A pulmonary embolism thrombus segmentation method and system based on a deep learning network

Abstract

The application relates to a pulmonary embolism thrombus segmentation method and system based on a deep learning network. The method obtains a processed pulmonary arteriography image as sample data and divides a training set and a verification set; an original image is obtained by sampling the training set, tubular structure enhancement is carried out based on a Hessian matrix, and a blood vessel enhanced image is obtained; the original image and the blood vessel enhanced image are spliced into a double-channel lung image and input into a deep learning network, and a thrombus segmentation probability graph is output; a loss weight is dynamically updated through a same variance uncertainty learning strategy by combining an overlap constraint loss, a shape constraint loss and a classification constraint loss, and a pulmonary embolism thrombus segmentation model is trained to generate a thrombus three-dimensional segmentation result.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a method and system for pulmonary embolism thrombus segmentation based on deep learning networks. Background Technology

[0002] Pulmonary embolism is usually caused by a thrombus blocking the pulmonary artery or its branches. Pulmonary angiography can visualize the pulmonary artery and its branches, and is an important basis for the imaging analysis of pulmonary embolism. Accurate segmentation of the thrombus area helps in subsequent determination of thrombus location, volume assessment, and auxiliary diagnostic and treatment analysis.

[0003] Current methods for segmenting pulmonary embolism thrombi include manual delineation, traditional image processing methods, and deep learning-based automatic segmentation methods. Manual delineation is labor-intensive, time-consuming, and easily affected by differences in physician experience.

[0004] With the development of deep learning technology, medical image segmentation methods based on 3D convolutional neural networks are increasingly being applied to pulmonary embolism thrombus segmentation. These methods can utilize the spatial contextual information in 3D lung images, exhibiting better continuous structure perception capabilities compared to 2D slice analysis. However, existing deep learning-based pulmonary embolism thrombus segmentation methods still have certain limitations.

[0005] On the one hand, pulmonary embolism thrombi typically account for a low percentage of voxels in 3D lung images, with background voxels far outnumbering thrombus voxels. During training, the model is easily dominated by the background region, tending to predict voxels as background, leading to missed thrombus segmentation, especially microthrombi located in distal branches or small vessels. On the other hand, pulmonary vessels have distinct tubular and tree-like branching structures, but existing segmentation networks mostly rely solely on image data for feature learning, failing to adequately incorporate prior knowledge of pulmonary vascular anatomy. This results in insufficient segmentation stability of the model at the ends of vascular branches, in narrow vascular lumens, and in areas with ambiguous boundaries. Summary of the Invention

[0006] To overcome, at least to some extent, the problem of insufficient accuracy in pulmonary embolism thrombus segmentation based on three-dimensional convolutional neural networks in related technologies, this application provides a pulmonary embolism thrombus segmentation method and system based on deep learning networks.

[0007] The proposed solution is as follows:

[0008] According to a first aspect of the embodiments of this application, a method for pulmonary embolism thrombus segmentation based on a deep learning network is provided, comprising: The processed pulmonary angiography images are used as sample data, and the sample data is divided into a training set and a validation set. Sampling is performed on the training set, and the sampling results are used as the original images; Construct a Hessian matrix, and perform tubular structure enhancement processing on the original image based on the Hessian matrix to obtain a blood vessel enhanced image; The original image and the enhanced vascular image are stitched together to obtain a dual-channel lung image; The dual-channel lung images are input into a pre-constructed deep learning network, which outputs a thrombus segmentation probability map. Based on the thrombus segmentation probability map, calculate the overlap constraint loss to reduce missed thrombus segmentation, the morphological constraint loss to constrain the continuity of thrombus boundaries, and the classification constraint loss to enhance the attention of difficult samples. Based on the homoscedastic uncertainty learning strategy, the loss weights of the overlapping constraint loss, morphological constraint loss and classification constraint loss are dynamically updated, and the deep learning network is iteratively trained until convergence based on the dynamically updated loss weights to obtain the pulmonary embolism thrombus segmentation model. The target lung 3D image is input into the pulmonary embolism thrombus segmentation model, the target thrombus segmentation probability map is output, and the pulmonary embolism thrombus 3D segmentation result is generated based on the target thrombus segmentation probability map.

[0009] Preferably, the method further includes: Acquire pulmonary angiography images and / or virtual angiography lung images as three-dimensional lung images to be processed; The three-dimensional lung image to be processed is subjected to grayscale truncation processing according to a preset grayscale truncation range. The lower limit of the preset grayscale truncation range corresponds to the grayscale range of the lung parenchyma, and the upper limit of the preset grayscale truncation range corresponds to the grayscale range of the enhanced blood vessels. The three-dimensional lung images to be processed after grayscale truncation are normalized to obtain normalized three-dimensional lung images. The normalized 3D lung image is randomly cropped according to the preset 3D cropping size to obtain the cropped 3D lung image. Online data augmentation processing is performed on the cropped three-dimensional lung image block to obtain an enhanced three-dimensional lung image; the online data augmentation processing includes one or more of random rotation processing, random flipping processing, and random elastic deformation processing; Enhanced three-dimensional lung images and their corresponding thrombus annotation maps were used as the sample data.

[0010] Preferably, the original image is subjected to tubular structure enhancement processing based on the Hessian matrix to obtain a vascular enhancement image, including: The original image is subjected to multi-scale smoothing processing to obtain smoothed original images at multiple scales; The second-order gray-level change information of the voxel neighborhood in the smoothed original image at each scale is extracted respectively, and the Hessian matrix at the corresponding scale is constructed based on the second-order gray-level change information. Based on the Hessian matrix, Frangi filtering is performed on the neighborhood of each voxel to determine the degree of matching between the neighborhood of each voxel and the pulmonary vascular tubular structure, thereby obtaining vascular tubular response images at various scales. The tubular response images of blood vessels at multiple scales are fused to obtain fused tubular response images of blood vessels; The fused vascular tubular response image is subjected to background suppression and intensity normalization to obtain the enhanced vascular image.

[0011] Preferably, the deep learning network includes an encoder and a decoder; The encoder has a channel attention module embedded at its end; the encoder is used to extract multi-level lung image features from dual-channel lung images, and at the end, the channel attention module performs channel attention recalibration on the multi-level lung image features to obtain thrombus enhancement features. The decoder has a global attention module embedded in its deep feature region; the decoder is used to upsample and recover the thrombus enhancement features, and capture the global dependencies in the thrombus enhancement features in the deep feature region through the global attention module, establish global associations between distal voxels, and output a thrombus segmentation probability map.

[0012] Preferably, the method further includes: With the goal of balancing positive and negative samples in thrombus segmentation, a joint loss function is constructed. The joint loss function includes: an overlap constraint loss to reduce missed thrombus segmentation, a morphological constraint loss to constrain the continuity of thrombus boundaries, and a classification constraint loss to enhance the attention of difficult samples.

[0013] Preferably, the loss of the joint loss function is calculated based on the thrombus segmentation probability map, including: Obtain the thrombus annotation map corresponding to the original image, and perform a voxel-level comparison between the thrombus segmentation probability map and the thrombus annotation map to determine the thrombus overlapping area, the thrombus missing segmentation area, and the thrombus missegmentation area. The overlap constraint loss is calculated based on the thrombus overlap region, the thrombus unsegmented region, and the thrombus missegmented region, and the penalty weight corresponding to the thrombus unsegmented region is made greater than the penalty weight corresponding to the thrombus missegmented region. Boundary extraction is performed on the thrombus segmentation probability map and the thrombus annotation map to obtain the predicted thrombus boundary and the labeled thrombus boundary. The morphological constraint loss is calculated based on the boundary deviation between the predicted thrombus boundary and the labeled thrombus boundary. Based on the prediction confidence of each voxel in the thrombus segmentation probability map and the category labeling of each voxel in the thrombus labeling map, the easy-to-segment sample voxels and difficult-to-segment sample voxels are determined. The classification constraint loss is obtained by reducing the contribution weight of easily classifiable voxels in the classification loss and increasing the contribution weight of difficult voxels in the classification loss.

[0014] Preferably, based on the homoscedastic uncertainty learning strategy, the loss weights of the overlap constraint loss, morphological constraint loss, and classification constraint loss are dynamically updated, including: Learnable uncertainty parameters are set for the overlap constraint loss, morphological constraint loss, and classification constraint loss, respectively; In each round of iterative training, the overlap constraint loss, morphological constraint loss and classification constraint loss for the current round are calculated based on the thrombus segmentation probability map and thrombus annotation map corresponding to the original image. Based on the learnable uncertainty parameters, the current loss weights corresponding to the overlap constraint loss, the morphological constraint loss, and the classification constraint loss are determined respectively, wherein the current loss weights are negatively correlated with the uncertainty of the corresponding loss. According to the current loss weight, the overlapping constraint loss, the morphological constraint loss and the classification constraint loss are weighted and fused, and combined with the constraint terms corresponding to each learnable uncertainty parameter, to obtain the dynamic joint loss value; Backpropagation is performed based on the dynamic joint loss value to synchronously update the network parameters of the deep learning network and each of the learnable uncertainty parameters; The loss weights of the overlap constraint loss, the morphological constraint loss, and the classification constraint loss are re-determined based on the updated learnable uncertainty parameters.

[0015] Preferably, the method further includes: The pulmonary embolism thrombus segmentation model was validated based on the validation set.

[0016] Preferably, the pulmonary embolism thrombus segmentation model is validated based on the validation set, including: Input the three-dimensional lung images from the validation set into the pulmonary embolism thrombus segmentation model to obtain the validation thrombus segmentation probability map; The verification thrombus segmentation probability map is binarized according to a preset probability threshold to obtain the verification thrombus three-dimensional segmentation result. The three-dimensional segmentation results of the verified thrombus are compared with the corresponding thrombus annotation information at the voxel level to obtain the verification overlapping region, the verification missing segmentation region, and the verification missegmentation region. Based on the verification overlap region, the verification missed segmentation region, and the verification missegmentation region, calculate the overlap similarity evaluation value, the recall evaluation value, and the precision evaluation value; Extract the predicted boundary of the three-dimensional segmentation result of the verified thrombus and the labeled boundary of the thrombus annotation information, and calculate the boundary accuracy evaluation value based on the boundary deviation between the predicted boundary and the labeled boundary; The validation results of the pulmonary embolism thrombus segmentation model are determined based on the overlap similarity evaluation value, the recall evaluation value, the precision evaluation value, and the boundary precision evaluation value.

[0017] According to a second aspect of the embodiments of this application, a pulmonary embolism thrombus segmentation system based on a deep learning network is provided, comprising: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute the program stored in the memory; The memory is used to store a program, which is at least used to execute a deep learning network-based pulmonary embolism thrombus segmentation method as described in any of the preceding claims.

[0018] The technical solution provided in this application may include the following beneficial effects: By performing tubular structure enhancement processing based on the Hessian matrix on the original image, enhanced vascular images are obtained. These images are then stitched together to form a dual-channel lung image. This allows the deep learning network to learn the grayscale features of the 3D lung image while simultaneously acquiring prior knowledge of the pulmonary vascular tubular structure, thereby enhancing the model's ability to identify the vascular region containing the thrombus and reducing interference from the background region on thrombus segmentation. Simultaneously, overlap constraint loss, morphological constraint loss, and classification constraint loss are used to constrain thrombus segmentation omissions, thrombus boundary continuity, and the ability to identify difficult samples, respectively. This ensures that the model training process balances the detection of small thrombi, boundary segmentation accuracy, and missegmentation suppression. Furthermore, a homoscedastic uncertainty learning strategy is used to dynamically update the weights of each loss, avoiding training bias caused by manually fixed weights and improving the model's training stability and generalization ability. Therefore, the accuracy, continuity, and reliability of 3D pulmonary embolism thrombus segmentation results can be improved, especially the segmentation effect of thrombi in distal branches, small vessels, and areas with blurred boundaries.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] Figure 1 This is a flowchart illustrating a method for pulmonary embolism thrombus segmentation based on a deep learning network, provided in one embodiment of this application. Figure 2 This is a schematic diagram of a pulmonary embolism thrombus segmentation system based on a deep learning network, provided in one embodiment of this application. Detailed Implementation

[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0023] Example 1 Figure 1 This is a flowchart illustrating a deep learning network-based pulmonary embolism thrombus segmentation method according to an embodiment of this application. (Refer to...) Figure 1 A method for pulmonary embolism thrombus segmentation based on deep learning networks, comprising: S1. Obtain the processed pulmonary angiography images as sample data, and divide the sample data into training set and validation set; S2. Sample the data in the training set and use the sampling results as the original images; S3. Construct the Hessian matrix, and perform tubular structure enhancement processing on the original image based on the Hessian matrix to obtain the enhanced blood vessel image. S4. The original image and the enhanced vascular image are stitched together to obtain a dual-channel lung image; S5. Input the dual-channel lung images into a pre-constructed deep learning network and output a thrombus segmentation probability map; S6. Based on the thrombus segmentation probability map, calculate the overlap constraint loss used to reduce missed thrombus segmentation, the morphological constraint loss used to constrain the continuity of thrombus boundaries, and the classification constraint loss used to enhance the attention of difficult samples. S7. Based on the homoscedastic uncertainty learning strategy, the loss weights of overlap constraint loss, morphological constraint loss and classification constraint loss are dynamically updated, and the deep learning network is iteratively trained until convergence based on the dynamically updated loss weights to obtain the pulmonary embolism thrombus segmentation model. S8. Input the target lung 3D image into the pulmonary embolism thrombus segmentation model, output the target thrombus segmentation probability map, and generate the pulmonary embolism thrombus 3D segmentation result based on the target thrombus segmentation probability map.

[0024] For ease of understanding, the following explains some key terms in this embodiment: Pulmonary angiography refers to imaging data obtained by X-ray or CT scanning after injecting a contrast agent into the pulmonary artery. This image clearly shows the structure of the pulmonary artery and its branches, and is an important basis for the diagnosis and assessment of pulmonary embolism.

[0025] The Hessian matrix is ​​a square matrix composed of the second-order partial derivatives of a multivariable function. In image processing, this matrix is ​​used to describe the gray-level variation characteristics of local regions of an image, and it is particularly valuable for identifying structural features in images, such as edges, corners, or tubular structures.

[0026] Tubular structure enhancement refers to the use of specific image processing algorithms, such as Frangi filtering based on the Hessian matrix, to enhance tubular structures in images, making them more prominent. This process helps to highlight tissue structures with tubular morphology, such as blood vessels.

[0027] Deep learning networks are artificial neural networks that contain multiple processing layers. These networks can automatically extract features by learning from large amounts of data and perform tasks such as classification and segmentation. In this embodiment, the network is used to identify and segment thrombus regions from lung images.

[0028] A thrombus segmentation probability map is an image output by a deep learning network. Each voxel (or pixel) in this image is assigned a probability value between 0 and 1, representing the likelihood that the voxel belongs to a thrombus region. The higher the probability value, the greater the likelihood that the voxel is a thrombus.

[0029] Overlap constraint loss is a loss function used to measure the degree of overlap between the predicted segmentation result and the ground truth annotation. This loss function aims to optimize the model to maximize the overlap between the predicted thrombus region and the actual thrombus region, thereby reducing the chance of missing thrombus segmentation.

[0030] Morphological constraint loss is a loss function used to measure the difference between the predicted segmentation boundary and the true labeled boundary. This loss function aims to optimize the model so that its predicted thrombus boundaries are smoother and more continuous, conforming to the actual morphological characteristics of the thrombus.

[0031] Classification constraint loss is a loss function used to measure the accuracy of a model in classifying voxels of different categories (such as thrombus voxels and background voxels). This loss function aims to optimize the model's ability to identify thrombus voxels, with particular attention to difficult samples that are hard to classify correctly.

[0032] Homoscedastic uncertainty learning strategy refers to a strategy used to dynamically adjust the weights of different loss functions in multi-task learning. This strategy introduces learnable uncertainty parameters to automatically adjust the contribution of each loss function to the total loss based on its degree of uncertainty, thereby avoiding training bias caused by manually setting weights and improving the training stability and generalization ability of the model.

[0033] Specifically, this method for pulmonary embolism thrombus fragmentation includes the following steps: First, processed pulmonary angiography images are acquired as sample data. This sample data can be images that have undergone preliminary processing such as filtering, cropping, or simple grayscale adjustment. Subsequently, this sample data is divided into a training set and a validation set, where the training set is used for model learning and the validation set is used to evaluate the model's performance.

[0034] Furthermore, during the model training phase, samples are taken from the training set, and the sampled results are used as raw images. This sampling can be done by randomly selecting images from the training set or by selecting them in a preset order. These raw images will serve as input for subsequent processing and the deep learning network.

[0035] To enhance tubular structure information in images, a Hessian matrix is ​​constructed, and the original image is then processed using this Hessian matrix to enhance the tubular structures, resulting in an enhanced vascular image. For example, the second-order partial derivatives of each voxel in the image can be calculated to form a Hessian matrix, and then tubular structures can be identified and enhanced through simple eigenvalue analysis or a preset threshold. This processing aims to highlight the morphological features of blood vessels.

[0036] Subsequently, the original image and the enhanced vascular image are stitched together to obtain a dual-channel lung image. For example, the original image can be used as the first channel, and the enhanced vascular image as the second channel, and the two can be merged along the channel dimension. This dual-channel image will then serve as the input to a deep learning network.

[0037] The dual-channel lung image is input into a pre-constructed deep learning network. This deep learning network is preferably 3DU-Net, but can also be replaced by V-Net, nnU-Net, or UNETR, etc., and contains multiple convolutional and pooling layers to extract features from the input image. After processing, the network outputs a thrombus segmentation probability map. Each voxel in this probability map represents the probability that it belongs to a thrombus.

[0038] During training, various loss functions are calculated based on the thrombus segmentation probability map. These loss functions include overlap constraint loss to reduce missed thrombus segmentation, morphological constraint loss to constrain the continuity of thrombus boundaries, and classification constraint loss to enhance the attention given to difficult samples. For example, overlap constraint loss can employ Dice loss or Jaccard loss, morphological constraint loss can use distance-transform-based loss, and classification constraint loss can use cross-entropy loss or Focal loss. These loss functions optimize the model's segmentation performance from different perspectives.

[0039] To balance the impact of different loss functions on model training, the loss weights of the overlap constraint loss, morphological constraint loss, and classification constraint loss are dynamically updated based on the homoscedastic uncertainty learning strategy. For example, an initial weight can be preset for each loss function, and manually adjusted during training based on its contribution to the total loss or its changing trend. Based on the dynamically updated loss weights, the deep learning network is iteratively trained until convergence, resulting in a pulmonary embolism thrombus segmentation model. Convergence typically means that the model's performance on the validation set no longer shows significant improvement.

[0040] Finally, the 3D image of the target lung is input into the pulmonary embolism thrombus segmentation model. This target lung 3D image can be any CT or contrast-enhanced image of the lung to be analyzed. The model outputs a target thrombus segmentation probability map, which represents the probability that each voxel in the target image belongs to a thrombus. Based on this target thrombus segmentation probability map, binarization is performed by setting a probability threshold to generate the 3D segmentation result of the pulmonary embolism thrombus, thereby achieving automatic identification and localization of the thrombus region.

[0041] The technical solution in this embodiment enhances the original image using tubular structure enhancement based on the Hessian matrix to obtain enhanced vascular images. The original image and the enhanced vascular images are then stitched together to form a dual-channel lung image. This allows the deep learning network to learn the grayscale features of the 3D lung image while simultaneously acquiring prior knowledge of the pulmonary vascular tubular structure, thereby enhancing the model's ability to identify the vascular region containing the thrombus and reducing interference from the background region on thrombus segmentation. Simultaneously, overlap constraint loss, morphological constraint loss, and classification constraint loss are used to constrain thrombus segmentation omissions, thrombus boundary continuity, and the ability to identify difficult samples, respectively. This ensures that the model training process can balance the detection of small thrombi, boundary segmentation accuracy, and missegmentation suppression. Furthermore, a homoscedastic uncertainty learning strategy is used to dynamically update the weights of each loss, avoiding training bias caused by manually fixing weights and improving the model's training stability and generalization ability. Therefore, the accuracy, continuity, and reliability of the 3D segmentation results of pulmonary embolism thrombi can be improved, especially the segmentation effect of thrombi in distal branches, small vessels, and areas with blurred boundaries.

[0042] Example 2 In some embodiments, the method further includes: Acquire pulmonary angiography images and / or virtual angiography lung images as three-dimensional lung images to be processed; The three-dimensional lung image to be processed is truncated according to the preset grayscale truncation range. The lower limit of the preset grayscale truncation range corresponds to the grayscale range of the lung parenchyma, and the upper limit of the preset grayscale truncation range corresponds to the grayscale range of the enhanced blood vessels. The three-dimensional lung images to be processed after grayscale truncation are normalized to obtain normalized three-dimensional lung images. The normalized 3D lung image is randomly cropped according to the preset 3D cropping size to obtain the cropped 3D lung image. Online data augmentation processing is performed on the cropped 3D lung image block to obtain an enhanced 3D lung image; the online data augmentation processing includes one or more of random rotation processing, random flipping processing, and random elastic deformation processing; Enhanced 3D lung images and their corresponding thrombus annotation maps were used as sample data.

[0043] Specifically, this application, when performing pulmonary embolism thrombus segmentation, first acquires raw lung imaging data. This imaging data can be clinically common pulmonary angiography images, which enhance vascular visualization through contrast agent injection, aiding in the observation of vascular structures and thrombi. Alternatively, it can be virtual angiography lung images, which simulate the angiography effect from non-contrast CT images using specific image processing techniques, providing a more diverse data source for model training and thus improving the model's generalization ability. These images typically exist in the form of three-dimensional data, containing rich spatial information.

[0044] To focus on the tissue structures most relevant to the pulmonary embolism thrombus segmentation task and remove irrelevant background noise or high-density bone interference, grayscale truncation is performed on the acquired 3D lung images to be processed. This processing is achieved by setting a preset grayscale truncation range. Specifically, the lower limit of this range is set to the grayscale range corresponding to the lung parenchyma, aiming to retain the main tissue information of the lung; while the upper limit corresponds to the grayscale range of enhanced vessels, ensuring that the contrast-enhanced vessels and any thrombus areas within them are completely preserved. This grayscale truncation effectively reduces data redundancy, highlights the target area, and provides a cleaner input for subsequent processing.

[0045] After grayscale truncation, images are normalized to eliminate inconsistencies in grayscale value ranges caused by different imaging devices, scanning parameters, or individual patient differences. Normalization maps the grayscale values ​​of an image to a uniform standard range (e.g., [0, 1] or [-1, 1]), ensuring that the data input to the deep learning network has a consistent scale and distribution. This helps accelerate model convergence, improve training stability, and enhance the model's adaptability to data from different sources. Common normalization methods include min-max normalization or Z-score normalization.

[0046] Since raw 3D lung images are typically large, directly inputting them into a deep learning network can lead to excessive computational resource consumption or memory overflow. Therefore, random cropping is performed on the normalized 3D lung images. This cropping operation is performed according to a preset 3D cropping size, randomly selecting a sub-region from the original image as the cropped 3D lung image block. Random cropping not only decomposes large images into manageable blocks but also increases the diversity of training data to some extent, forcing the model to focus on local features, thereby improving the model's generalization ability and effectively mitigating overfitting.

[0047] To further expand the training dataset and improve the model's robustness and generalization ability, online data augmentation was performed on the cropped 3D lung image patches. Online data augmentation means transforming the data in real time during model training, rather than pre-generating a large amount of augmented data. This process can include random rotation, which simulates images from different positions or scanning angles by rotating the images at different angles; random flipping, which increases the symmetry variation of the data by flipping the images along different axes; and random elastic deformation, which simulates the minute deformations of biological tissues in the real world by applying small, random, nonlinear deformations to the pixel positions of the images, thereby making the model more robust to deformation. These augmentation operations can effectively increase the diversity of training samples, reduce the model's dependence on specific data patterns, and thus improve the model's performance when faced with unknown data.

[0048] After the aforementioned series of preprocessing and data augmentation steps, the resulting enhanced 3D lung images, along with their corresponding thrombus annotation maps labeled by professional doctors or experts, constitute the sample data used for training the deep learning network. The thrombus annotation maps provide the true label of whether each voxel belongs to a thrombus, and are a key component of supervised learning. This carefully prepared sample data provides high-quality, diverse, and representative input to the deep learning network, ensuring that the model can effectively learn the characteristics of pulmonary embolism thrombi and accurately segment them.

[0049] Example 3 In some embodiments, tubular structure enhancement processing is performed on the original image based on the Hessian matrix to obtain a vascular enhancement image, including: Multi-scale smoothing processing is performed on the original image to obtain smoothed original images at multiple scales; Second-order gray-level variation information of voxel neighborhood in smoothed original images at each scale is extracted, and Hessian matrix of corresponding scale is constructed based on the second-order gray-level variation information. Based on the Hessian matrix, Frangi filtering is applied to the neighborhood of each voxel to determine the degree of matching between the neighborhood of each voxel and the pulmonary vascular tubular structure, thus obtaining vascular tubular response images at various scales. The tubular response images of blood vessels at multiple scales are fused to obtain fused tubular response images of blood vessels; Background suppression and intensity normalization were performed on the fused vascular tubular response images to obtain enhanced vascular images.

[0050] Specifically, multi-scale smoothing of the original images aims to eliminate random noise and provide contextual information at different scales for subsequent feature extraction. Since the diameters of pulmonary vessels vary considerably, a single-scale smoothing kernel may not be sufficient to capture all vessel sizes. By smoothing at multiple scales, such as using Gaussian filters with different standard deviations, a series of smoothed images at different levels of detail can be generated, thus preparing the image for subsequent steps to capture vascular structures of varying sizes.

[0051] Building upon this, a Hessian matrix can be constructed from the smoothed original images at various scales by calculating the second-order partial derivatives of each voxel's neighborhood. The Hessian matrix is ​​a symmetric matrix describing the local curvature of a function; its eigenvalues ​​and eigenvectors reflect the direction and intensity of grayscale changes in local image regions. For tubular structures, their cross-sections typically exhibit a pattern of gradually decreasing grayscale values ​​from the center to the edges, which is specifically reflected in the eigenvalues ​​of the Hessian matrix, thus providing a mathematical basis for identifying tubular structures.

[0052] Subsequently, Frangi filtering is applied to the neighborhood of each voxel based on the constructed Hessian matrix. The Frangi filter is a classic vascular enhancement algorithm that utilizes the eigenvalues ​​of the Hessian matrix to quantify the degree of matching between the voxel neighborhood and the ideal tubular structure. By analyzing the relationships between eigenvalues, the Frangi filter can effectively distinguish between tubular structures (such as blood vessels) and sheet-like structures (such as organ edges) or speckled structures (such as noise), and suppress non-vascular regions, thereby obtaining vascular tubular response images sensitive to vascular structure responses at various scales.

[0053] To comprehensively utilize vascular information at different scales, tubular vascular response images obtained at multiple scales are fused. Since smoothing and Frangi filtering at each scale provide optimal response for vessels of specific sizes, fusion processing, such as taking the maximum response value of each voxel in the response images at each scale, ensures that vessels of different diameters are effectively enhanced, thereby generating a fused tubular vascular response image containing more comprehensive and robust vascular information.

[0054] Finally, background suppression processing is performed on the fused vascular tubular response images to further remove residual responses from non-vascular regions, making the vascular structures more prominent. For example, a threshold can be set to zero for response values ​​below that threshold. Subsequently, intensity normalization processing is performed to adjust the grayscale value range of the images to a preset standard range (e.g., 0 to 1). This helps to eliminate the influence of intensity differences between different images, making the features of the enhanced vascular images more stable and facilitating subsequent processing and learning by deep learning networks.

[0055] Example 4 In some embodiments, the deep learning network includes an encoder and a decoder; The encoder has a channel attention module embedded at its end; the encoder is used to extract multi-level lung image features from dual-channel lung images, and at the end, the channel attention module performs channel attention recalibration on the multi-level lung image features to obtain thrombus enhancement features. The decoder embeds a global attention module in its deep feature region; the decoder is used to upsample and recover the thrombus enhancement features, and captures the global dependencies in the thrombus enhancement features in the deep feature region through the global attention module, establishes global associations between distal voxels, and outputs a thrombus segmentation probability map.

[0056] The encoder incorporates a channel attention module at its end. This module is a mechanism designed to enable the network to adaptively learn the importance of each feature channel. It weights the channel dimensions of the input feature map, allowing the network to focus more on feature channels that contribute more to the current task (e.g., thrombus segmentation) while suppressing less important channels. This module typically processes these features at the end of the encoder after they have extracted multiple levels of features to enhance their discriminative power. Specifically, the encoder extracts multiple levels of lung image features from dual-channel lung images and performs channel attention recalibration on these features at the end using the channel attention module to obtain thrombus enhancement features. The encoder receives dual-channel lung images as input and progressively extracts lung image features at different scales through multiple convolution and pooling operations. These features contain rich content ranging from low-level texture edges to high-level semantic information. At the end of the encoder, the channel attention module processes these multiple levels of features, adaptively recalibrating them by calculating the weight of each channel. This recalibration enables the network to highlight thrombus-related feature channels and suppress background or noise channels, thereby generating thrombus enhancement features that are more discriminative of thrombus regions.

[0057] The decoder embeds a global attention module in its deep feature regions. This global attention module aims to capture long-distance dependencies between any two locations in the feature map, thereby establishing global contextual information. Embedding the global attention module in the deep feature regions of the decoder allows the network to comprehensively consider the contextual information of the entire image, rather than just local regions, during feature recovery and upsampling. This is particularly important for segmenting irregularly shaped and widely distributed thrombus structures, contributing to more coherent and accurate segmentation results. Specifically, the decoder is used to upsample and recover the thrombus enhancement features, and captures global dependencies in the thrombus enhancement features in the deep feature regions through the global attention module, establishing global associations between distal voxels and outputting a thrombus segmentation probability map. The decoder receives the thrombus enhancement features output by the encoder and progressively performs upsampling operations to restore the spatial resolution of the feature map. In this process, the global attention module in the deep regions of the decoder plays a crucial role. It captures global dependencies in the thrombus enhancement features by calculating the interrelationships between all voxels in the feature map, thereby establishing associations between distal voxels. This global correlation enables the network to understand the overall structure and distribution of thrombi, avoiding local misjudgments and ensuring the consistency and completeness of the segmentation results. Finally, the decoder outputs the probability that each voxel belongs to a thrombus, i.e., the thrombus segmentation probability map.

[0058] Example 5 In some embodiments, the method further includes: With the goal of balancing positive and negative samples in thrombus segmentation, a joint loss function is constructed. The joint loss function includes: overlap constraint loss to reduce missed thrombus segmentation, morphological constraint loss to constrain the continuity of thrombus boundaries, and classification constraint loss to enhance the attention of difficult samples.

[0059] A joint loss function combines multiple loss terms of different types into a comprehensive loss function. In training deep learning models, a single loss function may not fully capture the complexity of segmentation tasks. For example, pixel-level classification losses may be insensitive to boundary details, while region-based losses may be insensitive to small targets. By constructing a joint loss function, multiple performance metrics of the model can be optimized simultaneously. For instance, it can simultaneously focus on pixel classification accuracy, region overlap, boundary fineness, and handling of difficult samples, enabling the model to learn more robust and accurate segmentation capabilities from different dimensions.

[0060] Overlap constraint loss, used to reduce thrombus omission segmentation, measures the degree of overlap between predicted and ground truth regions and imposes a greater penalty for omissions. Omissions occur when the model fails to correctly identify an actual thrombus region, incorrectly predicting it as background. In pulmonary embolism diagnosis, omissions can lead to serious clinical consequences. For example, region overlap-based loss functions such as Dice or Jaccard loss can be used. These functions directly reflect the similarity between the predicted and ground truth regions, and optimizing this loss maximizes the overlap between the predicted and ground truth regions, effectively reducing the thrombus omission rate.

[0061] Morphological constraint losses, used to constrain the continuity of thrombus boundaries, aim to encourage models to generate smoother, more biologically morphologically consistent thrombus boundaries. Boundary information of thrombi is crucial for accurately assessing their size, shape, and location. However, traditional pixel-level classification loss functions often fail to effectively constrain the smoothness and continuity of segmentation boundaries, potentially resulting in jagged, discontinuous, or overly coarse segmentation boundaries. This can be addressed by calculating the distance between the predicted and true boundaries, gradient information, or by utilizing morphological operations to define the loss term, thereby improving the clinical interpretability and accuracy of segmentation results.

[0062] The classification constraint loss, designed to enhance the focus on difficult samples, aims to adjust the loss contribution weights of different samples, enabling the model to pay more attention to those difficult samples that are hard to classify correctly during training. During training, the model tends to correctly classify samples with obvious and easily distinguishable features (easy samples), while it is prone to errors in classifying samples with ambiguous features, high similarity to the background, or located in boundary regions (difficult samples). If difficult samples are not given sufficient attention, the model may not be able to fully learn their discriminative features. For example, strategies such as Focal Loss or OHEM (Online Hard Example Mining) can be used to dynamically increase the loss weight of difficult samples or decrease the loss weight of easy samples, thereby forcing the model to better learn and distinguish these key difficult samples and improving overall segmentation performance.

[0063] Based on this, the loss of the joint loss function is calculated according to the thrombus segmentation probability map, including: Obtain the thrombus annotation map corresponding to the original image, and perform a voxel-level comparison between the thrombus segmentation probability map and the thrombus annotation map to determine the overlapping thrombus area, the thrombus missed segmentation area, and the thrombus missegmentation area. Calculate the overlap constraint loss based on the overlapping region of thrombus, the unsegmented region of thrombus, and the missegmented region of thrombus, and make the penalty weight corresponding to the unsegmented region of thrombus greater than the penalty weight corresponding to the missegmented region of thrombus. Boundary extraction is performed on the thrombus segmentation probability map and the thrombus annotation map to obtain the predicted thrombus boundary and the labeled thrombus boundary. The morphological constraint loss is calculated based on the boundary deviation between the predicted thrombus boundary and the labeled thrombus boundary. Based on the prediction confidence of each voxel in the thrombus segmentation probability map and the category labeling of each voxel in the thrombus labeling map, the easy-to-segment sample voxels and difficult-to-segment sample voxels are determined. The classification constraint loss is obtained by reducing the contribution weight of easily classifiable voxels in the classification loss and increasing the contribution weight of difficult voxels in the classification loss.

[0064] This application refines the calculation method of the joint loss function to address the specific characteristics of thrombus segmentation tasks. By using voxel-level comparison to identify overlapping, missed, and mis-segmented thrombus regions, and applying a higher penalty weight to missed segments, the risk of missed thrombus segmentation is effectively reduced, ensuring comprehensive thrombus identification in clinical diagnosis. Simultaneously, by extracting predicted and labeled thrombus boundaries and calculating their deviations, the morphological constraint loss guides the model to learn more accurate thrombus edge information, resulting in more continuous and accurate segmentation boundaries. Furthermore, by identifying and increasing the contribution weight of difficult sample voxels in the classification loss, the model can more effectively focus on those hard-to-distinguish thrombus regions or edge voxels, thereby enhancing the model's robustness and ability to identify complex thrombus morphologies. This comprehensive loss calculation strategy allows the deep learning network to optimize its parameters more comprehensively and precisely during training, ultimately obtaining a more accurate and reliable pulmonary embolism thrombus segmentation model, significantly improving segmentation accuracy and clinical applicability.

[0065] Furthermore, based on the homoscedastic uncertainty learning strategy, the loss weights of the overlap constraint loss, morphological constraint loss, and classification constraint loss are dynamically updated, including: Learnable uncertainty parameters are set for overlap constraint loss, morphological constraint loss, and classification constraint loss, respectively; In each round of iterative training, the overlap constraint loss, morphological constraint loss and classification constraint loss for the current round are calculated based on the thrombus segmentation probability map and thrombus annotation map corresponding to the original image. Based on each learnable uncertainty parameter, the current loss weights corresponding to the overlap constraint loss, morphological constraint loss, and classification constraint loss are determined respectively, wherein the current loss weights are negatively correlated with the uncertainty of the corresponding loss. Based on the current loss weights, the overlap constraint loss, morphological constraint loss, and classification constraint loss are weighted and fused together, and combined with the constraint terms corresponding to each learnable uncertainty parameter, to obtain the dynamic joint loss value; Backpropagation is performed based on dynamic joint loss values ​​to synchronously update the network parameters and learnable uncertainty parameters of the deep learning network. The loss weights of overlap constraint loss, morphological constraint loss, and classification constraint loss are redefined based on the updated learnable uncertainty parameters.

[0066] Specifically, this method first sets learnable uncertainty parameters for the overlap constraint loss, morphological constraint loss, and classification constraint loss, respectively. These learnable uncertainty parameters are variables that the model can automatically adjust during training, and they are typically designed to reflect the noise level or reliability of the corresponding loss term. For example, a larger uncertainty parameter value for a loss term indicates greater noise or lower reliability, and vice versa.

[0067] During each iteration of training, the system calculates the specific values ​​of overlap constraint loss, morphological constraint loss, and classification constraint loss for the current round based on the thrombus segmentation probability map and thrombus annotation map corresponding to the original image. Then, based on each learnable uncertainty parameter, it determines the current loss weights corresponding to the overlap constraint loss, morphological constraint loss, and classification constraint loss, respectively. In this mechanism, the current loss weight is negatively correlated with the uncertainty level of the corresponding loss. This means that for loss terms that the model considers to have high uncertainty (i.e., high noise or low reliability), their corresponding weights are automatically reduced; while for loss terms with low uncertainty (i.e., low noise or high reliability), their weights are correspondingly increased. This dynamic adjustment mechanism allows the model to more intelligently allocate the importance of each loss term.

[0068] Next, the system weights and fuses the overlap constraint loss, morphological constraint loss, and classification constraint loss according to the current loss weights, and combines this with the constraint terms corresponding to each learnable uncertainty parameter to obtain a dynamic joint loss value. This constraint term is typically a regularization term used to prevent the uncertainty parameters from being too large or too small, thereby stabilizing the training process. Finally, backpropagation is performed based on the dynamic joint loss value to synchronously update the network parameters and each learnable uncertainty parameter of the deep learning network. This means that in each iteration, not only is the model's segmentation ability improved, but the model's "trust" in each loss term is also continuously learned and optimized. Based on the updated learnable uncertainty parameters, the system redetermines the loss weights of the overlap constraint loss, morphological constraint loss, and classification constraint loss, preparing for the next round of training iterations.

[0069] Example 6 In some embodiments, the method further includes: The pulmonary embolism thrombus segmentation model was validated using a validation set, including: Input the validation lung 3D images from the validation set into the pulmonary embolism thrombus segmentation model to obtain the validation thrombus segmentation probability map; The validation thrombus segmentation probability map is binarized according to a preset probability threshold to obtain the validation thrombus three-dimensional segmentation result. The three-dimensional segmentation results of the verified thrombus are compared with the corresponding thrombus annotation information at the voxel level to obtain the verification overlapping area, the verification missing segmentation area, and the verification missegmentation area. Based on the verification overlapping region, the verification missed segmentation region, and the verification missegmentation region, calculate the overlap similarity evaluation value, recall evaluation value, and precision evaluation value; Extract the predicted boundary and the labeled boundary of the thrombus annotation information to verify the 3D segmentation results of the thrombus, and calculate the boundary accuracy evaluation value based on the boundary deviation between the predicted boundary and the labeled boundary. The validation results of the pulmonary embolism thrombus segmentation model were determined based on the overlap similarity evaluation value, recall evaluation value, precision evaluation value, and boundary precision evaluation value.

[0070] Specifically, during the validation process, the validation lung 3D images from the validation set are first input into the trained pulmonary embolism thrombus segmentation model. The validation lung 3D images are usually preprocessed lung 3D images. The model receives these images as input, performs feature extraction and upsampling recovery through its internal encoder-decoder structure, and finally generates a probability value between 0 and 1 for each voxel in the output layer, representing the confidence that the voxel is a thrombus, thus obtaining the validation thrombus segmentation probability map.

[0071] Subsequently, to obtain the final binarized segmentation result, the validation thrombus segmentation probability map needs to be binarized according to a preset probability threshold to obtain the validation thrombus 3D segmentation result. The preset probability threshold is usually determined based on experience or by optimization on a small portion of validation data, for example, 0.5. For each voxel in the validation thrombus segmentation probability map, if its probability value is greater than or equal to the threshold, it is marked as 1 (thrombus), otherwise it is marked as 0 (non-thrombus), thus forming a binarized 3D image.

[0072] To evaluate the accuracy of the segmentation model, the 3D segmentation results of the thrombus were compared with the corresponding thrombus annotation information at the voxel level. The thrombus annotation information is usually generated manually or by semi-automatic tools by professional doctors or experts. Voxel-level comparison involves logical operations to accurately identify areas where the predicted results are consistent with or inconsistent with the actual situation, thereby obtaining the verification overlapping areas (true positives), verification missed segmentation areas (false negatives), and verification mis-segmentation areas (false positives).

[0073] Based on this, overlap similarity evaluation value, recall evaluation value, and precision evaluation value are calculated according to the validation overlapping region, validation missed segmentation region, and validation missegmentation region. The overlap similarity evaluation value (e.g., Dice similarity coefficient) measures the degree of spatial overlap between the predicted result and the ground truth annotation; a higher value indicates better overlap. The recall evaluation value measures the model's ability to correctly identify all true thrombus voxels, i.e., its ability to avoid missed diagnoses. The precision evaluation value measures how many voxels predicted as thrombi are actually thrombi, i.e., its ability to avoid misdiagnoses. These indicators are calculated using the number of true positive, false positive, and false negative voxels.

[0074] Furthermore, to more precisely evaluate the segmentation results, the predicted boundaries and labeled boundaries of the thrombus 3D segmentation results are extracted to verify the thrombus annotation information. A boundary accuracy evaluation value is then calculated based on the boundary deviation between the predicted and labeled boundaries. Boundary extraction can be achieved through image processing algorithms (such as Canny edge detection, Sobel operator, or morphological operations) from the binarized segmentation results and annotation information. Boundary deviation calculation can employ metrics such as Hausdorff distance or mean surface distance to quantify the difference between the predicted and true boundaries, reflecting the level of detail in the segmentation results.

[0075] Finally, the validation results of the pulmonary embolism thrombus segmentation model are determined based on the overlap similarity evaluation value, recall evaluation value, precision evaluation value, and boundary precision evaluation value. These evaluation values ​​are typically aggregated and analyzed, for example, by setting acceptance criteria for each indicator or calculating a weighted average of these indicators to obtain a comprehensive score.

[0076] Example 7 A pulmonary embolism thrombus segmentation system based on deep learning networks, referring to Figure 2 ,include: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute programs stored in memory. A memory for storing a program, which is at least used to execute a deep learning network-based pulmonary embolism thrombus segmentation method as described in any of the above embodiments.

[0077] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0078] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.

[0079] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0080] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0081] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0082] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0083] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0084] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0085] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for pulmonary embolism thrombus segmentation based on deep learning networks, characterized in that, include: The processed pulmonary angiography images are used as sample data, and the sample data is divided into a training set and a validation set. Sampling is performed on the training set, and the sampling results are used as the original images; Construct a Hessian matrix, and perform tubular structure enhancement processing on the original image based on the Hessian matrix to obtain a blood vessel enhanced image; The original image and the enhanced vascular image are stitched together to obtain a dual-channel lung image; The dual-channel lung images are input into a pre-constructed deep learning network, which outputs a thrombus segmentation probability map. Based on the thrombus segmentation probability map, calculate the overlap constraint loss to reduce missed thrombus segmentation, the morphological constraint loss to constrain the continuity of thrombus boundaries, and the classification constraint loss to enhance the attention of difficult samples. Based on the homoscedastic uncertainty learning strategy, the loss weights of the overlapping constraint loss, morphological constraint loss and classification constraint loss are dynamically updated, and the deep learning network is iteratively trained until convergence based on the dynamically updated loss weights to obtain the pulmonary embolism thrombus segmentation model. The target lung 3D image is input into the pulmonary embolism thrombus segmentation model, the target thrombus segmentation probability map is output, and the pulmonary embolism thrombus 3D segmentation result is generated based on the target thrombus segmentation probability map.

2. The method according to claim 1, characterized in that, The method further includes: Acquire pulmonary angiography images and / or virtual angiography lung images as three-dimensional lung images to be processed; The three-dimensional lung image to be processed is subjected to grayscale truncation processing according to a preset grayscale truncation range. The lower limit of the preset grayscale truncation range corresponds to the grayscale range of the lung parenchyma, and the upper limit of the preset grayscale truncation range corresponds to the grayscale range of the enhanced blood vessels. The three-dimensional lung images to be processed after grayscale truncation are normalized to obtain normalized three-dimensional lung images. The normalized 3D lung image is randomly cropped according to the preset 3D cropping size to obtain the cropped 3D lung image. Online data augmentation processing is performed on the cropped three-dimensional lung image block to obtain an enhanced three-dimensional lung image; the online data augmentation processing includes one or more of random rotation processing, random flipping processing, and random elastic deformation processing; Enhanced three-dimensional lung images and their corresponding thrombus annotation maps were used as the sample data.

3. The method according to claim 1, characterized in that, The original image is enhanced with tubular structures based on the Hessian matrix to obtain a vascular enhancement image, including: The original image is subjected to multi-scale smoothing processing to obtain smoothed original images at multiple scales; The second-order gray-level change information of the voxel neighborhood in the smoothed original image at each scale is extracted respectively, and the Hessian matrix at the corresponding scale is constructed based on the second-order gray-level change information. Based on the Hessian matrix, Frangi filtering is performed on the neighborhood of each voxel to determine the degree of matching between the neighborhood of each voxel and the pulmonary vascular tubular structure, thereby obtaining vascular tubular response images at various scales. The tubular response images of blood vessels at multiple scales are fused to obtain fused tubular response images of blood vessels; The fused vascular tubular response image is subjected to background suppression and intensity normalization to obtain the enhanced vascular image.

4. The method according to claim 1, characterized in that, The deep learning network includes an encoder and a decoder; The encoder has a channel attention module embedded at its end; the encoder is used to extract multi-level lung image features from dual-channel lung images, and at the end, the channel attention module performs channel attention recalibration on the multi-level lung image features to obtain thrombus enhancement features. The decoder has a global attention module embedded in its deep feature region; the decoder is used to upsample and recover the thrombus enhancement features, and capture the global dependencies in the thrombus enhancement features in the deep feature region through the global attention module, establish global associations between distal voxels, and output a thrombus segmentation probability map.

5. The method according to claim 2, characterized in that, The method further includes: With the goal of balancing positive and negative samples in thrombus segmentation, a joint loss function is constructed. The joint loss function includes: an overlap constraint loss to reduce missed thrombus segmentation, a morphological constraint loss to constrain the continuity of thrombus boundaries, and a classification constraint loss to enhance the attention of difficult samples.

6. The method according to claim 5, characterized in that, Based on the thrombus segmentation probability map, the loss of the joint loss function is calculated, including: Obtain the thrombus annotation map corresponding to the original image, and perform a voxel-level comparison between the thrombus segmentation probability map and the thrombus annotation map to determine the thrombus overlapping area, the thrombus missing segmentation area, and the thrombus missegmentation area. The overlap constraint loss is calculated based on the thrombus overlap region, the thrombus unsegmented region, and the thrombus missegmented region, and the penalty weight corresponding to the thrombus unsegmented region is made greater than the penalty weight corresponding to the thrombus missegmented region. Boundary extraction is performed on the thrombus segmentation probability map and the thrombus annotation map to obtain the predicted thrombus boundary and the labeled thrombus boundary. The morphological constraint loss is calculated based on the boundary deviation between the predicted thrombus boundary and the labeled thrombus boundary. Based on the prediction confidence of each voxel in the thrombus segmentation probability map and the category labeling of each voxel in the thrombus labeling map, the easy-to-segment sample voxels and difficult-to-segment sample voxels are determined. The classification constraint loss is obtained by reducing the contribution weight of easily classifiable voxels in the classification loss and increasing the contribution weight of difficult voxels in the classification loss.

7. The method according to claim 6, characterized in that, Based on the homoscedastic uncertainty learning strategy, the loss weights of the overlap constraint loss, morphological constraint loss, and classification constraint loss are dynamically updated, including: Learnable uncertainty parameters are set for the overlap constraint loss, morphological constraint loss, and classification constraint loss, respectively; In each round of iterative training, the overlap constraint loss, morphological constraint loss and classification constraint loss for the current round are calculated based on the thrombus segmentation probability map and thrombus annotation map corresponding to the original image. Based on the learnable uncertainty parameters, the current loss weights corresponding to the overlap constraint loss, the morphological constraint loss, and the classification constraint loss are determined respectively, wherein the current loss weights are negatively correlated with the uncertainty of the corresponding loss. According to the current loss weight, the overlapping constraint loss, the morphological constraint loss and the classification constraint loss are weighted and fused, and combined with the constraint terms corresponding to each learnable uncertainty parameter, to obtain the dynamic joint loss value; Backpropagation is performed based on the dynamic joint loss value to synchronously update the network parameters of the deep learning network and each of the learnable uncertainty parameters; The loss weights of the overlap constraint loss, the morphological constraint loss, and the classification constraint loss are re-determined based on the updated learnable uncertainty parameters.

8. The method according to claim 1, characterized in that, The method further includes: The pulmonary embolism thrombus segmentation model was validated based on the validation set.

9. The method according to claim 8, characterized in that, The pulmonary embolism thrombus segmentation model is validated based on the validation set, including: Input the three-dimensional lung images from the validation set into the pulmonary embolism thrombus segmentation model to obtain the validation thrombus segmentation probability map; The verification thrombus segmentation probability map is binarized according to a preset probability threshold to obtain the verification thrombus three-dimensional segmentation result. The three-dimensional segmentation results of the verified thrombus are compared with the corresponding thrombus annotation information at the voxel level to obtain the verification overlapping region, the verification missing segmentation region, and the verification missegmentation region. Based on the verification overlap region, the verification missed segmentation region, and the verification missegmentation region, calculate the overlap similarity evaluation value, the recall evaluation value, and the precision evaluation value; Extract the predicted boundary of the three-dimensional segmentation result of the verified thrombus and the labeled boundary of the thrombus annotation information, and calculate the boundary accuracy evaluation value based on the boundary deviation between the predicted boundary and the labeled boundary; The validation results of the pulmonary embolism thrombus segmentation model are determined based on the overlap similarity evaluation value, the recall evaluation value, the precision evaluation value, and the boundary precision evaluation value.

10. A pulmonary embolism thrombus segmentation system based on deep learning networks, characterized in that, include: Processor and memory; The processor and memory are connected via a communication bus: The processor is used to call and execute the program stored in the memory; The memory is used to store a program, which is at least used to execute the pulmonary embolism thrombus segmentation method based on a deep learning network as described in any one of claims 1-9.

Images