dose bath safety threshold of 10 Gy, such as T = 0.2) and a smoothing factor τ = 0.05 are defined; an adaptive physical guidance soft mask M is constructed using a differentiable sigmoid function. Dose ∈R 1×H×W×D ; .

[0084] (4) Dose-characteristic weighting: The initially extracted dose distribution spatial pattern features F Conv_Dose The calculated adaptive physical guidance soft mask M Dose Perform element-wise Hadamard product to complete physical prior weighting, and obtain the dose feature representation F, which is physically prior filtered and focuses on the high-dose region. Dose .

[0085] (5) Cross-modal feature fusion: The dose characteristic obtained after dose characteristic weighting is represented as F. Dose After feature projection through a 1×1×1 3D learnable point convolution, it is then combined with the extracted shallow features F of the 3D planning CT image. CT The summation and fusion yields a unified embedding representation that integrates anatomical structure information (from 3D planning CT images) and treatment planning information (from 3D helical tomography dose maps). The tensor shape is B×64×128×128×64, which is the superficial combined dose embedding feature F. embed , represented as F embed =F CT +W proj *F Dose .

[0086] Step 3: Evolution of features of the dose-modulated 3D Swin Transformer backbone network: (1) Initialization: In the shallow combined dose embedding feature F embed After being fed into the 4-stage dose modulation 3D Swin Transformer backbone network, the shallow joint dose embedding feature F is... embed Assigning values ​​to the backbone network feature F backbone This serves as the starting point for the input in the subsequent four stages of processing.

[0087] (2) Four-stage cycle of the backbone network: The four-stage feature extraction of the dose-modulated 3D Swin Transformer backbone network is enabled, iterating from stage 1 to stage 4. In each stage, the backbone network features F of the current stage are extracted. backbone The system is divided into non-overlapping 3D local windows Ω, with a window size of W×W×W. Within each window, the relative position offset of dose modulation, dynamic window size adjustment, and self-attention calculation are performed independently in sequence. Finally, the deep spatial comprehensive feature map is output by the fourth stage located at the top layer.

[0088] (3) Calculate the dose modulation relative position bias (DM-RPB): A dose modulation relative position bias (DM-RPB) mechanism is introduced. Specifically, for each local three-dimensional window Ω, the input three-dimensional helical tomography dose map tensor X is used to... Dose_norm Slice out the corresponding local dose block D Ω Calculate the mean dose D within the window Ω. mean Maximum dose D max and dose standard deviation D std The formulas for calculating these three dose statistics are as follows: ; ; ; Construct a dimensional adaptive dose adjustment factor φ that reflects the spatial gradient of the low-dose bath within the window. Dose The calculation formula is as follows: ; in, α These are learnable scalar coefficients that are automatically updated via backpropagation within the network, with 1 as the baseline and ε=10. -6 tanh(·) is the Tanh activation function; Standard 3D relative position bias standard Multiply by the calculated dimensional adaptive dose adjustment factor φ Dose Multiplicative weighting is performed to complete self-attention mapping and generate the relative position bias of dose modulation. DM , used for calculating the self-attention matrix of dose-informed modulation.

[0089] Let the extracted query, key, and value matrices be respectively... Let the standard three-dimensional relative position offset matrix be... The formula for calculating the dose-informed modulation self-attention matrix is ​​as follows: .

[0090] (4) Predicting the three-dimensional segmentation probability map G of key cardiac substructures sub : The deep feature map output from stage 2 is upsampled and restored to its original size H×W×D through 3D transposed convolution using a lightweight upsampling decoder. The number of channels is set to K=4 (representing the background, whole heart, left ventricular basal segment, and LAD vascular region). The output is a 3D segmentation probability map of the whole heart, left ventricular basal segment, and LAD vascular region, i.e., a 3D segmentation probability map G of key cardiac substructures. sub .

[0091] (5) Dynamic window size adjustment: In stages 3 and 4 of the dose-modulated 3D Swin Transformer backbone network, the window size of the 3D attention mechanism no longer remains constant. At this point, a cardiac substructure-guided adaptive window attention (CAG-AWA) mechanism is introduced to read the 3D segmentation probability map G of the cardiac substructure predicted by the tail in stage 2. sub The model calculates the mean probability of cardiac substructure segmentation; in the forward feature transfer of stages 3 and 4, the model calculates the size M of each dynamic window based on the calculated mean probability of cardiac substructure segmentation and the discretization calculation logic. adaptive .

[0092] Each dynamic window is M in size. adaptive The discretization calculation logic is as follows: ; In this formula, the baseline window size Mbase is set to 7, β is a learnable dilation control weighting coefficient, and MeanPixels calculates the mean probability of the key substructure mask within the local block region. Using this formula, the window size automatically expands to 9 or 11 in irradiated hazardous areas with complex anatomical boundaries, and remains at 7 or shrinks to 5 in normal tissue areas.

[0093] Thus, the size of the Swing local sliding attention window can be dynamically and adaptively adjusted between 5×5×5, 7×7×7, 9×9×9 and 11×11×11. For regions with more complex or important cardiac substructures, the window size will be increased accordingly so that the model can capture the feature information of the region more comprehensively; while for regions with relatively simple substructures, the window size will be kept smaller to improve computational efficiency.

[0094] (6) Self-attention execution: The window data is re-divided or processed using dynamically calculated window sizes, and standard 3D Swin self-attention computation is performed in conjunction with dose-modulated relative positional biases. This means that the attention range dynamically changes with the importance of anatomical structures, and the attention weights are modulated by local dose distribution characteristics.

[0095] Step 4: One-dimensional multimodal asynchronous feature encoding and multi-head cross-attention fusion: (1) Perform three-dimensional global average pooling on the deep spatial comprehensive feature map output by the dose modulation 3D Swin Transformer backbone network to reduce the dimensionality and obtain a one-dimensional spatial anatomy-dose total feature vector F. spatial ∈R 512 .

[0096] (2) A multilayer perceptron (MLP) consisting of three fully connected layers, InstanceNorm1D, and LeakyReLU is used to process the input one-dimensional clinical variable vector V. clin_norm ∈R 214 The projection is a one-dimensional clinical variable feature vector F clin ∈R 512 .

[0097] (3) First, a set of one-dimensional convolutional layers is used to process the input baseline standard 12-lead electrocardiogram time tensor X. ECG_norm ∈R 12×5000 Temporal downsampling is performed, followed by the use of a temporal network consisting of four standard Transformer Encoders to capture long-term heart rhythm and waveform abnormalities, thereby encoding a one-dimensional electrophysiological feature vector F. ECG ∈R 512 .

[0098] (4) Define the learnable projection matrix W Q W K W V ∈R 512×512 The spatial anatomy-dose total feature vector F spatial ∈R 512 As a query, the one-dimensional clinical variable feature vector F clin ∈R 512 With one-dimensional electrophysiological eigenvector F ECG ∈R 512 The feature vector F obtained after concatenation context =[F clin ;F ECG ]∈R 1024 Different linear layers are mapped to keys and values, respectively. These are then input into a multi-head cross-attention layer to perform high-order modality fusion, outputting a fused feature vector F. fusion ∈R 512 .

[0099] Query, Key, Value, and Fusion Feature Vector F fusion The calculation formula is as follows: ; ; ; .

[0100] Step 5: Monte Carlo random inactivation inference and high-confidence prediction output: F will be fused with feature vector F fusion ∈R 512 The input is a fully connected classification layer with a Monte Carlo random deactivation mechanism, where the random deactivation rate of the Monte Carlo random deactivation mechanism is p=0.3. This fully connected classification layer automatically triggers an iterative controller to process the input fused feature vector F. fusion ∈R 512 The forward classification prediction is performed independently and repeatedly N times (e.g., N=10). The Monte Carlo random deactivation mechanism remains enabled throughout the test inference phase to ensure that the output of the fully connected classification layer has confidence, thereby obtaining the N classification probability prediction values. Finally, calculate the predicted probability values ​​for these N classification iterations. algebraic mean Y represents the probability of developing early subclinical cardiac injury. prob And calculate the predicted classification probability values ​​for these N iterations. discrete variance The uncertainty score Y is the result of the system prediction. uncert。

[0101] Algebraic Mean and discrete variance The calculation formula is as follows: ; ; Where σ(·) is the standard binary classification sigmoid activation function.

[0102] To verify the significant advancements and substantial features of this invention compared to existing technologies, the researchers conducted comprehensive benchmark and ablation experiments on a real clinical cohort comprising 400 left breast cancer patients undergoing TOMO radiotherapy at two centers. The internal development set included 300 cases from the first center, divided into a training set and an independent internal test set (n=30). The external validation set included 100 completely independent patient cases from the Affiliated Hospital of Xiangnan University, used to assess the cross-center generalization ability of the system.

[0103] 1. Horizontal benchmark comparison experiment with current mainstream / cutting-edge prediction schemes.

[0104] The complete system proposed in this invention was rigorously compared with four prediction schemes commonly used in clinical practice and at the forefront of the medical computer vision field: Comparison with Option 1 (Traditional Dosimetric Scalar Regression Model): Extract the dosimetric scalar parameters most commonly used by radiation oncologists, such as mean cardiac dose (MHD) and mean LAD vascular dose, and construct a traditional multivariate stepwise logistic regression model.

[0105] Comparison Scheme 2 (Traditional 3D Radiomics Machine Learning Scheme): Within the whole-organ ROI of the heart delineated by TPS, 107 traditional omics features containing first-order statistics, 3D shape and texture (GLCM, GLSZM) are extracted using Pyradiomics, and the classic Random Forest (RF) classifier is used for feature selection and classification.

[0106] Comparison Scheme 3 (Pure 3D Convolutional Neural Network End-to-End Model): The classic 3DResNet-18 is used as the backbone network for image classification. The preprocessed 3D planning CT and 3D dose map are directly concatenated along the channel dimension as input, without multimodal or auxiliary tasks.

[0107] Comparison Scheme 4 (Standard Medical Transformer Architecture): Adopts the standard 3D SwinUNETR classification network variant from the MONAI library, without special "dose-informed relative position bias" and "ECG multimodal fusion".

[0108] Option 5 (the complete system proposed in this invention): a full-featured version integrating a dose embedding module, dose modulation relative position bias, adaptive window attention, multi-task assisted segmentation, and a 12-lead ECG cross-attention post-fusion head.

[0109] The objective and quantitative indicators for evaluating the dual-center independent cohort are shown in Table 1: Table 1: Horizontal Comparison of Comprehensive Performance Indicators of the Prediction System of this Invention with Current Mainstream / Cutting-Edge Technologies

[0110] Comparative experimental quantitative conclusion analysis: As can be seen from the extremely high granularity scientific data in Table 1, the present invention (Solution 5) exhibits an overwhelming advantage in all indicators.

[0111] (1) In the internal test set, the ROC-AUC of the present invention reached an extremely high 0.852, representing an absolute improvement of 15.8% compared to the current clinical normal tissue scalar model of complications (Scheme 1, AUC=0.694). The DeLong test showed that this improvement was statistically significant ($P < 0.01$).

[0112] (2) On the external validation set of the second center (Xiangnan University Affiliated Hospital) that had not participated in the training, in the face of data drift caused by different equipment and different drawing habits, the AUC of the traditional deep learning schemes (Scheme 3 and Scheme 4) showed a significant cliff drop (dropped to 0.732 and 0.751 respectively), while the scheme of this invention remained stable at a high level of 0.784, and the Brier score was as low as 0.176 (significantly lower than all other comparative variants), which strongly confirms that the "dose notification mechanism and high-order late-stage multimodal fusion" proposed in this invention can greatly filter out hardware-specific noise and has extremely strong clinical generalization and promotion value.

[0113] 2. System-level ablation study of the core technological innovation components of this invention.

[0114] To further demonstrate that the various improved features (DEM, DM-RPB, CAG-AWA, MTCSA, ECG modality) independently designed in this invention are not simply structural stacking, but each possesses a clear and substantial technical contribution, the researchers of this invention performed rigorous ablation experiments on an internal independent test set (n=30): Variant Model M1 (Pure Image Baseline Network): This model removes all the improvements of this invention and reverts to the standard 3DSwin Transformer classification network. The input is only a simple physical stitching channel of anatomical CT and normalized dose map, without any multi-task or multimodal cascading.

[0115] Variant model M2 (w / o DM-RPB-removed dose modulation relative position bias): In the complete system of this invention, the introduced dose gradient adjustment factor operation is abolished, and the standard three-dimensional relative position bias matrix is ​​restored, which depends only on the distance between three-dimensional spatial grid points.

[0116] Variant model M3 (w / o CAG-AWA-removed adaptive window attention): On the full system, the dynamic scaling mechanism of the window size guided by the segmentation probability mask is cancelled, and the 3D Swin attention window size of all stages is forcibly fixed to a standard constant of 7 × 7 × 7.

[0117] Variant model M4 (w / o MTCSA - remove multi-task auxiliary segmentation supervision): During the training phase, the parallel auxiliary segmentation decoder is directly removed, that is, the multi-task adjustment weight r is directly set to 0, and the network falls back to pure binary classification cross-entropy loss with strong supervision training.

[0118] Variant model M5 (w / o 12-lead ECG - removal of one-dimensional ECG electrophysiological modality): One-dimensional ECG feature encoding branches are removed in the late fusion head, and multi-head cross-attention is applied only to the spatial anatomy-dose total feature vector F. spatial Compared with one-dimensional clinical variable characteristics F clin It takes place between these two periods.

[0119] Variant Model M6 (Complete Full-Function Architecture Model of the Invention): A complete system version containing all innovative and improved components.

[0120] The quantitative data records for each stripping and ablation step are shown in Table 2: Table 2: Contribution and Gain Table of Experimental Performance Indicators of the Core Improved Components of the System of the Invention in the Peel-off Ablation Experiment

[0121] Substantive technical analysis of quantitative data from ablation experiments: The quantitative data from the ablation experiments directly established the inventive merit and necessity of each innovative improvement in this invention: (1) The technical necessity of DM-RPB (dose modulation relative position offset): When we remove the dose relative position offset unique to this invention (the complete model M6 is downgraded to variant M2), the AUC drops dramatically, from 0.852 to 0.793. This strongly demonstrates that relying solely on conventional spatial relative distance is completely insufficient to cope with the low-dose radiation immersion (low-dose bath) mode unique to TOMO (Tomotherapy of Helical Motion) and characterized by highly heterogeneous spatial spiral accumulation. It is necessary to rely on the dimensionless adaptive dose adjustment factor φ designed in this invention. Dose Only by performing pixel-level correction on the attention weight map can the model lock onto the winning features.

[0122] (2) High-order information complementarity of 12-lead digital electrocardiogram: In the field of medical classification, how to efficiently integrate multiple modalities has always been a challenge. Experiments show that when we remove the branches of the baseline one-dimensional 12-lead electrocardiogram signal (the complete model M6 is downgraded to variant M5), the model's discrimination AUC drops by 4.3% (from 0.852 to 0.809), and the F1-Score also declines significantly. This key finding has significant technical implications, as it confirms with objective scientific data that the hidden electrophysiological abnormalities implied in the baseline electrocardiogram (such as potential weak T-wave inversion, weak ST-segment normality, and continuous temporal characteristics that cannot be reflected by scalar indicators like heart rate variability) have extremely strong physical-pathological information complementarity with the geometric dose distribution of radiation in three-dimensional space. The two are deeply integrated through the multi-head cross-attention Late Fusion architecture specially designed in this invention, which can perfectly eliminate the inherent blind spots of single-modal prediction, forming a significant technological leap of "1+1>2".

[0123] Compared with the prior art, the significant progress and substantial technical advantages of this invention are as follows: 1. Overcoming the technical bottleneck of lack of physical priors: Traditional medical image classification networks can only passively learn grayscale features. The invention's unique "dose embedding module" and "dose modulation relative position bias" establish a "physical guidance channel" within the self-attention mechanism, allowing the network to directly perceive the complex three-dimensional spatial heterogeneity of the low-dose bath in TOMO spiral radiotherapy, and to give significant adaptive weight amplification to key anatomical areas on the edge of dose steepness.

[0124] 2. Multi-task constraints significantly improve anatomical boundary perception: The auxiliary segmentation network introduced in this invention forces the backbone network to extract high-quality anatomical boundary representations of cardiac substructures (such as the LAD and left ventricular basal segment) before making classification decisions, which solves the common problems of pure end-to-end classification networks that are prone to background noise and lack clinical interpretability.

[0125] 3. Pioneering four-dimensional modal high-order late-stage fusion of anatomy, dose, electrophysiology and clinical: This invention is the first to integrate the high-order temporal waveform characteristics of a one-dimensional standard 12-lead electrocardiogram signal into radiotoxicity prediction. It utilizes a multi-head cross-attention head to deeply integrate the hidden electrophysiological pathological signals with spatial geometric dose characteristics. Ablation experiments have confirmed that this fusion strategy can directly raise the upper limit of the discrimination AUC.

[0126] The present invention also proposes a computer device, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus, and the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform the operation corresponding to the above-mentioned method for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy.

[0127] The present invention also proposes a computer-readable storage medium storing at least one executable instruction that causes a processor to perform operations corresponding to the above-described method for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy.

[0128] It should be specifically noted that the system and method provided by this invention only offer a probability prediction of cardiac injury, falling under the category of risk assessment, rather than direct disease diagnosis. This invention essentially combines computer software with medical technology, constituting a data processing and analysis system, and is not a disease diagnosis method; therefore, it falls within the scope of patent protection.

[0129] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A system for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy, characterized in that, include: The data input and preprocessing module is responsible for acquiring and preprocessing the patient's original three-dimensional planning CT image, original three-dimensional helical tomography dose map, original baseline standard 12-lead electrocardiogram time series signal and clinical variables, and outputting three-dimensional planning CT image tensor, three-dimensional helical tomography dose distribution tensor, baseline standard 12-lead electrocardiogram time series tensor and one-dimensional clinical variable vector. The dose embedding module is responsible for extracting and fusing shallow spatial features from the 3D planning CT image tensor and the 3D helical tomography dose distribution tensor through physical prior mask weighting and a learnable projection matrix, and outputting shallow joint dose embedding features. The dose modulation 3D Swin Transformer backbone network is responsible for using a 4-stage 3D Swin Transformer that integrates a dose modulation relative position bias mechanism and a cardiac substructure-guided adaptive window attention mechanism to perform 4-stage deep feature extraction on shallow joint dose embedding features and output a deep spatial comprehensive feature map. The multi-task cardiac substructure attention block is connected to the intermediate stage feature output of the 4-stage 3D Swin Transformer. It is responsible for reconstructing the 3D segmentation probability map and the 3D prediction mask of the key cardiac substructure through a lightweight upsampling decoder. The 3D segmentation probability map of the key cardiac substructure serves as the calculation basis for the cardiac substructure-guided adaptive window attention mechanism during inference. The 3D prediction mask of the key cardiac substructure serves as an auxiliary segmentation task to implement joint training with strong supervision during training. The multimodal late-stage cross-attention fusion head is responsible for extracting features from the one-dimensional clinical variable vector and the baseline standard 12-lead ECG time tensor, respectively. Then, through the multi-head cross-attention mechanism, the extracted one-dimensional clinical variable feature vector, one-dimensional electrophysiological feature vector and the spatial anatomy-dose total feature vector obtained by three-dimensional global average pooling of the deep spatial comprehensive feature map are fused in a high-order late stage to output the fused feature vector. The uncertainty prediction output head is responsible for propagating the fused feature vector forward multiple times and using the Monte Carlo Dropout mechanism to ensure the confidence level of the prediction results. Finally, it outputs the probability of early subclinical cardiac injury and the uncertainty score of the system prediction results.

2. The prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy according to claim 1, characterized in that, The data input and preprocessing module includes: The multimodal heterogeneous data acquisition unit is responsible for acquiring the patient's original three-dimensional planning CT images and corresponding original three-dimensional helical tomography radiotherapy dose maps in the radiotherapy planning system; reading the patient's original baseline standard 12-lead electrocardiogram timing signals; and collecting 24-dimensional clinical variables including patient age, BMI, history of hypertension, whether anthracycline / cyclophosphamide chemotherapy drugs were used and their cumulative dose, and peak left ventricular outflow tract velocity. The 3D image axial rigid registration unit is responsible for using the SimpleITK toolbox to perform 3D image axial rigid registration on the acquired original 3D planning CT image as the reference image. The resampling unit is responsible for resampling the registered 3D planning CT image and the 3D spiral tomography dose map to a resolution of 256×256×128 voxels, with the voxel spatial physical size fixed at 1mm×1mm×2mm. The outlier window clipping unit is responsible for uniformly clipping the window width and window level of the resampled 3D planning CT image to [-200, 300] HU and performing conventional Z-score normalization; extracting the absolute dose value from the resampled 3D helical tomography dose map, dividing the absolute dose value by the patient's actual prescribed total dose scalar value, and linearly mapping it to the [0, 1] normalization interval to obtain the 3D planning CT image tensor and the 3D helical tomography dose map tensor; The ECG digital filtering unit is responsible for removing baseline drift and electromyography artifacts from the original baseline standard 12-lead ECG digital signal by passing a 0.5-100Hz bandpass filter and a 50Hz power frequency notch filter; uniformly extracting the first 5000 time-domain sampling points, performing maximum and minimum value normalization, and reshaping it into a 12×5000 baseline standard 12-lead ECG time series tensor; The clinical variable one-hot encoding unit is responsible for encoding the categorical variables in the 24-dimensional clinical variable indicators of the patient using One-Hot encoding, standardizing the continuous scalar variables in the 24-dimensional clinical variable indicators of the patient using mean and variance, and then combining them into a one-dimensional clinical variable vector.

3. The prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy according to claim 1, characterized in that, The dose embedding module includes: The CT feature extraction branch is responsible for extracting features from the tensor of the 3D planning CT image through 3×3×3 convolution, and outputting the shallow features of the 3D planning CT image. The dose feature extraction branch is responsible for extracting features from the tensor of the three-dimensional helical tomography dose map through 3×3×3 convolution, and outputting the shallow features of the three-dimensional helical tomography dose map. The physical prior dose soft mask calculation unit is responsible for calculating the adaptive physical guidance soft mask based on the raw data of the three-dimensional helical tomography dose map tensor, the normalized physical dose threshold constant, and the smoothing factor. The dose feature weighting unit is responsible for performing element-wise Hadamard product between the initially extracted dose distribution spatial pattern features and the calculated soft mask to achieve physical prior weighting and output the dose feature representation focused on the high dose region after physical prior filtering. The cross-modal feature fusion unit is responsible for projecting the dose feature representation obtained after physical prior weighting through a 1×1×1 three-dimensional learnable point convolution, and then adding it to the shallow features of the extracted three-dimensional planned CT image, outputting a unified embedded representation that integrates anatomical structure information and treatment plan information, namely, shallow joint dose embedding feature.

4. The prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy according to claim 1, characterized in that, The dose modulation 3D Swin Transformer backbone network includes: The 4-stage 3D Swin Transformer is responsible for feature evolution. It performs four stages of deep feature extraction on the backbone network features formed after shallow joint dose embedding feature assignment, progressively transforming the input backbone network features into more semantically informative feature representations. Specifically, in each of the four stages, a series of Transformer blocks and window attention mechanisms are used to progressively extract and transform features from the input image. From low-level edge and texture features to high-level semantic and structural features, the abstraction level of the features is gradually improved. Furthermore, through changes in feature map size and channel number adjustment at different stages, multi-scale feature fusion is achieved, ultimately outputting a deep spatial comprehensive feature map. The dose modulation relative position bias mechanism is responsible for optimizing the position bias in the self-attention calculation during the four-stage deep feature extraction process, thereby enhancing the model's attention to and discrimination ability of different dose regions in cardiac images. Specifically, it first calculates three dose statistics—mean dose, maximum dose, and standard deviation—within each window based on the input three-dimensional helical tomotherapy dose map tensor and window coordinates. Then, it calculates the ratio of the dose range to the standard deviation, scales it using the Tanh activation function, and adds a baseline value of 1 and a learnable parameter. α The dimensional adaptive dose adjustment factor reflecting the spatial gradient of the low-dose bath within the window is calculated. Finally, the standard 3D relative position bias is multiplied by the calculated dimensional adaptive dose adjustment factor, and multiplicative weighting is performed to complete the self-attention mapping and generate the relative position bias of dose modulation, which is used to calculate the self-efficacy matrix of dose informative modulation. The cardiac substructure-guided adaptive window attention mechanism is responsible for dynamically adjusting the size of the attention window during the four-stage deep feature extraction process. This allows the model to adaptively focus on substructure regions of different sizes and complexities in cardiac images, improving the ability to identify and segment fine cardiac structures. Specifically, in stages 3 and 4, the mean segmentation probability of the cardiac substructure is calculated using the 3D segmentation probability map of key cardiac substructures predicted from the tail of stage 2. Then, based on the calculated mean segmentation probability and discretization logic, the size of each dynamic window is calculated. For regions with more complex or important cardiac substructures, the window size will be increased accordingly, while for regions with relatively simple substructures, the window size will remain smaller. By combining the dose-modulated relative position bias mechanism with the cardiac substructure-guided adaptive window attention mechanism, the 4-stage 3D Swin Transformer model can better capture the complex features and structural information in cardiac images.

5. The prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy according to claim 1, characterized in that, The multi-task cardiac substructure attention block includes a parallel auxiliary segmentation decoder derived at the deep feature map transition point of the second stage of the 4-stage 3D Swin Transformer. The auxiliary segmentation decoder consists of consecutive 2×2×2 three-dimensional transposed convolutions and residual convolution blocks. It is responsible for upsampling the deep feature map output from the second stage of the 4-stage 3D Swin Transformer back to its original size H×W×D, with the number of channels set to K=4. It outputs three-dimensional segmentation probability maps and three-dimensional prediction masks for the whole heart, the left ventricular basal segment, and the LAD vascular region, i.e., the three-dimensional segmentation probability maps and three-dimensional prediction masks for key cardiac substructures. The auxiliary segmentation decoder participates in the global loss function penalty during the training period, forcing the bottom backbone of the 4-stage 3D Swin Transformer to learn high-risk anatomical boundaries.

6. The prediction system for early subclinical cardiac injury of left breast cancer after helical tomotherapy according to claim 1, characterized in that, The multimodal late-stage cross-attention fusion head includes: The one-dimensional clinical variable encoder uses a multilayer perceptron consisting of three fully connected layers, InstanceNorm1D, and LeakyReLU to extract features from the input one-dimensional clinical variable vector and output a one-dimensional clinical variable feature vector. A one-dimensional electrocardiogram encoder includes a set of one-dimensional convolutional layers and a temporal network consisting of four standard Transformer Encoder layers; the one-dimensional convolutional layers are responsible for temporal downsampling the input baseline standard 12-lead electrocardiogram temporal tensor; the temporal network is responsible for capturing long-term heart rhythm and waveform abnormalities in the temporal downsampling results and outputting a one-dimensional electrophysiological feature vector. The three-dimensional global average pooling unit is responsible for performing three-dimensional global average pooling on the deep spatial comprehensive feature map output by the dose modulation 3D Swin Transformer backbone network, and outputting the spatial anatomy-dose total feature vector. The multi-head cross-attention late-stage fusion unit is responsible for using the multi-head cross-attention mechanism to perform high-order late-stage fusion of one-dimensional clinical variable feature vectors, one-dimensional electrophysiological feature vectors, and spatial anatomy-dose total feature vectors, and outputting a fused feature vector.

7. The prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy according to claim 1, characterized in that, The uncertainty prediction output header includes: The fully connected classification layer, connected to the output of the multimodal late-stage cross-attention fusion head, is responsible for performing multiple independent random inactivation forward propagations on the fused feature vectors to infer a series of classification probability predictions. Then, by calculating this series of classification probability prediction values algebraic average and discrete variance Ultimately, the probability of early subclinical cardiac injury and the uncertainty score of the system prediction results are obtained; The Monte Carlo random deactivation mechanism, deployed before the fully connected classification layer, is responsible for causing the fully connected classification layer to perform forward classification prediction with a random deactivation rate of p=0.3 during test inference, in order to ensure the confidence of the output results.

8. A method for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy based on the prediction system for early subclinical cardiac injury in left breast cancer after helical tomotherapy as described in any one of claims 1-7, characterized in that, Includes the following steps: Step 1: Acquisition and preprocessing of multimodal heterogeneous data: (1) Rigid registration of three-dimensional images: acquire the original three-dimensional planning CT image of the patient in the radiotherapy planning system and the corresponding original three-dimensional helical tomography radiotherapy dose map; use the SimpleITK toolbox to perform rigid registration of three-dimensional images on the original three-dimensional planning CT image as the reference image. (2) Resampling and outlier window clipping: The resolution of the registered three-dimensional planning CT image and the three-dimensional spiral tomography dose map was resampled to 256×256×128 voxels, and the physical size of the voxel space was fixed at 1mm×1mm×2mm. The window width and window level of the resampled 3D planning CT image were uniformly cropped to [-200, 300] HU and normalized using conventional Z-score. The absolute dose value of the resampled 3D helical tomography dose map was extracted, and the absolute dose value was divided by the patient's actual total prescribed dose scalar value and linearly mapped to the normalized interval [0, 1] to obtain the 3D planning CT image tensor and the 3D helical tomography dose map tensor. (3) Digital filtering of electrocardiogram: Read the baseline standard 12-lead electrocardiogram digital signal, and remove baseline drift and electromyography artifacts by passing a 0.5-100Hz bandpass filter and a 50Hz power frequency notch filter; uniformly extract the first 5000 time domain sampling points, perform maximum and minimum value normalization, and reshape it into a 12×5000 baseline standard 12-lead electrocardiogram time series tensor; (4) One-Hot encoding of clinical variables: 24-dimensional clinical variables including patient age, BMI, history of hypertension, whether anthracycline / cyclophosphamide chemotherapy drugs were used and the cumulative dose, and peak flow velocity of the left ventricular outflow tract were collected. The categorical variables were encoded using One-Hot encoding, and the continuous scalar variables were standardized using mean and variance, and combined into a one-dimensional feature vector of clinical variables. Step 2: Shallow combined dose embedding feature extraction: (1) CT feature extraction: Feature extraction of the 3D planning CT image tensor is performed by 3×3×3 convolution to obtain shallow features of the 3D planning CT image; (2) Preliminary extraction of dose characteristics: Feature extraction of the three-dimensional helical tomography dose map tensor was performed by 3×3×3 convolution to obtain shallow features of the three-dimensional helical tomography dose map. (3) Calculate the physical prior dose soft mask: Based on the raw data of the three-dimensional helical tomography dose map tensor, the normalized physical dose threshold constant, and the smoothing factor, an adaptive physical-guided soft mask is calculated to force the model to explicitly perceive the radiotherapy risk area in the forward computation. (4) Dose-characteristic weighting: The initially extracted spatial pattern features of dose distribution are combined with the calculated adaptive physical guidance soft mask by element-wise Hadamard product to complete the physical prior weighting, and the dose feature representation focused on the high dose region after physical prior screening is obtained. (5) Cross-modal feature fusion: The dose feature representation obtained after dose feature weighting is projected through a 1×1×1 three-dimensional learnable point convolution and then fused with the extracted shallow features of the three-dimensional planning CT image to obtain a unified embedding representation that integrates anatomical structure information and treatment plan information, namely, shallow joint dose embedding feature. Step 3: Evolution of features of the dose-modulated 3D Swin Transformer backbone network: (1) Initialization: After the shallow joint dose embedding features are fed into the 4-stage dose modulation 3D Swin Transformer backbone network, the shallow joint dose embedding features are assigned to the backbone network features as the starting point for the subsequent four stages of processing. (2) Four-stage cycle of the backbone network: The four-stage feature extraction of the dose modulation 3D Swin Transformer backbone network is enabled. The process iterates from stage 1 to stage 4. In each stage, the backbone network features of the current stage are divided into non-overlapping 3D local windows with a window size of W×W×W. Within each window, the dose modulation relative position bias calculation, dynamic window size adjustment, and self-attention calculation are performed independently in sequence. Finally, the deep spatial comprehensive feature map is output by stage 4 at the top layer. (3) Calculate the relative positional offset of dose modulation: A dose modulation relative position bias mechanism is introduced. Within each self-attention calculation window, a corresponding local dose block is sliced ​​from the input 3D helical tomography dose map tensor, and the mean dose, maximum dose, and standard deviation of the dose in the current window are calculated. The ratio of the range to the standard deviation of the dose is calculated, scaled by the Tanh activation function, and a baseline value of 1 and a learnable parameter are added. α The dimensional adaptive dose adjustment factor reflecting the spatial gradient of the low-dose bath within the window is calculated. The standard 3D relative position bias is multiplied by the calculated dimensional adaptive dose adjustment factor, and multiplicative weighting is performed to complete the self-attention mapping, generate the relative position bias of dose modulation, and use it to calculate the dose-information modulation self-attention matrix. (4) Predicting the three-dimensional segmentation probability map of key cardiac substructures: The deep feature map output from the second stage is upsampled and restored to its original size H×W×D by the three-dimensional transposed convolution of the lightweight upsampling decoder. The number of channels is set to K=4, and the three-dimensional segmentation probability map of the whole heart, the basal segment of the left ventricle, and the LAD vascular region is output, which is the three-dimensional segmentation probability map of key cardiac substructures. (5) Dynamic window size adjustment: A cardiac substructure-guided adaptive window attention mechanism is introduced. Based on the 3D segmentation probability map of the cardiac substructure obtained from the tail prediction in stage 2, the mean segmentation probability of the cardiac substructure is calculated. In the forward feature transfer in stages 3 and 4, the model calculates the size of each dynamic window based on the calculated mean segmentation probability of the cardiac substructure and the discretization calculation logic, thereby dynamically and discretically adjusting the size of the Swing local sliding attention window between 5×5×5, 7×7×7, 9×9×9, and 11×11×11. For regions with more complex or important cardiac substructures, the window size will be increased accordingly so that the model can capture the feature information of the region more comprehensively; while for regions with relatively simple substructures, the window size will remain smaller to improve computational efficiency. (6) Self-attention execution: The window data is re-divided or processed using dynamically calculated window sizes, and standard 3D Swin self-attention operations are performed in conjunction with dose-modulated relative positional biases; this means that the attention range changes dynamically with the importance of anatomical structures, and the attention weights are modulated by local dose distribution characteristics. Step 4: One-dimensional multimodal asynchronous feature encoding and multi-head cross-attention fusion: (1) Perform three-dimensional global average pooling on the deep spatial comprehensive feature map output by the dose modulation 3D Swin Transformer backbone network to reduce the dimension to obtain a one-dimensional spatial anatomy-dose total feature vector. (2) A multilayer perceptron consisting of three fully connected layers, InstanceNorm1D and LeakyReLU is used to project the input one-dimensional clinical variable vector into a one-dimensional clinical variable feature vector; (3) First, a set of one-dimensional convolutional layers is used to perform temporal downsampling on the input baseline standard 12-lead electrocardiogram temporal tensor. Then, a temporal network composed of 4 layers of standard Transformer Encoder is used to capture long-term heart rhythm and waveform abnormalities, thereby encoding a one-dimensional electrophysiological feature vector. (4) Define the learnable projection matrix W Q W K W V The spatial anatomy-dose total feature vector is used as the query. The feature vector obtained by concatenating the one-dimensional clinical variable feature vector and the one-dimensional electrophysiological feature vector is mapped to keys and values ​​through different linear layers. The input is then fed into a multi-head cross-attention layer to perform higher-order modality fusion and output the fused feature vector. Step 5: Monte Carlo random inactivation inference and high-confidence prediction output: The fused feature vector is input into a fully connected classification layer equipped with a Monte Carlo random deactivation mechanism. The random deactivation rate of the Monte Carlo random deactivation mechanism is p=0.

3. The fully connected classification layer automatically triggers an iterative controller to perform N independent and repeated forward classification predictions on the input fused feature vector. The Monte Carlo random deactivation mechanism remains enabled throughout the test and inference phases to ensure that the output of the fully connected classification layer has confidence, thereby obtaining N classification probability prediction values. Finally, calculate the predicted probability values ​​for these N classification iterations. algebraic average As the final risk probability of early subclinical cardiac injury, the predicted values ​​of these N classification probabilities are calculated. discrete variance As an uncertainty score of the system's prediction results; ； ； Where σ(·) is the standard binary classification sigmoid activation function.

9. A computer device, comprising: The system includes a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface communicate with each other via the communication bus. The memory stores at least one executable instruction, which causes the processor to perform the operation corresponding to the method for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy as described in claim 8.

10. A computer-readable storage medium storing at least one executable instruction that causes a processor to perform an operation corresponding to the method for predicting early subclinical cardiac injury in left breast cancer after helical tomotherapy as described in claim 8.

Images