A method, apparatus, medium, and device for embolus detection in pulmonary artery imaging

Abstract

The application discloses a kind of embolus detection method, device, medium and equipment in pulmonary artery imaging, it is related to pulmonary artery imaging detection technical field.The history pulmonary artery imaging of pulmonary embolism patient is acquired as sample image and is determined mark first, then with Res2Net network as its backbone module in encoder, and after the last layer of decoder of U-Net network model jump connection, increase multi-hole spatial pyramid pooling module and attention mechanism module, build improved U-Net network model and train to obtain embolus detection model.Last, the pulmonary artery imaging to be detected is input into embolus detection model and embolus is detected.The application extracts the embolus feature under different scales in pulmonary artery imaging at fine granularity level, obtains the multi-scale information of embolus context information, and is more focused on high contribution feature channel, to realize the segmentation of embolus relatively accurate, improve the accuracy and precision of embolus detection in pulmonary artery imaging.

Description

Technical Field

[0001] This application relates to the field of pulmonary artery imaging detection technology, and in particular to a method, apparatus, medium and device for embolus detection in pulmonary artery imaging. Background Technology

[0002] Acute pulmonary embolism (APE) is a serious and potentially life-threatening disease, ranking as the third leading cause of death after malignant tumors and myocardial infarction. According to reports, the mortality rate of APE is 20%–30%, but with timely diagnosis, early detection, and treatment, the mortality rate can be reduced to below 8%. However, while common symptoms of APE include dyspnea, chest pain, cough, syncope, and hemoptysis, the triad of dyspnea, chest pain, and hemoptysis occurs in less than 30% of cases. Due to its low specificity, clinical diagnosis is challenging. Many doctors, limited by their experience, often fail to diagnose APE immediately, resulting in a persistently high rate of missed diagnoses.

[0003] Currently, the effective method for diagnosing acute pulmonary embolism is computer-to-pulmonary angiography (CTPA). Typically, a single patient will have 300 to 500 CTPA images scanned, each of which needs to be annotated by a doctor. This method consumes too much manpower and time, which is not conducive to the rapid diagnosis and treatment of acute pulmonary embolism. Therefore, if computer segmentation of emboli in the pulmonary artery can be used, it will effectively reduce the workload of doctors and thus facilitate the early diagnosis of acute pulmonary embolism.

[0004] Despite the significant advancements in computer vision and medical imaging in recent years, the segmentation of acute pulmonary embolism (APE) remains a major challenge. Over the past decade, numerous researchers have employed traditional segmentation methods such as region-based, boundary-based, and contour-based approaches to segment APE. However, the rapid development of deep learning technology in recent years has enabled substantial progress in addressing many problems that traditional medical image processing methods struggle with. Medical image segmentation methods based on convolutional neural networks have garnered widespread attention due to their self-learning capabilities, automatic feature extraction, and broad applicability, offering new insights for APE segmentation.

[0005] Huang et al. proposed PENet, a 77-layer network structure for end-to-end detection of acute pulmonary embolism (APE) images. Trained on an independently collected dataset and validated on another dataset, it achieved a detection accuracy of 0.84, making an innovative contribution to the automatic detection of APE. Wen Z et al. proposed a network based on a modified U-Net for effective semantic segmentation of APE CTPA images. In their work, they used partial weights of a pre-trained VGG16 model to initialize the parameters of the U-Net network's contraction path, significantly reducing training time and improving the network's generalization ability. This research also redefined the loss function, combining Focal Loss and cross-entropy loss to address the uneven quality of experimental datasets. Ultimately, automatic segmentation of APE CTPA images was achieved.

[0006] However, due to the large individual variability in CTPA images of acute pulmonary embolism, the segmentation results vary significantly, resulting in low sensitivity and high false positive rate of current convolutional neural network-based methods, leading to low accuracy in embolism detection in current pulmonary artery imaging. Summary of the Invention

[0007] Therefore, it is necessary to provide a method, apparatus, medium, and device for embolism detection in pulmonary artery imaging to address the aforementioned technical problems.

[0008] The following technical solution is adopted in this specification:

[0009] This specification provides a method for embolus detection in pulmonary artery imaging, including:

[0010] Historical pulmonary artery imaging of patients with acute pulmonary embolism was obtained as sample images, and the actual location of the embolus in each sample image was labeled.

[0011] Based on the U-Net network model, Res2Net network is used as the backbone module in its encoder, and a porous spatial pyramid pooling module and an attention mechanism module are added after the last layer skip connection of the U-Net network model decoder to construct an improved U-Net network model.

[0012] The sample images are input into the improved U-Net network model to obtain the predicted embolus location; and the improved U-Net network model is trained with the optimization objective of minimizing the deviation between the actual embolus location and the predicted embolus location in the sample images to obtain the embolus detection model.

[0013] The pulmonary artery image to be detected is input into the embolism detection model. The encoder extracts embolism features at different scales of the pulmonary artery image to obtain multi-scale feature maps of emboli of different sizes. The decoder fuses the multi-scale feature maps of different sizes to obtain a fine feature map of the emboli. The porous spatial pyramid pooling module extracts features from the fine feature map at different scales and performs linear fusion of each feature channel to obtain an enhanced feature map of the emboli. The attention mechanism module assigns weights according to the contribution of different feature channels to embolism segmentation and weights each feature channel of the enhanced feature map to obtain a weighted feature map. The output layer performs embolism detection on the weighted feature map to obtain the embolism location.

[0014] Optionally, acquiring historical pulmonary artery imaging data of patients with acute pulmonary embolism as sample images specifically includes:

[0015] Acquire historical CT pulmonary artery images of patients with acute pulmonary embolism, perform quality checks on the historical CT pulmonary artery images, and use historical CT pulmonary artery images that meet the quality requirements as sample images.

[0016] In cases where the clinical and imaging data of patients with acute pulmonary embolism are incomplete, or the historical pulmonary artery imaging is not a 1 mm thin-slice image, or the historical pulmonary artery imaging contains artifacts, or the embolic signs in the historical pulmonary artery imaging cannot be marked, the corresponding historical pulmonary artery imaging quality is determined to be non-compliant.

[0017] Optionally, the method further includes:

[0018] Convert the sample image to HU value, and set the window level of the sample image to 100HU and the window width to 900HU;

[0019] The adjusted sample images are centered and standardized to obtain sample images that follow a standard normal distribution, and a mask is set in the lungs;

[0020] The sample image is enhanced through various operations such as rotation, horizontal mirror flip, vertical mirror flip, and sharpening.

[0021] Optionally, the step of extracting embolus features at different scales of pulmonary artery imaging to be detected through an encoder, and obtaining multi-scale feature maps of emboli of different sizes, specifically includes:

[0022] By decomposing convolution operations at different scales within a single residual block using Res2Net networks of different sizes in the encoder, emboli features at different scales of the pulmonary artery imaging to be detected are extracted at a fine-grained level. Feature combinations of different receptive field sizes are obtained and fused to acquire multi-scale feature maps of emboli at different sizes.

[0023] Optionally, the step of extracting features from the fine feature map at different scales using a porous spatial pyramid pooling module and linearly fusing each feature channel to obtain the enhanced feature map of the thrombus specifically includes:

[0024] The fine feature map is convolved by a porous spatial pyramid pooling module, and dilated convolutions with different dilated convolutional layer sampling rates corresponding to dilated rates are performed in multiple separate parallel branches.

[0025] Global average pooling is performed on the convolution results to obtain multi-scale information of the pulmonary artery imaging emboli to be detected, and linear fusion is performed on each feature channel to obtain the enhanced feature map of the emboli.

[0026] Optionally, the step of assigning weights to the contribution of different feature channels to embolus segmentation through the attention mechanism module, and obtaining a weighted feature map by weighting each feature channel of the enhanced feature map, specifically includes:

[0027] The attention mechanism module compresses the enhanced feature map along the spatial dimension, reducing the dimensionality of the feature map for each feature channel;

[0028] The feature maps of each feature channel after dimensionality reduction and compression are assigned different weights according to the contribution of different feature channels to thrombus segmentation. The weights corresponding to each feature channel are multiplied with the enhanced feature map channel by channel to obtain the weighted feature map.

[0029] This specification provides an embolus detection device for pulmonary artery imaging, comprising:

[0030] The acquisition module is used to acquire historical pulmonary artery images of patients with acute pulmonary embolism as sample images and determine the annotation of the actual embolus location in each sample image;

[0031] The building blocks are used to construct an improved U-Net network model based on the U-Net network model, using the Res2Net network as the backbone module in its encoder, and adding a porous spatial pyramid pooling module and an attention mechanism module after the last layer skip connection of the U-Net network model's decoder.

[0032] The training module is used to input sample images into the improved U-Net network model to obtain the predicted embolus location; and to train the improved U-Net network model with the optimization objective of minimizing the deviation between the actual embolus location and the predicted embolus location in the sample image, so as to obtain the embolus detection model.

[0033] The detection module inputs the pulmonary artery image to be detected into the embolism detection model. The encoder extracts embolism features at different scales from the pulmonary artery image, obtaining multi-scale feature maps of emboli at different sizes. The decoder fuses these multi-scale feature maps to obtain a fine-grained feature map of the emboli. A porous spatial pyramid pooling module extracts features from the fine-grained feature map at different scales and linearly fuses each feature channel to obtain an enhanced feature map of the emboli. An attention mechanism module assigns weights based on the contribution of different feature channels to embolism segmentation and weights each feature channel of the enhanced feature map to obtain a weighted feature map. The output layer performs embolism detection on the weighted feature map to determine the emboli location.

[0034] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the embolus detection method in pulmonary artery imaging described above.

[0035] This specification provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the embolus detection method in pulmonary artery imaging described above.

[0036] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects:

[0037] First, historical pulmonary artery images of patients with pulmonary embolism were acquired as sample images and labeled. Then, based on the U-Net network model, a Res2Net network was used as the backbone module in its encoder. A porous spatial pyramid pooling module and an attention mechanism module were added after the final skip connections of the U-Net network decoder to construct an improved U-Net network model, which was then trained to obtain the embolism detection model. Finally, the pulmonary artery images to be detected were input into the embolism detection model for embolism detection.

[0038] This invention utilizes the multi-scale enhancement capability of the Res2Net network in the trained embolism detection model to extract embolism features at different scales in pulmonary artery imaging at a fine-grained level. It leverages the multi-scale capture capability of the porous spatial pyramid pooling module to obtain multi-scale information of pulmonary artery imaging context. Furthermore, it utilizes the focusing capability of the attention mechanism module to enable the embolism detection model to focus more on the feature channels that contribute more when detecting and segmenting emboli, thereby reducing the impact of individual differences on the detection results and achieving more accurate segmentation of emboli. This improves the accuracy and precision of embolism detection in pulmonary artery imaging. Attached Figure Description

[0039] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0040] Figure 1 This is a schematic diagram of a method for embolus detection in pulmonary artery imaging provided in this specification.

[0041] Figure 2 This is a schematic diagram illustrating the acquisition and data filtering of historical pulmonary artery imaging data provided in this specification.

[0042] Figure 3 This is a schematic diagram of the structure of a U-Net network model provided in this specification;

[0043] Figure 4 This document provides a schematic diagram of a data convolution process.

[0044] Figure 5 This document provides a schematic diagram of a residual network principle.

[0045] Figure 6 This document provides a comparative diagram of ResNet and Res2Net network modules.

[0046] Figure 7 This is a schematic diagram of a stage module provided in this specification;

[0047] Figure 8 This is a schematic diagram of an ASPP module connection provided in this manual;

[0048] Figure 9 This is a schematic diagram of an SE module connection provided in this specification;

[0049] Figure 10 This is a schematic diagram of an image preprocessing workflow provided in this specification;

[0050] Figure 11 This document provides a schematic diagram of an improved U-Net network model structure.

[0051] Figure 12a This is a schematic diagram of the segmentation index curve of an embolus detection model provided in this specification;

[0052] Figure 12b This specification provides a schematic diagram of a CE-Net network segmentation index curve.

[0053] Figure 12c This document provides a schematic diagram of a U-Net network segmentation index curve.

[0054] Figure 12d This is a schematic diagram of the training index curve of an embolus detection model provided in this specification;

[0055] Figure 12e This document provides a schematic diagram of a CE-Net network training metric curve.

[0056] Figure 12f This document provides a schematic diagram of a U-Net network training metric curve.

[0057] Figure 13 This is a radar chart illustrating the segmentation results of three types of networks provided in this specification.

[0058] Figure 14 This is a schematic diagram of a model segmentation result provided in this specification;

[0059] Figure 15 This is a schematic diagram of an embolus detection device in pulmonary artery imaging provided in this specification.

[0060] Figure 16 This is a schematic diagram of a computer device for implementing an embolus detection method in pulmonary artery imaging, as provided in this specification. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0062] Acute pulmonary embolism (APE) refers to the blockage of the pulmonary artery by a blood clot, usually causing severe symptoms within a short period of time. APE has a rapid onset and high mortality and disability rates, ranking third among cardiovascular causes of death. Because of its insidious onset and nonspecific symptoms, it is often overlooked. Without timely treatment, the mortality rate can reach 25%–35%, while timely and effective treatment can reduce the mortality rate to below 10%. CT pulmonary angiography is the "gold standard" for diagnosing APE, but it has some drawbacks. Typically, a single patient requires 300–500 CT PA images, resulting in a heavy workload for radiologists and a time-consuming and labor-intensive diagnostic process. Furthermore, due to fatigue from prolonged image reading and differences in experience among physicians, missed diagnoses and misdiagnoses are common, hindering timely diagnosis and early treatment of APE. Computer-aided diagnosis (CAD) systems can assist in the rapid diagnosis of APE, but their low sensitivity and high false positive rate have limited their widespread clinical application. Deep learning networks are an important method in the field of artificial intelligence and are increasingly being applied to the auxiliary diagnosis and assessment of acute pulmonary embolism (APE). This invention proposes to construct an intelligent APE detection model based on deep learning networks to improve the accuracy of APE diagnosis and the efficiency of physician diagnosis.

[0063] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0064] Figure 1 This is a schematic diagram of a method for embolus detection in pulmonary artery imaging as described in this specification, which specifically includes the following steps:

[0065] S101: Obtain historical pulmonary artery imaging data of patients with acute pulmonary embolism as sample images, and label the actual embolus location in each sample image.

[0066] Generally, the detection of pulmonary embolism involves detecting the presence of emboli (insoluble blood vessels) in the pulmonary artery, which obstruct its lumen. Therefore, the presence of emboli in the pulmonary artery can be used to determine the presence of pulmonary embolism. Based on this, neural network models can be used to detect emboli in pulmonary artery imaging.

[0067] Therefore, in one or more embodiments of this specification, the server may first acquire historical pulmonary artery imaging data of patients with acute pulmonary embolism to create sample images for model training. Here, the pulmonary artery imaging may refer to CTPA images.

[0068] Specifically, the business platform's server can directly acquire or indirectly acquire CTPA images of patients who have undergone CTPA examinations and have been clinically diagnosed with acute pulmonary embolism from third-party data platforms. Then, the acquired images can be quality checked, and historical pulmonary artery imaging (CTPA images) that meet the quality standards can be used as sample images.

[0069] CTPA images can be stored in the image archiving and communication system (PACS) of the business platform or third-party data platform. The following example can be used as a reference when the business platform or third-party data platform acquires CTPA images.

[0070] Scanning equipment: Siemens dual-source 64-slice spiral CT (Somatom Definition 60320, SIEMENS, Germany).

[0071] Scanning Procedure: The patient was placed in a supine position with both arms raised. Iohexol at a concentration of 350 mg I / mL was used as the contrast agent, injected via the antecubital vein at a flow rate of 5.0 mL / s using a dual-barrel high-pressure injector, for a total injection of 50 mL. This was followed by a flushing with 30 mL of normal saline at the same flow rate to reduce contrast agent residue. Intelligent tracking automatic trigger scanning technology was used, with the trigger point set at the maximum level of the main pulmonary artery trunk, a threshold of 100 HU, and a delay time of 2 seconds. The scanning range extended from the thoracic inlet to 1.0 cm below the costophrenic angles on both sides. After the pulmonary artery phase scan, images of the pulmonary venous phase were immediately acquired from the feet towards the head.

[0072] Scanning parameters: tube voltage 120kV, automatic milliampere control technology, detector width 0.625mm×64. Standard reconstruction algorithm, 1mm slice thickness, 512×512 matrix. Mediastinal window width 465HU, window level 65HU. Lung window width 1515HU, window level -480HU.

[0073] The quality control measures mentioned can be: (1) confirming whether all patients underwent CTPA examination, with the diagnostic criteria for PE referring to the 2019 European Society of Cardiology and European Respiratory Society guidelines for the diagnosis and management of acute pulmonary embolism. (2) confirming the completeness of all patients' clinical and imaging data. (3) confirming whether the acquired historical pulmonary artery images were 1.0 mm thin-slice images. (4) confirming the presence of obvious artifacts in the historical pulmonary artery images that would prevent embolism labeling. (5) confirming the presence of obvious embolic signs of acute pulmonary embolism in the historical pulmonary artery images, enabling embolism labeling.

[0074] During the quality inspection process, the following situations are considered unqualified in terms of historical pulmonary artery imaging quality: CTPA examination and PE diagnostic criteria not referring to the guidelines for the diagnosis and management of acute pulmonary embolism; incomplete clinical and imaging data; historical pulmonary artery images not being 1.0mm thin-slice images; historical pulmonary artery images containing obvious artifacts; and historical pulmonary artery images showing unclear embolic signs that cannot be labeled. Conversely, the following situations are considered qualified in terms of historical pulmonary artery imaging quality.

[0075] After acquiring sample images, the server can determine the annotation of the actual emboli location within the images. For example, in response to user input, the server can annotate lesions on patient CTPA images in digital imaging and communications in medicine (DICOM) format using ITK-SNAP image segmentation software. First, a junior physician (e.g., with 2 years of experience in CT pulmonary artery imaging diagnosis) can manually delineate the emboli ROIs for all PE patients under a single-blind approach, paying attention to coronal, sagittal, and axial planes to ensure accurate segmentation. After annotation, a senior physician (with 12 years of experience in CT pulmonary artery imaging diagnosis) can review all annotated images. For difficult or uncertain cases, the results can be determined through discussion. The final annotated dataset reviewed by the senior physician serves as the gold standard for image annotation.

[0076] Figure 2 This is a schematic diagram illustrating the acquisition and data filtering of historical pulmonary artery imaging data as described in this specification. Figure 2 As can be seen, after retrospectively including CTPA images and going through layers of image removal, 303 CTPA images were left, which were then divided into training set, validation set and test set.

[0077] The server mentioned in this manual can be a server set up on a business platform, or a device such as a desktop computer or laptop computer capable of executing the solution described in this manual. For ease of explanation, the following description will only focus on the server as the execution subject.

[0078] S102: Based on the U-Net network model, using the Res2Net network as the backbone module in its encoder, and adding a porous spatial pyramid pooling module and an attention mechanism module after the last layer skip connection of the U-Net network model's decoder, an improved U-Net network model is constructed.

[0079] S103: Input the sample image into the improved U-Net network model to obtain the predicted embolus location; and train the improved U-Net network model with the optimization objective of minimizing the deviation between the actual embolus location and the predicted embolus location in the sample image to obtain the embolus detection model.

[0080] S104: The pulmonary artery image to be detected is input into the embolism detection model. The encoder extracts embolism features at different scales of the pulmonary artery image to obtain multi-scale feature maps of emboli of different sizes. The decoder fuses the multi-scale feature maps of different sizes to obtain a fine feature map of the emboli. The porous spatial pyramid pooling module extracts features from the fine feature map at different scales and performs linear fusion of each feature channel to obtain an enhanced feature map of the emboli. The attention mechanism module assigns weights according to the contribution of different feature channels to embolism segmentation and weights each feature channel of the enhanced feature map to obtain a weighted feature map. The output layer performs embolism detection on the weighted feature map to obtain the embolism location.

[0081] After obtaining the sample images and labeling the emboli as described above, an improved U-Net network model can be constructed. The improved U-Net network model can be trained based on the sample images to obtain an emboli detection model. The trained emboli detection model can then be used to detect emboli.

[0082] For the construction of the improved U-Net network model, medical images often involve patient privacy and are difficult to obtain, resulting in relatively limited data. The U-Net network model employs rich data augmentation strategies to expand the dataset, such as random rotation, flipping, and scaling, which improves the model's generalization ability and robustness. On the other hand, specifically for pulmonary artery imaging, the semantic information is relatively simple, with an information density far lower than that of everyday photographs. Its application is highly targeted, focusing solely on imaging the pulmonary arteries. Therefore, when detecting emboli, both high-level semantic information and low-level features are crucial. To address this, the U-Net network model's U-shaped structure contains skip connections, effectively combining low-resolution information (providing a basis for object category identification) and high-resolution information (providing a basis for accurate segmentation and localization).

[0083] Therefore, this invention constructs an improved U-Net network model based on the U-Net network model. The structure of the U-Net network model is as follows: Figure 3 As shown. By Figure 3 As can be seen, the U-Net network model has a U-shaped structure. The left side of the U-shaped network is the encoder structure, which includes convolution and downsampling processes. The right side of the U-shaped network is the decoder structure, which includes deconvolution and upsampling processes. The horizontal arrow in the middle represents the skip connections, which include copy and pruning processes.

[0084] By employing skip connections in the U-Net network model, pulmonary artery images downsampled at the same layer can be cropped to the same size and then added as feature layers during the decoding process. Simultaneously, the encoding process, after multiple downsampling steps, yields deep information corresponding to low-resolution data. This provides contextual semantic information about emboli within the entire pulmonary artery image pixels, aiding in emboli classification. Meanwhile, the shallow information corresponding to high-resolution data passed from the encoder to the decoder via skip connections provides more refined features for emboli location segmentation. Combining deep and shallow information improves the accuracy of emboli segmentation in pulmonary artery imaging.

[0085] Furthermore, in one or more embodiments of this specification, the encoder in the U-Net network model, which includes simple convolution and downsampling, can be replaced with a Res2Net network (e.g., Res2Net-34).

[0086] Res2Net is an improvement on ResNet, which was proposed in 2015 by Kaiming He and others from Microsoft Research. Before ResNet, all neural networks consisted of convolutional and pooling layers. The prevailing view at the time was that deeper networks performed better; however, experiments showed that as network depth increased, gradient vanishing or exploding phenomena and degradation problems occurred.

[0087] The ResNet network addresses these issues by incorporating Batch Normalization (BN) layers to resolve vanishing and exploding gradient phenomena. Image normalization is typically performed during image preprocessing to accelerate network convergence. Figure 4 As shown, Figure 4 This is a schematic diagram of a data convolution process in this specification. For the Conv1 layer, the input is a feature matrix that follows a certain distribution, but for the Conv2 layer, the input feature matrix does not necessarily follow a certain distribution. Therefore, a Batch Normalization (BN) layer is added to process the data of each channel to make it follow a standard Gaussian distribution with a mean of 0 and a variance of 1. This reduces the dependency between layers, thereby avoiding the occurrence of gradient vanishing and gradient exploding phenomena.

[0088] To address the degradation problem, residual networks are proposed as a solution. The principle of residual networks is as follows: Figure 5As shown. For traditional network structures, the input is X, and the features learned by the network are H(X). The residual structure used in ResNet does not directly learn the features H(X), but rather learns the residual F(X) between the features and the input. After obtaining F(x), the feature H(X) = F(X) + X can be calculated. The reason why the residual module can solve the degradation problem is that when the residual learned by the network is 0, this module is equivalent to performing an identity mapping, which will not lead to a decrease in network performance. Moreover, in actual network training, the residual is usually not 0.

[0089] The Res2Net network and the ResNet network have the same overall structure. Figure 6 This is a schematic diagram comparing a ResNet network module and a Res2Net network module in this specification. Figure 6 The left side shows a schematic of a ResNet network module, and the right side shows a schematic of a Res2Net network module. The Res2Net network can be obtained by transforming each module in the ResNet network.

[0090] Depend on Figure 6 As can be seen, the Res2Net network module includes 4 lines. The first line x1 is transmitted directly to y1 without any processing. The second line x2 is split into two lines after a 3×3 convolution. One line continues forward to y2, and the other line is transmitted to the third line. In this way, the third line obtains the information of the second line. The third and fourth lines can be processed in the same way.

[0091] As can be seen, based on this split-hybrid connection structure, the Res2Net network module constructs hierarchical residual connections within a single residual block, representing multi-scale features at a fine-grained level and increasing the receptive field of each network layer. That is, the output of the Res2Net network module contains combinations of different receptive field sizes. This structure is more advantageous for extracting global and local information from pulmonary artery imaging when detecting and segmenting emboli. Therefore, this invention replaces the encoder in the U-Net network model with a Res2Net network, using the Res2Net network to extract emboli features at different scales of pulmonary artery imaging at a fine-grained level.

[0092] Therefore, in one or more embodiments of this specification, a Res2Net network can be used as the backbone module in the encoder of the U-Net network model. The image is input into the Res2Net network corresponding to each backbone module, and image features are extracted at different scales. The Res2Net network increases its awareness of multi-scale features by decomposing the convolution operations at each scale. Features extracted at different scales are fused to obtain multi-scale features, enabling better feature fusion. The fused features are then used for classification or emboli detection tasks in pulmonary artery imaging. The Res2Net network solves the problems of small pulmonary emboli and indistinct features, enhances the overall model's ability to extract detailed information from small targets, and improves detection accuracy.

[0093] Figure 7 This is a schematic diagram of one type of stage module in this specification, namely... Figure 3 This diagram illustrates the stage module in the U-Net network model after replacing it with a Res2Net network. Figure 7 As can be seen, in the stage module, stride=2. After passing through the stage module, the size of the pulmonary artery imaging will be reduced, so the branches will also be downsampled to make them the same size as the main path. The number of input and output channels in the stage module is generally different, while in the normal module, stride=1, the branches do not undergo downsampling, and the number of input and output channels is the same.

[0094] Furthermore, in one or more embodiments of this specification, an Atrous Spatial Pyramid Pooling (ASPP) module and an Attention Excitation (SE) module may be added after the final skip connections of the decoder in the U-Net network model, such as... Figure 3 As shown.

[0095] Figure 8 This is a schematic diagram of the connection of an ASPP module in this specification. Through the ASPP module, dilated convolution with different dilation rates can be used to perform convolution and global average pooling operations on the input feature map. After the results are fused, the description of the multi-scale information corresponding to the pulmonary artery imaging context information is obtained.

[0096] The porous spatial pyramid pooling module uses multiple parallel dilated convolutional layers with different sampling rates, and the features extracted for each sampling rate are further processed in a separate branch.

[0097] Spatial Pyramid Pooling with Pores (ASPP) is derived from Spatial Pyramid Pooling (SPP) and includes four porous convolutions with different dilation rates, which can effectively capture multi-scale information. ASPP consists of one 1×1 convolution, three 3×3 convolutions (with dilation rates of 6, 12, and 18, respectively), and a global average pooling module.

[0098] Figure 9 This is a schematic diagram of an SE module connection in this specification. The SE module is added after the ASPP module. Through the SE module, different weights can be assigned according to the contribution of different feature channels of the sample image to embolus segmentation at the channel level. After training the model, more important feature information in pulmonary artery imaging can be obtained based on the SE module.

[0099] In the SE module, the first step is to compress the feature map along the spatial dimension using the "Squeeze" operation, reducing the dimensionality of the feature map for each channel to a real number. Since the dimensionality reduction method used is global average pooling, the real number for each channel represents the feature of the entire feature map in that channel, and the receptive field is the size of the entire feature map.

[0100] Then, the "Excitation" operation is used to activate the features. It learns the network parameters to assign different weights to each compressed feature. The weights learned in "Excitation" are multiplied with the input features channel by channel, thus assigning the importance of the learned channels to the initial features and completing the recalibration of channel importance.

[0101] The attention mechanism module adaptively adjusts the channel weights, enhancing the representational power of important feature channels and suppressing unimportant feature channels, thereby determining the contribution of different feature channels in the sample image to thrombus segmentation. Based on the contribution of different feature channels in the sample image to thrombus segmentation, the weights of each channel corresponding to different image locations in the channel domain are determined.

[0102] After constructing the improved U-Net network model, the server can input sample images into the improved U-Net network model to obtain predicted emboli locations. Then, the server can train the improved U-Net network model with the optimization objective of minimizing the deviation between the actual and predicted emboli locations in the sample images, thus obtaining an emboli detection model. Finally, the server can use the trained emboli detection model to detect emboli in the pulmonary artery images to be tested.

[0103] First, the encoder extracts emboli features from pulmonary artery imaging at different scales. The encoder primarily reduces the size of the feature maps and extracts high-level semantic information. During this process, the Res2Net network performs convolution and pooling operations on the input image to extract feature representations and obtain multi-scale feature maps of emboli at different sizes. With each layer in the encoder network, the image's width and height are halved, and the number of channels doubles. Specifically, the Res2Net networks in the encoder decompose convolution operations at different scales within a single residual block, extracting emboli features from pulmonary artery imaging at different scales at a fine-grained level. This results in feature combinations with different receptive field sizes, which are then fused to obtain multi-scale feature maps of emboli at different sizes.

[0104] Following the encoder, there is a bottleneck layer used to further compress the size of the feature map and increase the representational power of the features. This process is achieved through operations such as convolution and pooling.

[0105] Then, the feature map is received from the bottleneck layer and its size is gradually restored through upsampling. The role of the decoder is to restore the resolution and fuse low-level and high-level features to produce a more refined segmentation result and obtain a fine feature map of the plug.

[0106] Then, by using the porous spatial pyramid pooling module to convolve the fine feature map with dilated convolutions of different dilated convolutional layer sampling rates in multiple separate parallel branches, the contextual information of emboli in different receptive fields can be captured.

[0107] The convolution results are then subjected to global average pooling to obtain multi-scale information of the pulmonary artery imaging emboli context information to be detected. Then, 1×1 convolution is used to linearly fuse each channel to generate a high-resolution, semantically strong enhanced feature map that can fully characterize small targets. This helps the model better understand the semantics and structure of the image and improves segmentation performance.

[0108] After that, the attention mechanism module (Squeeze-and-Excitation block, SE block) assigns weights according to the contribution of different feature channels to embolus segmentation. The attention mechanism module can learn the relationship between channels through the training process and adaptively adjust the weight of each channel to enhance the representation ability of important feature channels and suppress unimportant feature channels.

[0109] Specifically, the attention mechanism module compresses the feature map along the spatial dimension of the enhanced feature map, reducing the dimensionality of the feature map for each feature channel.

[0110] The feature maps of each feature channel after dimensionality reduction and compression are assigned different weights according to the contribution of different feature channels to thrombus segmentation. The weights corresponding to each feature channel are multiplied with the enhanced feature map channel by channel. In this way, the importance of the learned channels is assigned to the input features, and a weighted feature map is obtained, thus completing the recalibration of the importance of the channels.

[0111] Output: The last layer of the network uses convolutional operations to transform the feature map into a predicted segmentation result. The output size is the same as the input image size, and each pixel represents the segmentation result of which category the corresponding location belongs to, thus obtaining the location of the thrombus.

[0112] based on Figure 1 The embolism detection method in pulmonary artery imaging, as shown, first acquires historical pulmonary artery images of patients with pulmonary embolism as sample images and determines their annotations. Then, based on the U-Net network model, using a Res2Net network as the backbone module in its encoder, and adding a porous spatial pyramid pooling module and an attention mechanism module after the final skip connections of the U-Net network model's decoder, an improved U-Net network model is constructed and trained to obtain the embolism detection model. Finally, the pulmonary artery images to be detected are input into the embolism detection model for embolism detection.

[0113] This invention utilizes the multi-scale enhancement capability of the Res2Net network in the trained embolism detection model to extract embolism features at different scales in pulmonary artery imaging at a fine-grained level. It leverages the multi-scale capture capability of the porous spatial pyramid pooling module to obtain multi-scale information of pulmonary artery imaging context. Furthermore, it utilizes the focusing capability of the attention mechanism module to enable the embolism detection model to focus more on the feature channels that contribute more when detecting and segmenting emboli, thereby reducing the impact of individual differences on the detection results and achieving more accurate segmentation of emboli. This improves the accuracy and precision of embolism detection in pulmonary artery imaging.

[0114] When applying the embolus detection method in pulmonary artery imaging provided in this manual, it is not necessary to follow the instructions. Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.

[0115] Furthermore, in one or more embodiments of this specification, before inputting the sample image into the improved U-Net network model for embolus detection, the server may preprocess the sample image, such as... Figure 10 As shown, Figure 10 This is a schematic diagram of an image preprocessing workflow in this specification.

[0116] The server first converts the sample images to HU values, setting the window level to 100HU and the window width to 900HU. Then, it centers and standardizes the adjusted images to obtain samples that follow a standard normal distribution, and applies a mask to the lungs. Finally, it enhances the images through rotation, horizontal mirroring, vertical mirroring, and sharpening. Training the embolus detection model with these preprocessed images accelerates convergence, and the masking of the lungs enhances the model's focus on lesion areas. Furthermore, the data augmentation allows for further expansion of the dataset, improving the generalization ability of the improved U-Net network model and the trained embolus detection model.

[0117] In addition, this manual also provides application examples. Experimental environment: The method in this example is programmed using Python 3.8, runs on Ubuntu 20.04 system, with an Intel Core i7 processor, uses the deep learning framework PyTorch 1.10 based on CUDA 11.4, and achieves GPU parallel acceleration on an Nvidia RTX 1080ti graphics card.

[0118] Statistical analysis: Data were analyzed using SPSS 26.0 (IBM SPSS Statistics, Version 26.0). Quantitative data were expressed as mean ± standard deviation (x ± s) or median and interquartile range, and analyzed using one-way ANOVA. Categorical data were expressed as percentages, and comparisons were performed using the χ² test. A p-value < 0.05 was considered statistically significant.

[0119] Clinical data comparison: This example included 303 patients diagnosed with APE, 178 males and 125 females, with an age of 52.68±18.73 years. The training set consisted of 212 patients, the validation set of 61 patients, and the test set of 30 patients. There were no statistically significant differences among the three groups in age, sex, comorbidities (hypertension, diabetes, coronary heart disease, pulmonary hypertension, deep vein thrombosis of the lower extremities, recent surgical history, and elevated D-dimer levels (P>0.05), as shown in Table 1.

[0120] Table 1 Comparison of clinical data of 303 patients

[0121] Age (years) 68±18 69±20 73±21.75 0.605 0.224 male 117(55.2) 38(62.3) 15(50.0) female 95(44.8) 23(37.7) 15(50.0) hypertension 84(39.6) 28(45.9) 14(46.7) 1.123 0.570 diabetes 44(20.8) 12(19.7) 8(26.7) 0.648 0.723 Coronary heart disease 69(32.5) 22(36.1) 16(53.3) 4.990 0.083 Pulmonary hypertension 38(17.9) 7(11.5) 3(10.0) 2.330 0.312 Deep vein thrombosis in the lower extremities 90(42.5) 29(47.5) 13(43.3) 0.500 0.779 Recent surgical history (within 3 months) 48(22.6) 11(18.0) 4(13.3) 1.736 0.420 D-dimer elevated 201(94.4) 60(98.4) 29(93.5) 1.792 0.408

[0122] Model training: In this embodiment, the embolus detection model is trained based on the improved U-Net network model, and the model performance is optimized by five-fold cross-validation. Figure 11 This is a schematic diagram of an improved U-Net network model structure in this specification.

[0123] Model performance evaluation: The embolus detection model was tested on a 30-sample test set and compared with the U-Net and CE-Net networks in terms of Dice coefficient, recall, precision, specificity, and IOU. The experimental results are shown in Table 2 (Ours indicates the method of this invention), and experimental data result curves were generated. Figure 12a ~f, and Figure 13 The diagram shows radar indicators of the segmentation results for the three types of networks. Among them, Figure 12a This is a schematic diagram of the segmentation index curve for one of the embolus detection models in this specification. Figure 12b This is a schematic diagram of a CE-Net network segmentation index curve used in this specification. Figure 12c This is a schematic diagram of a U-Net network segmentation index curve used in this specification. Figure 12d This is a schematic diagram of the training index curves for one of the embolus detection models in this specification. Figure 12e This is a schematic diagram of the training metric curves for one type of CE-Net network described in this specification. Figure 12f This is a schematic diagram of a training metric curve for a U-Net network as described in this specification.

[0124] Table 2. Experimental results of the three networks on the test set.

[0125] Ours 0.894 0.925 0.899 1.000 0.828 CE_Net 0.812 0.913 0.754 1.000 0.724 U-Net 0.704 0.905 0.701 1.000 0.667

[0126] Figure 14 This is a schematic diagram of a segmentation result of a model in this specification. The first column is the original CTPA image, the second column is the gold standard image annotation, the third column is the segmentation result of the embolus detection model of this invention, the fourth column is the segmentation result of the CE-Net network, and the fifth column is the segmentation result of the U-Net network.

[0127] Analysis: The embolus detection model in this embodiment demonstrates excellent detection performance in terms of Dice coefficient, precision, recall, specificity, and IOU. On the test set, the model achieves a Dice coefficient of 0.894, precision of 0.899, recall of 0.925, and Pos_IOU of 0.828. Its embolus detection performance surpasses that of existing U-Net and CE-Net models.

[0128] This embodiment compares the segmentation performance of the embolus detection model with that of the U-Net and CE-Net models. The changes in Dice coefficient, recall, accuracy, and specificity of the three networks during training are as follows: Figure 12aAs shown in ~c, the values ​​of these four indicators range from 0 to 1. A higher value indicates better segmentation. Because the number of negative samples in pulmonary embolism images far exceeds the number of positive samples, the specificity remains essentially 1 regardless of the network used. The images clearly show that in the early stages of training, recall, precision, and Dice coefficient are unstable and fluctuate greatly, but overall, each indicator shows an increasing trend. After approximately 150 epochs, the indicators tend to stabilize. Compared to the U-Net and CE-Net networks, the embolism detection model, due to its combination of ASPP and SE attention mechanisms, has certain advantages in medical image segmentation tasks, with significant improvements in Dice coefficient, recall, and precision. Dice coefficient, recall, and precision are important indicators for evaluating the quality of segmentation results. The Dice coefficient measures the similarity between the predicted result and the true label, recall measures the network's coverage of positive samples, and precision measures the network's prediction accuracy. The improvement in these metrics means that emboli detection models can more accurately locate and identify lesion areas in clinical diagnosis, buying valuable time for timely diagnosis and enabling subsequent treatment to proceed more quickly. By improving the quality of segmentation results, doctors can understand the condition more accurately and provide patients with more timely and effective medical care.

[0129] Figure 12d ~d indicates that the training loss gradually decreases as the network is continuously optimized during training. This shows that the network's ability to fit the training data is constantly improving, and it can better learn the data features. The F1-score is an important indicator for measuring the performance of a classification model, comprehensively considering precision and recall. During training, the F1-score gradually increases, indicating that the model's ability to identify and classify positive and negative samples is improving. pos_IOU (Intersection over Union of positive samples) refers to the degree of overlap between positive samples and the predicted results, and is also one of the key indicators for measuring the performance of a segmentation model. During training, as the network learns, pos_IOU gradually increases, indicating that the model can more accurately predict the position and shape of positive samples. Since there are far more negative samples than positive samples, neg_IOU (Intersection over Union of negative samples) will approach 1. This is because negative samples are generally background regions, and background regions overlap less with other background regions, so neg_IOU will approach 1. In summary, the decrease in training loss and the increase in F1-score and pos_IOU all represent the optimization and performance improvement of the model during training.

[0130] These results demonstrate that the embolism detection model of this invention can accurately detect pulmonary embolism, and that it outperforms the other two models in terms of Dice coefficient, recall, and precision. The radar chart further illustrates this advantage. Figure 14In the visualization of the segmentation results, the embolus detection model of this invention is most similar to the gold standard of image annotation compared to the other two networks, indicating that the segmentation performance of the embolus detection model of this invention is superior.

[0131] In summary, the embolus detection model of this invention demonstrates excellent performance in APE detection, outperforming existing U-Net and CE-Net models. This is significant for improving the accuracy and efficiency of physicians' APE diagnosis. This achievement provides a new solution for intelligent APE detection in CTPA images and can play an important role in clinical practice.

[0132] The above are one or more embodiments of the embolus detection method in pulmonary artery imaging provided in this specification. Based on the same idea, this specification also provides a corresponding embolus detection device in pulmonary artery imaging, such as... Figure 15 As shown.

[0133] Figure 15 A schematic diagram of an embolus detection device in pulmonary artery imaging provided in this specification includes:

[0134] The acquisition module 201 is used to acquire historical pulmonary artery images of patients with acute pulmonary embolism as sample images and to label the actual embolism location in each sample image.

[0135] Module 202 is used to build an improved U-Net network model based on the U-Net network model, using the Res2Net network as the backbone module in its encoder, and adding a porous spatial pyramid pooling module and an attention mechanism module after the last layer skip connection of the decoder of the U-Net network model.

[0136] The training module 203 is used to input sample images into the improved U-Net network model to obtain the predicted embolus location; and to train the improved U-Net network model with the optimization objective of minimizing the deviation between the actual embolus location and the predicted embolus location in the sample image, thereby obtaining the embolus detection model.

[0137] The detection module 204 is used to input the pulmonary artery image to be detected into the embolism detection model. The encoder extracts embolism features at different scales of the pulmonary artery image to obtain multi-scale feature maps of emboli of different sizes. The decoder fuses the multi-scale feature maps of different sizes to obtain a fine feature map of the emboli. The porous spatial pyramid pooling module extracts features from the fine feature map at different scales and performs linear fusion of each feature channel to obtain an enhanced feature map of the emboli. The attention mechanism module assigns weights according to the contribution of different feature channels to embolism segmentation and weights each feature channel of the enhanced feature map to obtain a weighted feature map. The output layer performs embolism detection on the weighted feature map to obtain the embolism location.

[0138] Optionally, the acquisition module 201 acquires historical CT pulmonary artery images of patients with acute pulmonary embolism, performs quality checks on the historical CT pulmonary artery images, and uses historical CT pulmonary artery images with acceptable quality as sample images. Specifically, if the clinical and imaging data of the patient with acute pulmonary embolism are incomplete, or the historical pulmonary artery images are not 1 mm thin-slice images, or the historical pulmonary artery images contain artifacts, or the embolic signs in the historical pulmonary artery images cannot be marked, it is determined that the corresponding historical pulmonary artery images do not meet the quality requirements.

[0139] Optionally, the device further includes: a preprocessing module 205, used to convert the sample image into HU values, set the window level of the sample image to 100HU, set the window width to 900HU, perform centering and standardization processing on the adjusted sample image to obtain a sample image that follows a standard normal distribution, and set a mask on the lungs to enhance the sample image through various operations such as rotation, horizontal mirror flip, vertical mirror flip, and sharpening.

[0140] Optionally, the detection module 204 decomposes the convolution operations at different scales at the corresponding size within a single residual block using Res2Net networks of different sizes in the encoder, extracts emboli features at different scales of the pulmonary artery imaging to be detected at a fine-grained level, obtains feature combinations of different receptive field sizes and fuses them, and obtains multi-scale feature maps of emboli at different sizes.

[0141] Optionally, the detection module 204 performs convolution on the fine feature map using a porous spatial pyramid pooling module, employing dilated convolution with different dilated convolutional layer sampling rates corresponding to dilated rates in multiple separate parallel branches. The convolution results are then subjected to global average pooling to obtain multi-scale information of the pulmonary artery imaging emboli context information to be detected. Furthermore, the feature channels are linearly fused to obtain the enhanced feature map of the emboli.

[0142] Optionally, the detection module 204 compresses the enhanced feature map along the spatial dimension direction through the attention mechanism module, reduces the dimensionality of the feature map of each feature channel, assigns different weights to the feature map of each feature channel after dimensionality reduction and compression according to the contribution of different feature channels to thrombus segmentation, and multiplies the weights corresponding to each feature channel with the enhanced feature map channel by channel to obtain a weighted feature map.

[0143] Specific limitations regarding the embolus detection device in pulmonary artery imaging can be found in the limitations of the embolus detection method in pulmonary artery imaging described above, and will not be repeated here. Each module in the aforementioned embolus detection device in pulmonary artery imaging can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0144] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided method for embolus detection in pulmonary artery imaging.

[0145] This instruction manual also provides Figure 16 The schematic diagram of the computer device shown is as follows: Figure 16 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 The provided method for embolus detection in pulmonary artery imaging.

[0146] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0147] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for embolus detection in pulmonary artery imaging, characterized in that, include: Historical pulmonary artery imaging of patients with acute pulmonary embolism was obtained as sample images, and the actual location of the embolus in each sample image was labeled. Based on the U-Net network model, Res2Net network is used as the backbone module in its encoder, and a porous spatial pyramid pooling module and an attention mechanism module are added after the last layer skip connection of the U-Net network model decoder to construct an improved U-Net network model. The sample images are input into the improved U-Net network model to obtain the predicted embolus location; and the improved U-Net network model is trained with the optimization objective of minimizing the deviation between the actual embolus location and the predicted embolus location in the sample images to obtain the embolus detection model. The pulmonary artery image to be detected is input into the embolism detection model. The encoder extracts embolism features at different scales of the pulmonary artery image to obtain multi-scale feature maps of emboli of different sizes. The decoder fuses the multi-scale feature maps of different sizes to obtain a fine feature map of the emboli. The porous spatial pyramid pooling module extracts features from the fine feature map at different scales and performs linear fusion of each feature channel to obtain an enhanced feature map of the emboli. The attention mechanism module assigns weights according to the contribution of different feature channels to embolism segmentation and weights each feature channel of the enhanced feature map to obtain a weighted feature map. The output layer performs embolism detection on the weighted feature map to obtain the embolism location.

2. The embolus detection method in pulmonary artery imaging as described in claim 1, characterized in that, The acquisition of historical pulmonary artery imaging data from patients with acute pulmonary embolism as sample images specifically includes: Acquire historical CT pulmonary artery images of patients with acute pulmonary embolism, perform quality checks on the historical CT pulmonary artery images, and use historical CT pulmonary artery images that meet the quality requirements as sample images. In cases where the clinical and imaging data of patients with acute pulmonary embolism are incomplete, or the historical pulmonary artery imaging is not a 1 mm thin-slice image, or the historical pulmonary artery imaging contains artifacts, or the embolic signs in the historical pulmonary artery imaging cannot be marked, the corresponding historical pulmonary artery imaging quality is determined to be non-compliant.

3. The embolus detection method in pulmonary artery imaging as described in claim 1, characterized in that, The method further includes: Convert the sample image to HU value, and set the window level of the sample image to 100HU and the window width to 900HU; The adjusted sample images are centered and standardized to obtain sample images that follow a standard normal distribution, and a mask is set in the lungs; The sample image is enhanced through various operations such as rotation, horizontal mirror flip, vertical mirror flip, and sharpening.

4. The embolus detection method in pulmonary artery imaging as described in claim 1, characterized in that, The step of extracting embolus features at different scales of pulmonary artery imaging using an encoder to obtain multi-scale feature maps of emboli of different sizes specifically includes: By decomposing convolution operations at different scales within a single residual block using Res2Net networks of different sizes in the encoder, emboli features at different scales of the pulmonary artery imaging to be detected are extracted at a fine-grained level. Feature combinations of different receptive field sizes are obtained and fused to acquire multi-scale feature maps of emboli at different sizes.

5. The embolus detection method in pulmonary artery imaging as described in claim 1, characterized in that, The process of extracting features from the fine feature map at different scales using a porous spatial pyramid pooling module and linearly fusing the feature channels to obtain the enhanced feature map of the thrombus specifically includes: The fine feature map is convolved by a porous spatial pyramid pooling module, and dilated convolutions with different dilated convolutional layer sampling rates corresponding to dilated rates are performed in multiple separate parallel branches. Global average pooling is performed on the convolution results to obtain multi-scale information of the pulmonary artery imaging emboli to be detected, and linear fusion is performed on each feature channel to obtain the enhanced feature map of the emboli.

6. The embolus detection method in pulmonary artery imaging as described in claim 1, characterized in that, The process involves assigning weights to the contribution of different feature channels to embolus segmentation through an attention mechanism module, and then weighting each feature channel of the enhanced feature map to obtain a weighted feature map. Specifically, this includes: The attention mechanism module compresses the enhanced feature map along the spatial dimension, reducing the dimensionality of the feature map for each feature channel; The feature maps of each feature channel after dimensionality reduction and compression are assigned different weights according to the contribution of different feature channels to thrombus segmentation. The weights corresponding to each feature channel are multiplied with the enhanced feature map channel by channel to obtain the weighted feature map.

7. An embolus detection device for pulmonary artery imaging, characterized in that, include: The acquisition module is used to acquire historical pulmonary artery images of patients with acute pulmonary embolism as sample images and determine the annotation of the actual embolus location in each sample image; The building blocks are used to construct an improved U-Net network model based on the U-Net network model, using the Res2Net network as the backbone module in its encoder, and adding a porous spatial pyramid pooling module and an attention mechanism module after the last layer skip connection of the U-Net network model's decoder. The training module is used to input sample images into the improved U-Net network model to obtain the predicted embolus location; and to train the improved U-Net network model with the optimization objective of minimizing the deviation between the actual embolus location and the predicted embolus location in the sample image, so as to obtain the embolus detection model. The detection module inputs the pulmonary artery image to be detected into the embolism detection model. The encoder extracts embolism features at different scales from the pulmonary artery image, obtaining multi-scale feature maps of emboli at different sizes. The decoder fuses these multi-scale feature maps to obtain a fine-grained feature map of the emboli. A porous spatial pyramid pooling module extracts features from the fine-grained feature map at different scales and linearly fuses each feature channel to obtain an enhanced feature map of the emboli. An attention mechanism module assigns weights based on the contribution of different feature channels to embolism segmentation and weights each feature channel of the enhanced feature map to obtain a weighted feature map. The output layer performs embolism detection on the weighted feature map to determine the emboli location.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 6.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 6.

Images