Deep learning-based pulmonary artery sarcoma and pulmonary artery embolism intelligent identification system

Abstract

The invention discloses an intelligent identification system for pulmonary arterial sarcoma and pulmonary arterial embolism based on deep learning, and belongs to the technical field of image processing. The system comprises a data acquisition module for acquiring a CTPA image data set containing acute pulmonary embolism, chronic pulmonary embolism and pulmonary arterial sarcoma for training, internal testing and external verification, and acquiring CTPA image data for data to be identified of a model application module; the data preprocessing module is used for preprocessing to obtain data meeting the requirements of subsequent modules; the deep learning model training module is realized by adopting a supervised convolutional neural network PVDNet framework, and the preprocessed data is used for training the model to obtain a discrimination model; an internal test module; an external verification module; and the model application module identifies the real to-be-identified data through the identification model. According to the method, corresponding disease identification can be carried out on the CTPA image through the deep learning model, and the identification reliability reaches the level of qualification of cardiovascular image experts.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a deep learning-based identification and recognition system for pulmonary artery sarcoma and pulmonary embolism. Background Art

[0002] Pulmonary thromboembolism (PTE) is the third most common cause of death after acute coronary syndrome and stroke. Chronic pulmonary thromboembolism (CPE) is a long-term complication of venous thromboembolism that may develop into chronic thromboembolic pulmonary hypertension, which requires different treatment options and management strategies compared with acute pulmonary embolism (APE). The identification and diagnosis of PTE is usually based on the demonstration of pulmonary artery filling defects on computed tomography pulmonary angiography (CTPA). However, an increasing number of cases have confirmed that pulmonary artery filling defects do not always represent PTE. Although rare, pulmonary artery sarcoma (PAS) is a malignant tumor that arises from the pulmonary artery and usually has a very poor prognosis. Early diagnosis of PAS is crucial because it can prevent inappropriate anticoagulation and allow timely surgical intervention, potentially improving outcomes. Although CTPA is useful in diagnosing PTE, PAS often mimics PTE due to similar filling defects, leading to frequent misdiagnosis and ineffective anticoagulation.

[0003] Previous studies have attempted to differentiate PAS from PTE by imaging features, but the number of patients in such studies was too small and the imaging results were nonspecific. With the development of artificial intelligence (AI) technology, its introduction into clinical workflow has been proposed to facilitate the diagnosis and treatment of diseases. Deep learning (DL) models have been used to evaluate APE, differentiate constrictive pericarditis from restrictive cardiomyopathy, hypertrophic cardiomyopathy from hypertensive heart disease, and segment and predict percutaneous recanalization in patients with chronic total occlusion in coronary CT angiography. Therefore, the use of DL models is expected to assist in the differential diagnosis of PAS and PTE. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the above-mentioned related art at least to a certain extent.

[0005] To this end, the purpose of the present invention is to provide an intelligent recognition system for pulmonary artery sarcoma and pulmonary embolism based on deep learning. APE, CPE and PAS are recognized for CTPA images through a deep learning model. The recognition reliability can reach the level of a senior cardiovascular radiologist with 20 years of experience, providing a reliable solution for the automatic recognition of APE, CPE and PAS.

[0006] In order to solve the above-mentioned technical problems, the present invention is achieved as follows: An embodiment of the present invention provides a deep learning-based intelligent recognition system for pulmonary artery sarcoma and pulmonary embolism, the system comprising: A data acquisition module, which obtains a three-dimensional volumetric CTPA image data set containing APE, CPE and PAS for use as a data set for a deep learning model training module, an internal testing module and an external verification module, and obtains CTPA image data for data to be identified in a model application module; The data preprocessing module preprocesses the data collected by the data acquisition module to obtain data that meets the use requirements of subsequent modules; The deep learning model training module is implemented using the supervised convolutional neural network PVDNet framework, and the PVDNet framework is trained with the preprocessed data to obtain a trained recognition model; An internal testing module is used to perform internal testing on the trained recognition model according to the corresponding internal testing data set; The external verification module is used to identify the trained recognition model according to the corresponding external verification data set, and then perform statistical analysis on the recognition results and the manual recognition results to verify the accuracy of the recognition model; The model application module recognizes the real data to be recognized through the recognition model and determines whether APE, CPE and / or PAS exist on the corresponding CTPA image.

[0007] In addition, the deep learning-based intelligent identification system for pulmonary artery sarcoma and pulmonary embolism according to the present invention may also have the following additional technical features: In some embodiments, the data acquisition module acquires 952 three-dimensional volume CTPA image data sets including APE, CPE and PAS; among them, 590 are used as training data sets for the deep learning model training module, 186 are used as test data sets for the internal testing module, and 176 are used as verification test sets for the external verification module.

[0008] In some embodiments, the inclusion criteria of the 952 three-dimensional volumetric CTPA image datasets include: the filling defect on CTPA is central, APE has been confirmed by follow-up after anticoagulation or thrombolytic therapy, CPE has been diagnosed by follow-up and endarterectomy or balloon pulmonary angioplasty, and PAS has been confirmed by biopsy or surgical pathology.

[0009] In some of the embodiments, the processing content of the data preprocessing module includes: segmenting each three-dimensional volume CTPA image, and then extracting the lung area from each segmented image through the three-dimensional bounding box extracted by the corresponding segmentation mask, retaining the pulmonary artery and the heart; then adjusting the size of the lung area and normalizing it.

[0010] In some of the embodiments, the lung region is resized to 192×192×192, the window center is 100, the window width is 700, a window operation is performed, and the resized lung region is subjected to minimum-maximum normalization.

[0011] In some of the embodiments, the supervised convolutional neural network PVDNet framework includes a pre-trained ResNet-50, a maximum pooling module, and a fully connected layer with a softmax activation function; The ResNet-50 contains 3.8 billion trainable parameters to extract instance features; The maximum pooling module uses a maximum pooling operation to perform feature aggregation; The fully connected layer generates probability scores for classification through a softmax activation function.

[0012] In some of the embodiments, the deep learning model training module receives the 3D volume data processed by the data preprocessing module and segments it along the z-axis to obtain a number of 2D slices to form a slice bag, and then inputs the slice bag into the supervised convolutional neural network PVDNet framework for model training.

[0013] In some of the embodiments, the deep learning model training module uses a five-fold cross-validation method to perform model training; in each fold, 80% of the training data is used as a development set for model training, and the remaining 20% ​​is used as a tuning set for model selection; the model that obtains the best results on the tuning set is used for subsequent internal test sets and external validation.

[0014] In some of the embodiments, the deep learning model training module uses an optimizer to train the entire network for 50 epochs, with an initial learning rate of 1e-5 and a small batch size of 1; instance balanced sampling is used to alleviate the impact of class imbalance, and class balanced loss is selected as the loss function to train the model; a simplified lon platform method based on adjusted set loss is used to adjust the learning rate;.

[0015] In some of the embodiments, the system further comprises a visualization module which uses gradient weighted class activation mapping to visualize the decision-making process of the model.

[0016] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A block diagram of a deep learning-based intelligent identification system for pulmonary artery sarcoma and pulmonary embolism disclosed in one embodiment of the present invention; Figure 2 The PVDNet model classification block diagram disclosed in one embodiment of the present invention; wherein (a) is the preprocessing process; (b) is the classification process; Figure 3 The model performance diagram of the internal test set disclosed in one embodiment of the present invention, wherein (a) is the ROC curve; (b) is the classification confusion matrix; Figure 4 A representative Grad-CAM heat map of the APE, CPE and PAS models disclosed in one embodiment of the present invention; Figure 5 ROC curves of different diseases identified by the PVDNet model and manually according to an embodiment of the present invention; (a) is acute pulmonary thromboembolism, (b) is chronic pulmonary thromboembolism; (c) is pulmonary artery sarcoma; Figure 6 The confusion matrix between the PVDNet model recognition disclosed in one embodiment of the present invention and four manual recognitions. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The embodiments of the present invention are described in detail below through specific embodiments and application scenarios in conjunction with the accompanying drawings.

[0020] See also Figure 1 As shown, in some embodiments of the present invention, a deep learning-based intelligent recognition system for pulmonary artery sarcoma and pulmonary embolism is provided, which recognizes PAS and PTE based on CTPA images based on a deep learning model. The present invention collects a number of image data and divides them into three parts, which are used for model training, internal testing, and external testing, respectively, so as to obtain a pulmonary artery sarcoma and pulmonary embolism recognition system that meets the accuracy requirements.

[0021] 1. Data Collection The present invention collected a three-dimensional volume dataset of CTPA images of 952 patients with APE, CPE and PAS from 15 hospitals (including 542 male patients with an average age of 61±15.4 years). Among them, the CTPA image datasets of 590 patients from January 2015 to December 2019 and 186 patients from January 2020 to December 2023 were used as training sets and internal test sets, respectively; the CTPA image datasets of 176 patients from January 2021 to December 2023 were used as external validation test sets. The criteria for inclusion of the above datasets are as follows: 1. The filling defect on CTPA is central, that is, a filling defect located in the lobar artery or larger pulmonary artery; 2. APE is diagnosed by follow-up after anticoagulation or thrombolytic therapy; 3. CPE is diagnosed by follow-up and pulmonary endarterectomy or balloon pulmonary angioplasty; 4. PAS is confirmed by biopsy or surgical cases. CTPA with poor image quality or severe artifacts was excluded. CTPA without Digital Imaging or Communications in Medicine (DICOM) format was excluded. Patients with loss of follow-up or diagnosed with other types of pulmonary embolism, including echinococcosis pulmonary embolism and pulmonary artery tumor embolism, pulmonary artery lipoma, Takayasu arteritis, fibrosing mediastinitis, and Behçet's disease were excluded.

[0022] 2. CTPA Scanning Protocol CTPA was performed using a multi-detector CT scanner. The entire chest was scanned from the apex to the diaphragm. Scanning parameters included: current 300550 mA with dose adjustment based on individual body mass index (BMI); tube voltage 100-120 kV, collimation 0.625-1 mm, and reconstruction increment 1 mm. Iodopyramide was injected intravenously using a mechanical injector at a dose of 370 mg / ml at a flow rate of 4.5-5.0 ml / sec. The region of interest was located at the level of the main pulmonary artery using an automated tracking technique with a predefined threshold of 80-100 HU.

[0023] 3. Image preprocessing All CTPA image datasets in DICOM format were preprocessed in several steps. Specifically, lung image segmentation was first performed on each CTPA scan using a deep learning approach, which can be implemented using existing medical imaging solution software. Lung regions were then extracted from each CT scan using the 3D bounding boxes extracted by the corresponding segmentation masks, retaining the pulmonary artery and heart. Each obtained lung region was resized to 192×192×192. The window center was 100, the window width was 700, and the window operation was performed with min-max normalization on the resized lung regions.

[0024] IV: Deep learning model and model training The present invention adopts the supervised convolutional neural network PVDNet framework for fine-grained classification. Figure 2 The pre-trained ResNet-5025, a 50-layer residual framework deep learning model with 3.8 billion trainable parameters, is used to extract instance features, and the maximum pooling operation is used for feature aggregation. The combined features are input into a fully connected layer, followed by a softmax activation function to generate a probability score for each type (APE, CPE, or PAS).

[0025] Specifically, for CTPA images, the above preprocessing is first performed on them, and the processed images are segmented on the z-axis to obtain several image slices. Each image slice obtained is copied into a three-channel image as a two-dimensional instance. All instances are formed into a slice bag. Each three-channel image in the bag is passed to the PVDNet model for training.

[0026] Model training was performed using a five-fold cross validation method. In each fold, the present invention split the patient-level training data, with approximately 80% of the training set data used as the training set for model construction and the remaining 20% ​​used as the tuning set for model selection. After training, the model that can obtain the best results on all tuning sets is applied to the internal test set and the external validation set.

[0027] The weights of ResNet-50 in the PVDNet model are pre-trained by ImageNet, and the classification layer (fully connected layer) is initialized using the Kaiming method. The entire network is trained for 50 epochs using the Adam optimizer, with an initial learning rate of 1e-5 and a mini-batch size of 1. An epoch refers to a training process in which all samples in the training dataset are passed through once (and only once). In an epoch, the training algorithm inputs all samples into the model in a set order for forward propagation, loss calculation, backpropagation, and parameter update. The learning rate is an important hyperparameter that controls the step size of weight update. A smaller learning rate means that the model updates the weights more cautiously during training; the mini-batch size is set to 1, and each gradient update is based on a single sample. Considering the small sample size of PAS in the training set, instance balanced sampling is used to mitigate the impact of class imbalance, and class balanced loss is selected as the loss function to train the model. During the training process, a simplified lon platform method based on the adjusted set loss is used to adjust the learning rate. If the loss no longer decreases or starts to increase, the learning rate is reduced; if the loss continues to decrease rapidly, the learning rate can be considered to be maintained or appropriately increased. Data augmentation is performed by a toolkit called batch generator. During the training phase, the performance of the model on the tuning set is evaluated by the macro F1 score. In the present invention, the average F1 of the 5 selected models on the tuning set is 0.838±0.051. By combining instance-balanced sampling, class-balanced loss function and a simplified method based on the tuning set loss to train the model, the problem of class imbalance in the training set can be effectively addressed, and the generalization ability and performance of the model can be improved.

[0028] V. Statistical Analysis In this part, the model judges the performance of the model in classifying PAS, APE and CPE. Python3.7 was used for statistical analysis. The performance of PAS, APE and CPE classification was evaluated by area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Bootstrap (1000 times) was used to estimate the 95% confidence interval (CI) of each evaluation indicator. In addition, the DeLong test was used to compare the relationship between the AUC of the model and the manual recognition and judgment results of radiologists of different years. The manual recognition and judgment is to manually recognize and judge the display results of CTPA images by several specific radiologists.

[0029] The number of true positive, false positive, true negative, and false negative results of the classification performance are also described in a 3×3 contingency table representing the confusion matrix by the prediction results on the internal test and external validation sets. Cohen's kappa coefficient was used to measure the agreement between the model prediction and the true label. If the kappa coefficient is 0~0.20, the level of agreement is interpreted as slight; fair, 0.21 to 0.40; moderate, 0.41 to 0.60; substantial, 0.61 to 0.80; almost perfect, 0.81 to 1.

[0030] VI: Analysis of research results The previously described 3D volumetric CTPA image dataset of 952 patients included 470 APE patients (249 males, mean age = 65.5 ± 15.2 years), 363 CPE patients (235 males, mean age = 58.4 ± 14.0 years), and 119 PAS patients (58 males, mean age = 51.5 ± 14.6 years).

[0031] The model performance on the internal test set is as follows: Table 2 shows the evaluation results of the classification models in the internal test set. According to the results, the PVDNet model predictions and true labels of the present invention are well correlated with the three diseases. The AUCs of APE, CPE, and PAS are 0.902 (95% CI: [0.855, 0.944]), 0.900 (95% CI: [0.852, 0.946]), and 0.972 (95% CI: [0.945, 0.994]), respectively.

[0032]

[0033] The ROC curve (that is, the receiver operating characteristic curve) is as follows Figure 3 In addition, in the internal validation set, there is a significant agreement between the model predictions and the ground truth (kappa = 0.687, p < 0.0001). Figure 3 (b) is the classification confusion matrix, which shows the number of true positive results, false positive results, true negative results, and false negative results predicted by the model.

[0034] To enhance the interpretability of the model, Gradient Weighted Class Activation Mapping (Grad-CAM) is employed. Figure 4Representative Grad-CAM heatmaps for the APE, CPE, and PAS models. For each sample, the original CT image (right) and the corresponding attention heatmap (left) are paired. The red in the heatmap highlights the activation areas associated with the predicted class, and the darker the color, the stronger the association. The orange arrow indicates the lesion in the original image at the actual location of the object. The Grad-CAM heatmap shows that the model of the present invention is able to pay attention to the abnormalities of the image when making decisions.

[0035] Model performance on the external validation set: The performance of the PVDNet model was further evaluated in an external validation set and compared with four radiologists. The evaluation results are summarized in Table 3. For PAS classification, PVDNet had good recognition ability, with an AUC of 0.973 (95% CI: [0.951, 0.992]), which was equivalent to the AUC of 0.943 (95% CI: [0.873, 0.997], p = 0.3076) for a senior radiologist with 20 years of experience in pulmonary vascular disease (SRPV), which was significantly better than the other three radiologists with different years of experience (JRDS, a junior resident with 3 years of experience: 0.741, 95% CI [0.644, 0.833], p < 0.0001; JRPV, a junior radiologist with 6 years of experience in pulmonary vascular disease: 0.793, 95% CI [0.696, 0.884], p < 0.0005; MRPV, a radiologist with 13 years of experience in pulmonary vascular disease: 0.872, 95% CI [0.790, 0.951], p < 0.05). In APE classification, the PVDNet model (0.835, 95% CI: [0.775, 0.893]) was superior to the 3-year-experienced junior resident JRDS (0.686, 95% CI: [0.613, 0.754], p<0.01) in terms of AUC. In CPE classification, the model's AUC was 0.809 (95% CI: [0.731, 0.884]), which was comparable to the 3-year-experienced junior resident JRDS's AUC of 0.735 (95% CI: [0.657, 0.808], p=0.2304). The ROC curves of the readers using the model of the present invention and the four manual readers are shown in Figure 2. Figure 5As shown. In the figure, APE is acute pulmonary thromboembolism, CPE is chronic pulmonary thromboembolism; PAS is pulmonary artery sarcoma; JRDS represents the recognition result of a junior resident with 3 years of experience; JRPV represents the recognition result of a junior radiologist with 6 years of experience in pulmonary vascular disease; MRPV represents the recognition result of a radiologist with 13 years of experience in pulmonary vascular disease; SRPV represents the recognition result of a senior radiologist with 20 years of experience in pulmonary vascular disease. As can be seen from the figure, the recognition result of the model of the present invention is significantly consistent with the recognition result of SRPV (kappa=0.651, p<0.0001), and SRPV has the highest level among the four manual (JRDS kappa=0.282, JRPV kappa=0.475, MRPVkappa=0.554, all p<0.0001), which also shows that the recognition accuracy of the model of the present invention is very good.

[0036]

[0037] Figure 6 The confusion matrix between the PVDNet model and the manual discrimination of three radiologists with different years of experience (JRDS represents a junior resident with 3 years of experience; JRPV represents a junior radiologist with 6 years of experience in pulmonary vascular disease; MRPV represents a senior radiologist with 13 years of experience in pulmonary vascular disease; SRPV represents a senior radiologist with 20 years of experience in treating pulmonary vascular disease) shows that PVDNet has the highest consistency with SPRV in diagnosing PAS and APE, but has low consistency with JRPV. The confusion matrix between PDVNet and four human readers in the external validation set. The kappa value of JRDS is 0.282; JRPV is 0.475; MRPV is 0.554; and SRPV is 0.651. APE is acute pulmonary thromboembolism, CPE is chronic pulmonary thromboembolism; PAS is pulmonary artery sarcoma.

[0038] In this paper, a three-dimensional deep learning model, PVDNet, was developed to discriminate PAS and PTE (considering APE and CPE) on CTPA. It can be found that in the internal test set, the AUC of PVDNet for PAS, APE and CPE were 0.972 (95% CI: [0.945, 0.994]), 0.902 (95% CI: [0.855, 0.944]) and 0.900 (95% CI: [0.852, 0.946]), respectively, and the results predicted by the model were significantly consistent with the results of clinical diagnosis. In the external validation set, PVDNet performed best in PAS discrimination, and its AUC was similar to that of a senior radiologist (SRPV) with 20 years of experience in pulmonary vascular disease (0.973 & 0.943, p = 0.3076), which was higher than that of another senior and two junior radiologists. There was a high degree of consistency in discrimination between PVDNet and SRPV, which was the highest among the four radiologists.

[0039] The design of the PVDNet model contains a max pooling operation for feature aggregation and a fully connected layer followed by a softmax activation function, which is specially tailored for fine-grained classification of lung diseases. PVDNet uses a pre-trained ResNet-50 architecture to extract instance features from CTPA images. ResNet-50 has the ability to handle very deep network structures, effectively alleviating the problem of gradient vanishing and improving the expressiveness and performance of the model. The use of residual connections can retain the original features, making the network learning smoother and more stable, thereby improving the accuracy and generalization ability of the model.

[0040] The present invention also uses gradient-weighted class activation mapping (Grad-CAM) to visualize the decision-making process of the model. These heat maps illustrate that the model pays attention to abnormalities in the image when making predictions, thereby improving the interpretability of the model. This feature enables radiologists to understand the basis of the model's predictions, thereby increasing confidence in its clinical applicability. In the internal test set, PVDNet had an AUC of (0.972, 0.902) for PAS, 0.902 for APE, and 0.900 for CPE, highlighting its strong recognition ability in these three pulmonary vascular diseases. The model has an unusually high accuracy in identifying PAS, which is critical for identifying this rare condition. For APE, its AUC is 0.902, and the model of the present invention can effectively distinguish it from other pulmonary vascular diseases. Although the AUC for CPE is slightly lower at 0.900, the model still shows strong differentiation ability. The Kappa statistic shows substantial consistency between the model's predictions and actual diagnoses, further supporting the reliability and clinical utility of the model. In external validation, PVDNet's performance was similar to that of a senior cardiovascular imaging expert with 20 years of experience in pulmonary vascular disease identification, and was superior to other physicians, including another senior and two junior radiologists in PAS classification, indicating that the model is capable of expert-level recognition, which is particularly important in situations where there is a shortage of specialist radiologists.

[0041] An advantage of the present invention is that all CTPA data are from multiple centers, enhancing the diversity of the samples and the robustness of the model. This broad-based approach is critical to ensuring the diagnostic ability of the model in different patient populations and imaging conditions. In summary, deep learning has the potential to revolutionize the diagnosis of PAS and PTE. The development of PVDNet represents an important step towards more accurate and accessible diagnostic tools in pulmonary vascular disease.

[0042] Parts of the present invention that are not described in detail may refer to prior art or are known to those skilled in the art.

[0043] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A deep learning-based intelligent recognition system for pulmonary artery sarcoma and pulmonary embolism, characterized in that: The system comprises: A data acquisition module, which obtains a three-dimensional volumetric CTPA image data set containing APE, CPE and PAS for use as a data set for a deep learning model training module, an internal testing module and an external verification module, and obtains CTPA image data for data to be identified in a model application module; The data preprocessing module preprocesses the data collected by the data acquisition module to obtain data that meets the use requirements of subsequent modules; The deep learning model training module is implemented using the supervised convolutional neural network PVDNet framework, and the PVDNet framework is trained with the preprocessed data to obtain a trained recognition model; An internal testing module is used to perform internal testing on the trained recognition model according to the corresponding internal testing data set; The external verification module is used to identify the trained recognition model according to the corresponding external verification data set, and then perform statistical analysis on the recognition results and the manual recognition results to verify the accuracy of the recognition model; The model application module recognizes the real data to be recognized through the recognition model and determines whether APE, CPE and / or PAS exist on the corresponding CTPA image.

2. The deep learning-based intelligent recognition system for pulmonary artery sarcoma and pulmonary embolism according to claim 1 is characterized in that: The data acquisition module acquires 952 three-dimensional volume CTPA image data sets including APE, CPE and PAS; 590 of them are used as training data sets for the deep learning model training module, 186 are used as test data sets for the internal testing module, and 176 are used as verification test sets for the external verification module.

3. The deep learning-based intelligent recognition system for pulmonary artery sarcoma and pulmonary embolism according to claim 2 is characterized in that: The inclusion criteria of the 952 3D volumetric CTPA image datasets included: the filling defect on CTPA was central, APE was confirmed by follow-up after anticoagulation or thrombolytic therapy, CPE was confirmed by follow-up and endarterectomy or balloon pulmonary angioplasty, and PAS was confirmed by biopsy or surgical pathology.

4. The deep learning-based intelligent recognition system for pulmonary artery sarcoma and pulmonary embolism according to claim 1 is characterized in that: The processing content of the data preprocessing module includes: segmenting each three-dimensional volume CTPA image, then extracting the lung area from each segmented image through the three-dimensional bounding box extracted by the corresponding segmentation mask, retaining the pulmonary artery and the heart; then adjusting the size of the lung area and then normalizing it.

5. The deep learning-based intelligent identification system for pulmonary artery sarcoma and pulmonary embolism according to claim 4 is characterized in that: The lung region was resized to 192×192×192, the window center was 100, and the window width was 700. A window operation was performed, and the resized lung region was subjected to minimum-maximum normalization.

6. The deep learning-based intelligent identification system for pulmonary artery sarcoma and pulmonary embolism according to claim 1, characterized in that: The supervised convolutional neural network PVDNet framework includes a pre-trained ResNet-50, a maximum pooling module, and a fully connected layer with a softmax activation function; The ResNet-50 contains 3.8 billion trainable parameters to extract instance features; The maximum pooling module uses a maximum pooling operation to perform feature aggregation; The fully connected layer generates probability scores for classification through a softmax activation function.

7. The deep learning-based intelligent identification system for pulmonary artery sarcoma and pulmonary embolism according to claim 6 is characterized in that: The deep learning model training module receives the 3D volume data processed by the data preprocessing module and segments it along the z-axis to obtain a number of 2D slices to form a slice bag, and then inputs the slice bag into the supervised convolutional neural network PVDNet framework for model training.

8. The deep learning-based intelligent identification system for pulmonary artery sarcoma and pulmonary embolism according to claim 1, characterized in that: The deep learning model training module uses a five-fold cross-validation method to perform model training; in each fold, 80% of the training data is used as the development set for model training, and the remaining 20% ​​is used as the tuning set for model selection; the model that obtains the best results on the tuning set is used for subsequent internal test sets and external validation.

9. The deep learning-based intelligent identification system for pulmonary artery sarcoma and pulmonary embolism according to claim 8, characterized in that: The deep learning model training module uses an optimizer to train the entire network for 50 epochs, with an initial learning rate of 1e-5 and a mini-batch size of 1; instance balanced sampling is used to mitigate the impact of class imbalance, and class balanced loss is selected as the loss function to train the model; a simplified lon platform method based on adjusted set loss is used to adjust the learning rate.

10. The deep learning-based intelligent identification system for pulmonary artery sarcoma and pulmonary embolism according to claim 1, characterized in that: The system also includes a visualization module that uses gradient-weighted class activation mapping to visualize the decision-making process of the model.

Images